chore(legal): regenerate THIRD-PARTY-NOTICES for the 0.1.15 dependency set

`cargo about generate about.hbs -o THIRD-PARTY-NOTICES.md` on the 0.1.15 lock: kigi crates at 0.1.15, chacha20 0.10.2. Claude-Session: https://claude.ai/code/session_01RUqV8HVGXcAETyYbZ2tPQ4
chore(release): bump workspace version to 0.1.15
2026-09-13 13:38:08 -04:00 · 2026-09-13 13:37:16 -04:00 · 2026-09-13 13:36:58 -04:00 · 2026-09-13 13:36:58 -04:00 · 2026-09-13 13:36:58 -04:00 · 2026-09-13 13:17:52 -04:00
216 changed files with 12269 additions and 1705 deletions
@@ -1,375 +0,0 @@
-name: Release
-
-# PRD F8, Gitea edition. On a v* tag push, build the single-file kigi binary
-# for the five supported targets, package kigi-<version>-<target-triple>.
-# {tar.gz|zip} (binary + LICENSE + NOTICE + THIRD-PARTY-NOTICES), and publish
-# them as a Gitea Release alongside SHA256SUMS. install.sh, install.ps1 and
-# auto_update::release_asset_name() all resolve these exact names — keep the
-# naming in lockstep with them.
-#
-# RUNNERS. Gitea has no hosted runners; every label below must be a runner you
-# registered yourself (docs/RELEASE.md has the register commands):
-#
-#   ubuntu-latest   Linux x86_64, container mode — the label the Dokploy box's
-#                   runner already advertises. Builds both Linux targets;
-#                   aarch64 is a cross-compile, so this leg installs the
-#                   aarch64 GNU toolchain. Because the image is a bare
-#                   runner-image, the Linux legs provision their own apt
-#                   packages and rustup (see the provision steps below).
-#                   Registering a dedicated `linux-x86_64:docker://...` runner
-#                   instead is fine — swap the three labels back if you do.
-#   macos-arm64     Apple Silicon Mac, host mode. Builds both Darwin targets;
-#                   x86_64 is a cross-compile against the same SDK. macOS
-#                   CANNOT be built anywhere else — Apple's SDK licence
-#                   restricts it to Apple hardware.
-#   windows-x86_64  Windows x86_64, host mode, with VS 2022 Build Tools
-#                   (C++ workload). Needed because the tree links ring,
-#                   mimalloc, zstd-sys and libz-sys against a static MSVC CRT
-#                   (-C target-feature=+crt-static).
-#
-# SHELLS. Every step names its shell explicitly and the Windows leg never uses
-# bash. On a normal Windows box `bash` resolves to WSL (not Git Bash), and
-# `pwsh` (PowerShell 7) is usually absent — so Windows steps use `powershell`,
-# which is the built-in 5.1. Keep those steps 5.1-compatible: no `&&`, no
-# ternary, no null-coalescing.
-#
-# Note: because this file exists, Gitea ignores .github/workflows entirely.
-#
-# PUBLISH ORDER. The release is created as a draft and only made public once
-# SHA256SUMS is attached, so `releases/latest` never hands the installer a
-# release whose checksum manifest is still missing.
-
-on:
-  push:
-    tags: ["v*"]
-
-permissions:
-  contents: write
-
-env:
-  CARGO_TERM_COLOR: always
-  # Matches the version pinned by the dotslash launcher in bin/protoc. That
-  # launcher is not used here: it has no Windows entry and needs a `cargo
-  # install dotslash` first, so every platform fetches protoc directly.
-  PROTOC_VERSION: "29.3"
-
-jobs:
-  prepare:
-    name: prepare draft release
-    runs-on: ubuntu-latest
-    outputs:
-      release_id: ${{ steps.draft.outputs.release_id }}
-    steps:
-      - uses: actions/checkout@v4
-
-      - name: Check tag matches workspace version
-        shell: bash
-        run: |
-          tag_version="${GITHUB_REF_NAME#v}"
-          cargo_version="$(sed -n 's/^version = "\(.*\)"$/\1/p' Cargo.toml | head -n 1)"
-          if [ "$tag_version" != "$cargo_version" ]; then
-            echo "Tag $GITHUB_REF_NAME does not match workspace version $cargo_version" >&2
-            exit 1
-          fi
-
-      - name: Create or reuse draft release
-        id: draft
-        shell: bash
-        env:
-          TOKEN: ${{ secrets.GITEA_TOKEN }}
-          API: ${{ github.server_url }}/api/v1/repos/${{ github.repository }}
-        run: |
-          set -eu
-          # The runner-image is bare; jq and curl are both used below.
-          if ! command -v jq >/dev/null || ! command -v curl >/dev/null; then
-            SUDO=""; command -v sudo >/dev/null && SUDO=sudo
-            export DEBIAN_FRONTEND=noninteractive
-            $SUDO apt-get update -qq
-            $SUDO apt-get install -y -qq --no-install-recommends jq curl ca-certificates
-          fi
-
-          # Idempotent so a re-run of a failed tag build reuses the same draft
-          # instead of erroring on a duplicate tag_name.
-          existing="$(curl -sS -H "Authorization: token $TOKEN" \
-            "$API/releases/tags/$GITHUB_REF_NAME")"
-          release_id="$(printf '%s' "$existing" | jq -r '.id // empty')"
-
-          if [ -z "$release_id" ]; then
-            prerelease=false
-            case "$GITHUB_REF_NAME" in *-*) prerelease=true ;; esac
-            created="$(curl -fsS -X POST \
-              -H "Authorization: token $TOKEN" \
-              -H "Content-Type: application/json" \
-              -d "{\"tag_name\":\"$GITHUB_REF_NAME\",\"name\":\"$GITHUB_REF_NAME\",\"draft\":true,\"prerelease\":$prerelease}" \
-              "$API/releases")"
-            release_id="$(printf '%s' "$created" | jq -r '.id')"
-            echo "Created draft release $release_id (prerelease=$prerelease)"
-          else
-            echo "Reusing release $release_id for $GITHUB_REF_NAME"
-          fi
-
-          echo "release_id=$release_id" >> "$GITHUB_OUTPUT"
-
-  build:
-    name: build (${{ matrix.target }})
-    needs: prepare
-    strategy:
-      fail-fast: false
-      matrix:
-        include:
-          - target: x86_64-unknown-linux-gnu
-            runner: ubuntu-latest
-            protoc: linux-x86_64
-            ext: tar.gz
-            linux: true
-          - target: aarch64-unknown-linux-gnu
-            runner: ubuntu-latest
-            protoc: linux-x86_64
-            ext: tar.gz
-            linux: true
-            cross_gnu: true
-          - target: aarch64-apple-darwin
-            runner: macos-arm64
-            protoc: osx-aarch_64
-            ext: tar.gz
-          - target: x86_64-apple-darwin
-            runner: macos-arm64
-            protoc: osx-aarch_64
-            ext: tar.gz
-          - target: x86_64-pc-windows-msvc
-            runner: windows-x86_64
-            protoc: win64
-            ext: zip
-    runs-on: ${{ matrix.runner }}
-    steps:
-      - uses: actions/checkout@v4
-
-      # ── linux container provisioning ──────────────────────────────────────
-      # The Linux legs run in a bare gitea runner-image, which ships git and
-      # node for the JavaScript actions but no compiler, no unzip (protoc) and
-      # no rustup. Host-mode macOS and Windows already have their toolchains
-      # installed, so this step is Linux-only.
-      - name: Provision Linux build deps
-        if: matrix.linux
-        shell: bash
-        run: |
-          set -eu
-          SUDO=""; command -v sudo >/dev/null && SUDO=sudo
-          export DEBIAN_FRONTEND=noninteractive
-          $SUDO apt-get update -qq
-          $SUDO apt-get install -y -qq --no-install-recommends \
-            build-essential pkg-config ca-certificates curl unzip tar jq
-
-      # ── protoc ────────────────────────────────────────────────────────────
-      - name: Install protoc (unix)
-        if: matrix.ext == 'tar.gz'
-        shell: bash
-        run: |
-          set -eu
-          url="https://github.com/protocolbuffers/protobuf/releases/download/v${PROTOC_VERSION}/protoc-${PROTOC_VERSION}-${{ matrix.protoc }}.zip"
-          curl -fsSL -o protoc.zip "$url"
-          rm -rf "$HOME/protoc"
-          unzip -q -o protoc.zip -d "$HOME/protoc"
-          chmod +x "$HOME/protoc/bin/protoc"
-          echo "$HOME/protoc/bin" >> "$GITHUB_PATH"
-
-      - name: Install protoc (windows)
-        if: matrix.ext == 'zip'
-        shell: powershell
-        run: |
-          $ErrorActionPreference = "Stop"
-          $url = "https://github.com/protocolbuffers/protobuf/releases/download/v$env:PROTOC_VERSION/protoc-$env:PROTOC_VERSION-${{ matrix.protoc }}.zip"
-          Invoke-WebRequest -Uri $url -OutFile protoc.zip
-          $dest = Join-Path $env:USERPROFILE "protoc"
-          if (Test-Path $dest) { Remove-Item $dest -Recurse -Force }
-          Expand-Archive protoc.zip -DestinationPath $dest -Force
-          Add-Content -Path $env:GITHUB_PATH -Value (Join-Path $dest "bin")
-
-      # ── toolchain ─────────────────────────────────────────────────────────
-      # rust-toolchain.toml pins the channel (1.97.0) and preinstalls both
-      # Linux targets; `rustup show` materialises exactly that.
-      - name: Install toolchain (unix)
-        if: matrix.ext == 'tar.gz'
-        shell: bash
-        run: |
-          set -eu
-          # No-op on the macOS host runner, which has rustup installed already.
-          # `--default-toolchain none` leaves the channel choice to
-          # rust-toolchain.toml, which `rustup show` then materialises.
-          if ! command -v rustup >/dev/null; then
-            curl --proto '=https' --tlsv1.2 -fsSL https://sh.rustup.rs \
-              | sh -s -- -y --no-modify-path --profile minimal --default-toolchain none
-            echo "$HOME/.cargo/bin" >> "$GITHUB_PATH"
-            export PATH="$HOME/.cargo/bin:$PATH"
-          fi
-          rustup show
-          rustup target add ${{ matrix.target }}
-
-      - name: Install toolchain (windows)
-        if: matrix.ext == 'zip'
-        shell: powershell
-        run: |
-          $ErrorActionPreference = "Stop"
-          rustup show
-          rustup target add ${{ matrix.target }}
-
-      # ring, mimalloc, zstd-sys and libz-sys are all cc-rs builds, so the
-      # cross target needs its own C toolchain — cc-rs picks these up from the
-      # target-suffixed vars.
-      - name: Install aarch64 cross toolchain
-        if: matrix.cross_gnu
-        shell: bash
-        run: |
-          set -eu
-          SUDO=""; command -v sudo >/dev/null && SUDO=sudo
-          $SUDO apt-get update -qq
-          $SUDO apt-get install -y -qq gcc-aarch64-linux-gnu g++-aarch64-linux-gnu
-          {
-            echo "CARGO_TARGET_AARCH64_UNKNOWN_LINUX_GNU_LINKER=aarch64-linux-gnu-gcc"
-            echo "CC_aarch64_unknown_linux_gnu=aarch64-linux-gnu-gcc"
-            echo "CXX_aarch64_unknown_linux_gnu=aarch64-linux-gnu-g++"
-            echo "AR_aarch64_unknown_linux_gnu=aarch64-linux-gnu-ar"
-          } >> "$GITHUB_ENV"
-
-      # ── build ─────────────────────────────────────────────────────────────
-      - name: Build kigi (unix)
-        if: matrix.ext == 'tar.gz'
-        shell: bash
-        run: cargo build --profile release-dist -p kigi-bin --locked --target ${{ matrix.target }}
-
-      - name: Build kigi (windows)
-        if: matrix.ext == 'zip'
-        shell: powershell
-        run: |
-          $ErrorActionPreference = "Stop"
-          cargo build --profile release-dist -p kigi-bin --locked --target ${{ matrix.target }}
-          if ($LASTEXITCODE -ne 0) { exit $LASTEXITCODE }
-
-      # ── package, upload, checksum ─────────────────────────────────────────
-      # Archives go straight to the release rather than through the artifact
-      # store: they are large, and the release is the only place install.sh
-      # can read them from anyway. Only the checksum line travels as an
-      # artifact, for the publish job to concatenate.
-      - name: Package and upload (unix)
-        if: matrix.ext == 'tar.gz'
-        shell: bash
-        env:
-          TOKEN: ${{ secrets.GITEA_TOKEN }}
-          API: ${{ github.server_url }}/api/v1/repos/${{ github.repository }}
-          RELEASE_ID: ${{ needs.prepare.outputs.release_id }}
-        run: |
-          set -eu
-          version="${GITHUB_REF_NAME#v}"
-          archive="kigi-${version}-${{ matrix.target }}.tar.gz"
-          staging="$(mktemp -d)"
-          cp "target/${{ matrix.target }}/release-dist/kigi" "$staging/kigi"
-          cp LICENSE "$staging/"
-          [ -f NOTICE ] && cp NOTICE "$staging/"
-          for f in THIRD-PARTY-NOTICES THIRD-PARTY-NOTICES.md; do
-            [ -f "$f" ] && cp "$f" "$staging/"
-          done
-          tar -C "$staging" -czf "$archive" .
-
-          curl -fsS -X POST \
-            -H "Authorization: token $TOKEN" \
-            -F "attachment=@${archive}" \
-            "$API/releases/$RELEASE_ID/assets?name=${archive}" > /dev/null
-          echo "Uploaded $archive"
-
-          if command -v sha256sum >/dev/null; then
-            sha256sum "$archive" > "checksum-${{ matrix.target }}.txt"
-          else
-            shasum -a 256 "$archive" > "checksum-${{ matrix.target }}.txt"
-          fi
-          cat "checksum-${{ matrix.target }}.txt"
-
-      - name: Package and upload (windows)
-        if: matrix.ext == 'zip'
-        shell: powershell
-        env:
-          TOKEN: ${{ secrets.GITEA_TOKEN }}
-          API: ${{ github.server_url }}/api/v1/repos/${{ github.repository }}
-          RELEASE_ID: ${{ needs.prepare.outputs.release_id }}
-        run: |
-          $ErrorActionPreference = "Stop"
-          $version = $env:GITHUB_REF_NAME.TrimStart("v")
-          $archive = "kigi-$version-${{ matrix.target }}.zip"
-          $staging = Join-Path $env:RUNNER_TEMP "staging"
-          if (Test-Path $staging) { Remove-Item $staging -Recurse -Force }
-          New-Item -ItemType Directory -Force -Path $staging | Out-Null
-          Copy-Item "target/${{ matrix.target }}/release-dist/kigi.exe" (Join-Path $staging "kigi.exe")
-          Copy-Item LICENSE $staging
-          if (Test-Path NOTICE) { Copy-Item NOTICE $staging }
-          foreach ($f in @("THIRD-PARTY-NOTICES", "THIRD-PARTY-NOTICES.md")) {
-            if (Test-Path $f) { Copy-Item $f $staging }
-          }
-          Compress-Archive -Path (Join-Path $staging "*") -DestinationPath $archive -Force
-
-          # System32\curl.exe, not PowerShell's `curl` alias for
-          # Invoke-WebRequest, which cannot do multipart uploads.
-          $api = "$env:API/releases/$env:RELEASE_ID/assets?name=$archive"
-          & "$env:SystemRoot\System32\curl.exe" -fsS -X POST `
-            -H "Authorization: token $env:TOKEN" `
-            -F "attachment=@$archive" $api
-          if ($LASTEXITCODE -ne 0) { throw "asset upload failed ($LASTEXITCODE)" }
-          Write-Host "Uploaded $archive"
-
-          # sha256sum's exact format, so the publish job can just concatenate.
-          # ASCII keeps the BOM out — a BOM would corrupt the first hash.
-          $hash = (Get-FileHash -Algorithm SHA256 -Path $archive).Hash.ToLowerInvariant()
-          "$hash  $archive" | Out-File -FilePath "checksum-${{ matrix.target }}.txt" -Encoding ascii
-          Get-Content "checksum-${{ matrix.target }}.txt"
-
-      - uses: actions/upload-artifact@v3
-        with:
-          name: checksum-${{ matrix.target }}
-          path: checksum-${{ matrix.target }}.txt
-          if-no-files-found: error
-
-  publish:
-    name: publish release
-    needs: [prepare, build]
-    runs-on: ubuntu-latest
-    steps:
-      - uses: actions/download-artifact@v3
-        with:
-          path: checksums
-
-      - name: Assemble SHA256SUMS
-        shell: bash
-        run: |
-          set -eu
-          # download-artifact v3 nests each artifact in its own directory. The
-          # Windows line arrives CRLF-terminated; strip CR so every line is LF
-          # for install.sh's `read -r hash name`.
-          find checksums -name 'checksum-*.txt' -exec cat {} + \
-            | tr -d '\r' | grep -v '^$' | sort -k2 > SHA256SUMS
-          lines="$(wc -l < SHA256SUMS)"
-          if [ "$lines" -ne 5 ]; then
-            echo "expected 5 checksum lines, got $lines" >&2
-            cat SHA256SUMS >&2
-            exit 1
-          fi
-          cat SHA256SUMS
-
-      - name: Attach SHA256SUMS and publish
-        shell: bash
-        env:
-          TOKEN: ${{ secrets.GITEA_TOKEN }}
-          API: ${{ github.server_url }}/api/v1/repos/${{ github.repository }}
-          RELEASE_ID: ${{ needs.prepare.outputs.release_id }}
-        run: |
-          set -eu
-          curl -fsS -X POST \
-            -H "Authorization: token $TOKEN" \
-            -F "attachment=@SHA256SUMS" \
-            "$API/releases/$RELEASE_ID/assets?name=SHA256SUMS" > /dev/null
-
-          # Flip the draft public only now that every asset install.sh looks
-          # for is attached.
-          curl -fsS -X PATCH \
-            -H "Authorization: token $TOKEN" \
-            -H "Content-Type: application/json" \
-            -d '{"draft":false}' \
-            "$API/releases/$RELEASE_ID" > /dev/null
-          echo "Published $GITHUB_REF_NAME"
@@ -14,6 +14,22 @@ official `kimi` CLI: binary `kigi`, config dir `~/.kigi`
 Never read or write `~/.kimi` (except the explicit one-time read-only
 import) or any `KIMI_*` env var.

+Upstream sync record (update on every sync): fork baseline is upstream
+snapshot `8adf901` (2026-07-16, ≈v0.2.102); bug fixes evaluated and
+selectively ported through upstream `77cd7eb` (2026-08-25, ≈v1.0.10) in
+the 0.1.13 cycle and through `3794978` (2026-09-09, ≈v1.0.24) in the
+0.1.15 cycle. Upstream is daily "Synced from monorepo" snapshots —
+the commit BODIES carry per-change bullet lists, and
+`crates/codegen/xai-grok-shell/changelogs/` maps releases to dates;
+start any future sync from those two, not from raw diffs.
+
+Known limitations left by the 0.1.15 sync (fix before relying on them):
+the Windows frame writer writes through `std::io::stderr`, so a console
+that stops reading still blocks exit-path prints behind std's stderr lock
+(unix silences fd 2 after a timed-out writer join); the bwrap command
+builder re-checks `Path::exists("/data")` and can omit the devbox bind on
+a stat error that the profile classification treated as present.
+
 ## Hard constraints

 - **Zero egress**: only `auth.kimi.com`, `api.kimi.com`, `api.moonshot.cn`,
@@ -195,6 +211,13 @@ client-side, no backend surface.
  (`MAX_SWARM_RUNTIME`) are bounded: the swarm blocks the caller's turn,
  so every wait needs a ceiling. Stragglers at the deadline are reported
  as still-running WITH their ids.
+- Underneath every fan-out (swarm, graph, raw Task) sits the
+  session-scoped spawn bound: `handle_subagent_request` holds one permit
+  from the parent session's semaphore for the whole child run
+  (kigi-tools `task/admission`: 32 default via
+  `KIGI_MAX_CONCURRENT_SUBAGENTS`, queue|fail via
+  `KIGI_SUBAGENT_LIMIT_BEHAVIOR`), so wide fan-outs queue instead of
+  exhausting file descriptors.
 - `ToolKind::AgentSwarm` is its own variant because `TemplateRenderer`'s
  `by_kind` map holds ONE tool name per kind — sharing `Task` would
  silently redirect `${{ tools.by_kind.task }}` in other tools' prompts.
@@ -228,7 +228,7 @@ version = "1.1.5"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "40c48f72fd53cd289104fc64099abca73db4166ad86ea0b4341abe65af83dadc"
 dependencies = [
- "windows-sys 0.61.2",
+ "windows-sys 0.60.2",
 ]

 [[package]]
@@ -251,7 +251,7 @@ checksum = "291e6a250ff86cd4a820112fb8898808a366d8f9f58ce16d1f538353ad55747d"
 dependencies = [
 "anstyle",
 "once_cell_polyfill",
- "windows-sys 0.61.2",
+ "windows-sys 0.60.2",
 ]

 [[package]]
@@ -1626,9 +1626,9 @@ checksum = "f079e83a288787bcd14a6aea84cee5c87a67c5a3e660c30f557a3d24761b3527"

 [[package]]
 name = "chacha20"
-version = "0.10.1"
+version = "0.10.2"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "d524456ba66e72eb8b115ff89e01e497f8e6d11d78b70b1aa13c0fbd97540a81"
+checksum = "65c35e4b699c7e15ccbe7ee35c005e4fc0a278d22238a2857e6ce2dadeda1b06"
 dependencies = [
 "cfg-if",
 "cpufeatures 0.3.0",
@@ -1810,7 +1810,7 @@ version = "3.1.1"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "faf9468729b8cbcea668e36183cb69d317348c2e08e994829fb56ebfdfbaac34"
 dependencies = [
- "windows-sys 0.61.2",
+ "windows-sys 0.48.0",
 ]

 [[package]]
@@ -2654,7 +2654,7 @@ dependencies = [
 "libc",
 "option-ext",
 "redox_users 0.5.2",
- "windows-sys 0.61.2",
+ "windows-sys 0.59.0",
 ]

 [[package]]
@@ -2980,7 +2980,7 @@ source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "39cab71617ae0d63f51a36d69f866391735b51691dbda63cf6f96d042b63efeb"
 dependencies = [
 "libc",
- "windows-sys 0.61.2",
+ "windows-sys 0.52.0",
 ]

 [[package]]
@@ -3958,7 +3958,7 @@ source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "b305d85504de270ad3525d726a6b69cc59ee7b2269b014387651107ab9f0755b"
 dependencies = [
 "bstr",
- "hashbrown 0.17.1",
+ "hashbrown 0.15.5",
 ]

 [[package]]
@@ -4497,9 +4497,9 @@ dependencies = [

 [[package]]
 name = "h2"
-version = "0.4.15"
+version = "0.4.19"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "6cb093c84e8bd9b188d4c4a8cb6579fc016968d14c99882163cd3ff402a4f155"
+checksum = "ef8e5e5a340588f4452631496976cf8636d4a7ecf600239fdc27615d2530bc16"
 dependencies = [
 "atomic-waker",
 "bytes",
@@ -5442,7 +5442,7 @@ dependencies = [

 [[package]]
 name = "kigi-acp-lib"
-version = "0.1.12"
+version = "0.1.15"
 dependencies = [
 "agent-client-protocol",
 "async-trait",
@@ -5456,7 +5456,7 @@ dependencies = [

 [[package]]
 name = "kigi-agent"
-version = "0.1.12"
+version = "0.1.15"
 dependencies = [
 "chrono",
 "dirs 6.0.0",
@@ -5486,7 +5486,7 @@ dependencies = [

 [[package]]
 name = "kigi-agent-lifecycle"
-version = "0.1.12"
+version = "0.1.15"
 dependencies = [
 "async-trait",
 "tokio",
@@ -5495,7 +5495,7 @@ dependencies = [

 [[package]]
 name = "kigi-auth"
-version = "0.1.12"
+version = "0.1.15"
 dependencies = [
 "async-trait",
 "http 1.4.2",
@@ -5508,7 +5508,7 @@ dependencies = [

 [[package]]
 name = "kigi-bin"
-version = "0.1.12"
+version = "0.1.15"
 dependencies = [
 "anyhow",
 "clap",
@@ -5543,7 +5543,7 @@ dependencies = [

 [[package]]
 name = "kigi-chat-state"
-version = "0.1.12"
+version = "0.1.15"
 dependencies = [
 "indexmap",
 "kigi-compaction",
@@ -5560,7 +5560,7 @@ dependencies = [

 [[package]]
 name = "kigi-codebase-graph"
-version = "0.1.12"
+version = "0.1.15"
 dependencies = [
 "ahash",
 "clap",
@@ -5596,7 +5596,7 @@ dependencies = [

 [[package]]
 name = "kigi-compaction"
-version = "0.1.12"
+version = "0.1.15"
 dependencies = [
 "anyhow",
 "async-trait",
@@ -5609,7 +5609,7 @@ dependencies = [

 [[package]]
 name = "kigi-config"
-version = "0.1.12"
+version = "0.1.15"
 dependencies = [
 "base64",
 "blake3",
@@ -5632,7 +5632,7 @@ dependencies = [

 [[package]]
 name = "kigi-config-types"
-version = "0.1.12"
+version = "0.1.15"
 dependencies = [
 "agent-client-protocol",
 "indexmap",
@@ -5646,7 +5646,7 @@ dependencies = [

 [[package]]
 name = "kigi-crash-handler"
-version = "0.1.12"
+version = "0.1.15"
 dependencies = [
 "backtrace",
 "libc",
@@ -5657,7 +5657,7 @@ dependencies = [

 [[package]]
 name = "kigi-env"
-version = "0.1.12"
+version = "0.1.15"
 dependencies = [
 "tracing",
 "url",
@@ -5665,7 +5665,7 @@ dependencies = [

 [[package]]
 name = "kigi-fast-worktree"
-version = "0.1.12"
+version = "0.1.15"
 dependencies = [
 "anyhow",
 "bytes",
@@ -5697,7 +5697,7 @@ dependencies = [

 [[package]]
 name = "kigi-file-utils"
-version = "0.1.12"
+version = "0.1.15"
 dependencies = [
 "anyhow",
 "aws-config",
@@ -5721,7 +5721,7 @@ dependencies = [

 [[package]]
 name = "kigi-fsnotify"
-version = "0.1.12"
+version = "0.1.15"
 dependencies = [
 "criterion",
 "dunce",
@@ -5742,7 +5742,7 @@ dependencies = [

 [[package]]
 name = "kigi-gix-status"
-version = "0.1.12"
+version = "0.1.15"
 dependencies = [
 "gix",
 "kigi-test-utils",
@@ -5752,7 +5752,7 @@ dependencies = [

 [[package]]
 name = "kigi-hooks"
-version = "0.1.12"
+version = "0.1.15"
 dependencies = [
 "fastrand",
 "kigi-config",
@@ -5771,7 +5771,7 @@ dependencies = [

 [[package]]
 name = "kigi-hooks-plugins-types"
-version = "0.1.12"
+version = "0.1.15"
 dependencies = [
 "serde",
 "serde_json",
@@ -5779,7 +5779,7 @@ dependencies = [

 [[package]]
 name = "kigi-http"
-version = "0.1.12"
+version = "0.1.15"
 dependencies = [
 "kigi-auth",
 "kigi-log",
@@ -5794,7 +5794,7 @@ dependencies = [

 [[package]]
 name = "kigi-hunk-tracker"
-version = "0.1.12"
+version = "0.1.15"
 dependencies = [
 "chrono",
 "dunce",
@@ -5815,14 +5815,14 @@ dependencies = [

 [[package]]
 name = "kigi-interjection-core"
-version = "0.1.12"
+version = "0.1.15"
 dependencies = [
 "serde",
 ]

 [[package]]
 name = "kigi-log"
-version = "0.1.12"
+version = "0.1.15"
 dependencies = [
 "anyhow",
 "chrono",
@@ -5840,7 +5840,7 @@ dependencies = [

 [[package]]
 name = "kigi-markdown"
-version = "0.1.12"
+version = "0.1.15"
 dependencies = [
 "anstyle",
 "anstyle-lossy",
@@ -5864,14 +5864,14 @@ dependencies = [

 [[package]]
 name = "kigi-markdown-core"
-version = "0.1.12"
+version = "0.1.15"
 dependencies = [
 "pulldown-cmark",
 ]

 [[package]]
 name = "kigi-mcp"
-version = "0.1.12"
+version = "0.1.15"
 dependencies = [
 "agent-client-protocol",
 "async-trait",
@@ -5908,7 +5908,7 @@ dependencies = [

 [[package]]
 name = "kigi-memory"
-version = "0.1.12"
+version = "0.1.15"
 dependencies = [
 "anyhow",
 "arc-swap",
@@ -5942,7 +5942,7 @@ dependencies = [

 [[package]]
 name = "kigi-mermaid"
-version = "0.1.12"
+version = "0.1.15"
 dependencies = [
 "fontdb",
 "image",
@@ -5960,7 +5960,7 @@ dependencies = [

 [[package]]
 name = "kigi-models"
-version = "0.1.12"
+version = "0.1.15"
 dependencies = [
 "kigi-env",
 "serde",
@@ -5970,7 +5970,7 @@ dependencies = [

 [[package]]
 name = "kigi-pager-minimal"
-version = "0.1.12"
+version = "0.1.15"
 dependencies = [
 "chrono",
 "crossterm",
@@ -5987,7 +5987,7 @@ dependencies = [

 [[package]]
 name = "kigi-pager-pty-harness"
-version = "0.1.12"
+version = "0.1.15"
 dependencies = [
 "alacritty_terminal",
 "anyhow",
@@ -6012,7 +6012,7 @@ dependencies = [

 [[package]]
 name = "kigi-pager-render"
-version = "0.1.12"
+version = "0.1.15"
 dependencies = [
 "agent-client-protocol",
 "anstyle",
@@ -6064,7 +6064,7 @@ dependencies = [

 [[package]]
 name = "kigi-paths"
-version = "0.1.12"
+version = "0.1.15"
 dependencies = [
 "camino",
 "serde",
@@ -6074,7 +6074,7 @@ dependencies = [

 [[package]]
 name = "kigi-prompt-queue"
-version = "0.1.12"
+version = "0.1.15"
 dependencies = [
 "serde",
 "serde_json",
@@ -6082,7 +6082,7 @@ dependencies = [

 [[package]]
 name = "kigi-proto-build"
-version = "0.1.12"
+version = "0.1.15"
 dependencies = [
 "anyhow",
 "pbjson-build",
@@ -6093,7 +6093,7 @@ dependencies = [

 [[package]]
 name = "kigi-ratatui-inline"
-version = "0.1.12"
+version = "0.1.15"
 dependencies = [
 "ansi-width",
 "anstyle-parse 0.2.7",
@@ -6110,7 +6110,7 @@ dependencies = [

 [[package]]
 name = "kigi-ratatui-textarea"
-version = "0.1.12"
+version = "0.1.15"
 dependencies = [
 "arboard",
 "chrono",
@@ -6131,7 +6131,7 @@ dependencies = [

 [[package]]
 name = "kigi-sampler"
-version = "0.1.12"
+version = "0.1.15"
 dependencies = [
 "async-openai",
 "async-stream",
@@ -6141,6 +6141,7 @@ dependencies = [
 "futures-util",
 "indexmap",
 "kigi-sampling-types",
+ "kigi-secrets",
 "kigi-test-support",
 "kigi-version",
 "reqwest 0.12.28",
@@ -6154,7 +6155,7 @@ dependencies = [

 [[package]]
 name = "kigi-sampling-types"
-version = "0.1.12"
+version = "0.1.15"
 dependencies = [
 "assert_matches",
 "async-openai",
@@ -6171,7 +6172,7 @@ dependencies = [

 [[package]]
 name = "kigi-sandbox"
-version = "0.1.12"
+version = "0.1.15"
 dependencies = [
 "anyhow",
 "chrono",
@@ -6186,13 +6187,14 @@ dependencies = [
 "serde",
 "serde_json",
 "serial_test",
+ "tempfile",
 "toml",
 "tracing",
 ]

 [[package]]
 name = "kigi-secrets"
-version = "0.1.12"
+version = "0.1.15"
 dependencies = [
 "regex",
 "serde_json",
@@ -6230,7 +6232,7 @@ dependencies = [

 [[package]]
 name = "kigi-shell"
-version = "0.1.12"
+version = "0.1.15"
 dependencies = [
 "agent-client-protocol",
 "anyhow",
@@ -6293,6 +6295,7 @@ dependencies = [
 "kigi-sampler",
 "kigi-sampling-types",
 "kigi-sandbox",
+ "kigi-secrets",
 "kigi-shared",
 "kigi-shell-base",
 "kigi-sqlite-journal",
@@ -6367,7 +6370,7 @@ dependencies = [

 [[package]]
 name = "kigi-shell-base"
-version = "0.1.12"
+version = "0.1.15"
 dependencies = [
 "anyhow",
 "chrono",
@@ -6392,7 +6395,7 @@ dependencies = [

 [[package]]
 name = "kigi-sqlite-journal"
-version = "0.1.12"
+version = "0.1.15"
 dependencies = [
 "libc",
 "rusqlite",
@@ -6403,7 +6406,7 @@ dependencies = [

 [[package]]
 name = "kigi-subagent-resolution"
-version = "0.1.12"
+version = "0.1.15"
 dependencies = [
 "kigi-sampling-types",
 "kigi-tool-types",
@@ -6418,7 +6421,7 @@ dependencies = [

 [[package]]
 name = "kigi-system-power"
-version = "0.1.12"
+version = "0.1.15"
 dependencies = [
 "windows-sys 0.59.0",
 "zbus",
@@ -6426,7 +6429,7 @@ dependencies = [

 [[package]]
 name = "kigi-test-support"
-version = "0.1.12"
+version = "0.1.15"
 dependencies = [
 "agent-client-protocol",
 "anyhow",
@@ -6448,7 +6451,7 @@ dependencies = [

 [[package]]
 name = "kigi-test-utils"
-version = "0.1.12"
+version = "0.1.15"
 dependencies = [
 "runfiles",
 "tracing",
@@ -6457,11 +6460,11 @@ dependencies = [

 [[package]]
 name = "kigi-token-estimation"
-version = "0.1.12"
+version = "0.1.15"

 [[package]]
 name = "kigi-tool-protocol"
-version = "0.1.12"
+version = "0.1.15"
 dependencies = [
 "kigi-tool-types",
 "serde",
@@ -6472,7 +6475,7 @@ dependencies = [

 [[package]]
 name = "kigi-tool-runtime"
-version = "0.1.12"
+version = "0.1.15"
 dependencies = [
 "anyhow",
 "async-trait",
@@ -6490,7 +6493,7 @@ dependencies = [

 [[package]]
 name = "kigi-tool-types"
-version = "0.1.12"
+version = "0.1.15"
 dependencies = [
 "minijinja",
 "schemars 1.2.1",
@@ -6500,7 +6503,7 @@ dependencies = [

 [[package]]
 name = "kigi-tools"
-version = "0.1.12"
+version = "0.1.15"
 dependencies = [
 "anyhow",
 "arc-swap",
@@ -6577,7 +6580,7 @@ dependencies = [

 [[package]]
 name = "kigi-tools-api"
-version = "0.1.12"
+version = "0.1.15"
 dependencies = [
 "kigi-proto-build",
 "kigi-tool-protocol",
@@ -6590,11 +6593,11 @@ dependencies = [

 [[package]]
 name = "kigi-tracing-macros"
-version = "0.1.12"
+version = "0.1.15"

 [[package]]
 name = "kigi-tty-utils"
-version = "0.1.12"
+version = "0.1.15"
 dependencies = [
 "libc",
 "nix 0.30.1",
@@ -6604,7 +6607,7 @@ dependencies = [

 [[package]]
 name = "kigi-tui"
-version = "0.1.12"
+version = "0.1.15"
 dependencies = [
 "agent-client-protocol",
 "ansi-to-tui",
@@ -6691,7 +6694,7 @@ dependencies = [

 [[package]]
 name = "kigi-update"
-version = "0.1.12"
+version = "0.1.15"
 dependencies = [
 "anyhow",
 "dunce",
@@ -6720,14 +6723,14 @@ dependencies = [

 [[package]]
 name = "kigi-version"
-version = "0.1.12"
+version = "0.1.15"
 dependencies = [
 "semver",
 ]

 [[package]]
 name = "kigi-workspace"
-version = "0.1.12"
+version = "0.1.15"
 dependencies = [
 "agent-client-protocol",
 "anyhow",
@@ -6806,7 +6809,7 @@ dependencies = [

 [[package]]
 name = "kigi-workspace-types"
-version = "0.1.12"
+version = "0.1.15"
 dependencies = [
 "base64",
 "chrono",
@@ -7305,7 +7308,7 @@ version = "0.6.1"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "536bfad37a309d62069485248eeaba1e8d9853aaf951caaeaed0585a95346f08"
 dependencies = [
- "windows-sys 0.61.2",
+ "windows-sys 0.60.2",
 ]

 [[package]]
@@ -7576,7 +7579,7 @@ version = "0.50.3"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "7957b9740744892f114936ab4a57b3f487491bbeafaf8083688b16841a4240e5"
 dependencies = [
- "windows-sys 0.61.2",
+ "windows-sys 0.59.0",
 ]

 [[package]]
@@ -7997,7 +8000,7 @@ source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "7d8fae84b431384b68627d0f9b3b1245fcf9f46f6c0e3dc902e9dce64edd1967"
 dependencies = [
 "libc",
- "windows-sys 0.61.2",
+ "windows-sys 0.48.0",
 ]

 [[package]]
@@ -8840,7 +8843,7 @@ dependencies = [

 [[package]]
 name = "ptyctl"
-version = "0.1.12"
+version = "0.1.15"
 dependencies = [
 "alacritty_terminal",
 "anyhow",
@@ -8858,7 +8861,7 @@ dependencies = [

 [[package]]
 name = "ptyctl-cli"
-version = "0.1.12"
+version = "0.1.15"
 dependencies = [
 "anyhow",
 "axum",
@@ -8992,7 +8995,7 @@ dependencies = [
 "once_cell",
 "socket2",
 "tracing",
- "windows-sys 0.61.2",
+ "windows-sys 0.52.0",
 ]

 [[package]]
@@ -9617,7 +9620,7 @@ dependencies = [
 "errno",
 "libc",
 "linux-raw-sys 0.12.1",
- "windows-sys 0.61.2",
+ "windows-sys 0.52.0",
 ]

 [[package]]
@@ -9687,7 +9690,7 @@ dependencies = [
 "security-framework 3.7.0",
 "security-framework-sys",
 "webpki-root-certs",
- "windows-sys 0.61.2",
+ "windows-sys 0.52.0",
 ]

 [[package]]
@@ -10536,7 +10539,7 @@ source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "c3d1e2c7f27f8d4cb10542a02c49005dbd6e93095799d6f3be745fae9f8fedd4"
 dependencies = [
 "libc",
- "windows-sys 0.61.2",
+ "windows-sys 0.60.2",
 ]

 [[package]]
@@ -10937,7 +10940,7 @@ dependencies = [
 "getrandom 0.4.3",
 "once_cell",
 "rustix 1.1.4",
- "windows-sys 0.61.2",
+ "windows-sys 0.52.0",
 ]

 [[package]]
@@ -11882,7 +11885,7 @@ checksum = "f2f6fb2847f6742cd76af783a2a2c49e9375d0a111c7bef6f71cd9e738c72d6e"
 dependencies = [
 "memoffset",
 "tempfile",
- "windows-sys 0.61.2",
+ "windows-sys 0.60.2",
 ]

 [[package]]
@@ -12414,15 +12417,15 @@ dependencies = [

 [[package]]
 name = "webbrowser"
-version = "1.2.1"
+version = "1.2.4"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "0fc95580916af1e68ff6a7be07446fc5db73ebf71cf092de939bbf5f7e189f72"
+checksum = "62c35be770821a214dbc362fc26908c853e776c0004294d0b10b8a6bad582f94"
 dependencies = [
- "core-foundation 0.10.1",
 "jni",
 "log",
 "ndk-context",
 "objc2 0.6.4",
+ "objc2-app-kit",
 "objc2-foundation 0.3.2",
 "url",
 "web-sys",
@@ -12561,7 +12564,7 @@ version = "0.1.11"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "c2a7b1c03c876122aa43f3020e6c3c3ee5c05081c9a00739faf7503aeba10d22"
 dependencies = [
- "windows-sys 0.61.2",
+ "windows-sys 0.48.0",
 ]

 [[package]]
@@ -76,7 +76,7 @@ members = [
 ]

 [workspace.package]
-version = "0.1.12"
+version = "0.1.15"
 edition = "2024"
 license = "Apache-2.0"

@@ -1,7 +1,7 @@
 # Security Policy

 Kigi is a community build; report vulnerabilities on the
-[Kigi issue tracker](https://git.zacharyzhang.com/ZacharyZhang-NY/Kigi-CLI/issues).
+[Kigi issue tracker](https://github.com/ZacharyZhang-NY/Kigi-CLI/issues).
 Issues there are public, so please leave out any detail that would make a
 report directly exploitable — open with the affected area and impact, and a
 maintainer will follow up for the rest.
@@ -252,11 +252,20 @@ pub fn by_name(name: &str) -> Option<AgentDefinition> {
    by_name_with_home(name, dirs::home_dir().as_deref(), kigi.as_deref())
 }

+/// An agent name is one file stem: a separator or `..` would resolve a file outside the searched directory with that directory's trust scope.
+fn is_single_path_component(name: &str) -> bool {
+    !name.is_empty() && name != "." && name != ".." && !name.contains(['/', '\\'])
+}
+
 fn by_name_with_home(
    name: &str,
    home: Option<&Path>,
    kigi_home: Option<&Path>,
 ) -> Option<AgentDefinition> {
+    if !is_single_path_component(name) {
+        tracing::warn!(name, "rejected agent name: not a single path component");
+        return None;
+    }
    // Check built-ins first — type-safe via BuiltinAgentName strum enum.
    // Legacy pre-rebrand agent types (persisted in old session files) are
    // mapped onto their current names first; see `canonical_agent_type`.
@@ -298,6 +307,10 @@ fn by_name_in_cwd_with_home(
    home: Option<&Path>,
    kigi_home: Option<&Path>,
 ) -> Option<AgentDefinition> {
+    if !is_single_path_component(name) {
+        tracing::warn!(name, "rejected agent name: not a single path component");
+        return None;
+    }
    if let Some(def) = load_project_definition_by_name(name, cwd) {
        return Some(def);
    }
@@ -827,6 +840,30 @@ mod tests {
        assert!(def.is_none());
    }

+    /// `../outside` must not load a file beyond the agents directory under that directory's scope.
+    #[test]
+    fn by_name_rejects_names_that_leave_the_agents_directory() {
+        let tmp = tempfile::tempdir().unwrap();
+        let project = tmp.path().join("project");
+        let agents_dir = project.join(".kigi").join("agents");
+        fs::create_dir_all(&agents_dir).unwrap();
+        write_agent_file(
+            &project,
+            "outside.md",
+            "outside",
+            "Lives outside the agents dir",
+        );
+        write_agent_file(&agents_dir, "inside.md", "inside", "Lives inside");
+        assert!(by_name_in_cwd_with_home("../../outside", &project, None, None).is_none());
+        assert!(by_name_in_cwd_with_home("..", &project, None, None).is_none());
+        assert!(by_name_in_cwd_with_home("a/b", &project, None, None).is_none());
+        assert!(by_name_with_home("../../outside", None, None).is_none());
+        assert_eq!(
+            by_name_in_cwd_with_home("inside", &project, None, None).map(|d| d.name),
+            Some("inside".to_string())
+        );
+    }
+
    #[test]
    fn test_discover_finds_md_files() {
        let tmp = tempfile::tempdir().unwrap();
@@ -17,6 +17,11 @@ pub enum AgentBuildError {
    #[error("IO error during agent construction: {0}")]
    IoError(#[from] std::io::Error),

+    /// The per-session runtime could not be built — fd or thread
+    /// exhaustion (EMFILE/EAGAIN), not a definition problem.
+    #[error("failed to build the session runtime: {0}")]
+    RuntimeBuild(std::io::Error),
+
    /// Carries template line numbers and surrounding context.
    #[error("template rendering error: {0}")]
    MiniJinjaError(#[from] minijinja::Error),
@@ -32,5 +32,5 @@ fn get_global_gitignore_path() -> Option<PathBuf> {
    git2::Config::open_default()
        .ok()
        .and_then(|cfg| cfg.get_path("core.excludesFile").ok())
-        .or_else(|| std::env::var_os("HOME").map(|h| PathBuf::from(h).join(".gitignore")))
+        .or_else(|| std::env::home_dir().map(|h| h.join(".gitignore")))
 }
@@ -146,27 +146,33 @@ async fn run_setup_command(json: bool) {
                    .expect("setup report has no non-serializable values");
                println!("{out}");
                if !report.configured {
-                    eprintln!(
-                        "Your team doesn't have a managed configuration yet. Ask a team admin to provision one."
+                    kigi_tui::best_effort_stderr::eprint_line(
+                        "Your team doesn't have a managed configuration yet. Ask a team admin to provision one.",
                    );
                }
            }
            Err(e) => {
-                eprintln!("Couldn't fetch managed configuration. {e}");
+                kigi_tui::best_effort_stderr::eprint_line(&format!(
+                    "Couldn't fetch managed configuration. {e}"
+                ));
                std::process::exit(1);
            }
        }
        return;
    }
    match managed_config::run_setup().await {
-        SetupOutcome::Installed => eprintln!("Applied managed configuration."),
+        SetupOutcome::Installed => {
+            kigi_tui::best_effort_stderr::eprint_line("Applied managed configuration.")
+        }
        SetupOutcome::NothingConfigured => {
-            eprintln!(
-                "Your team doesn't have a managed configuration yet. Ask a team admin to provision one."
+            kigi_tui::best_effort_stderr::eprint_line(
+                "Your team doesn't have a managed configuration yet. Ask a team admin to provision one.",
            );
        }
        SetupOutcome::Failed(e) => {
-            eprintln!("Couldn't apply managed configuration. {e}");
+            kigi_tui::best_effort_stderr::eprint_line(&format!(
+                "Couldn't apply managed configuration. {e}"
+            ));
            std::process::exit(1);
        }
    }
@@ -208,8 +214,8 @@ async fn run_leader_mgmt(args: LeaderMgmtArgs) -> Result<()> {
                println!("{info:#?}");
            } else {
                print_leader_descriptor(&descriptor);
-                eprintln!(
-                    "  (detailed info unavailable — leader does not advertise control capabilities)"
+                kigi_tui::best_effort_stderr::eprint_line(
+                    "  (detailed info unavailable — leader does not advertise control capabilities)",
                );
            }
            client.cancel();
@@ -220,7 +226,7 @@ async fn run_leader_mgmt(args: LeaderMgmtArgs) -> Result<()> {
 async fn kill_leaders() -> Result<()> {
    let leaders = kigi_shell::leader::discover_leaders().await;
    if leaders.is_empty() {
-        eprintln!("No leader candidates found.");
+        kigi_tui::best_effort_stderr::eprint_line("No leader candidates found.");
        return Ok(());
    }
    let mut killed = 0u32;
@@ -231,7 +237,9 @@ async fn kill_leaders() -> Result<()> {
        };
        if !kigi_shell::util::is_kigi_process(pid) {
            if let Some(ref lock) = d.lock_path {
-                eprintln!("  PID {pid} is not a kigi process, removing stale lock");
+                kigi_tui::best_effort_stderr::eprint_line(&format!(
+                    "  PID {pid} is not a kigi process, removing stale lock"
+                ));
                let _ = std::fs::remove_file(lock);
                cleaned += 1;
            }
@@ -240,19 +248,23 @@ async fn kill_leaders() -> Result<()> {
            }
            continue;
        }
-        eprintln!("  Killing leader PID {pid}");
+        kigi_tui::best_effort_stderr::eprint_line(&format!("  Killing leader PID {pid}"));
        if let Err(e) = kigi_shell::util::kill_process_by_pid(pid) {
-            eprintln!("  warning: failed to terminate PID {pid}: {e}");
+            kigi_tui::best_effort_stderr::eprint_line(&format!(
+                "  warning: failed to terminate PID {pid}: {e}"
+            ));
            continue;
        }
        killed += 1;
    }
    if killed > 0 {
-        eprintln!("Killed {killed} leader process(es).");
+        kigi_tui::best_effort_stderr::eprint_line(&format!("Killed {killed} leader process(es)."));
    } else if cleaned > 0 {
-        eprintln!("No live leader processes found (cleaned up {cleaned} stale lock(s)).");
+        kigi_tui::best_effort_stderr::eprint_line(&format!(
+            "No live leader processes found (cleaned up {cleaned} stale lock(s))."
+        ));
    } else {
-        eprintln!("No live leader processes found.");
+        kigi_tui::best_effort_stderr::eprint_line("No live leader processes found.");
    }
    Ok(())
 }
@@ -295,7 +307,7 @@ fn print_leader_descriptor(d: &LeaderDescriptor) {
        .map(|p| p.display().to_string())
        .unwrap_or_else(|| "?".into());
    let state = format!("{:?}", d.classification);
-    eprintln!("  PID {pid} ({state}) -- {sock}");
+    kigi_tui::best_effort_stderr::eprint_line(&format!("  PID {pid} ({state}) -- {sock}"));
 }
 fn leader_descriptor_json(d: &LeaderDescriptor) -> serde_json::Value {
    serde_json::json!(
@@ -1221,15 +1233,34 @@ fn main() {
            "Found crashed sessions from a previous run"
        );
    }
-    let runtime = tokio::runtime::Builder::new_multi_thread()
+    // Capped workers: tokio's per-core default pins 100+ thread slots on
+    // many-core shared hosts and later spawns die with EAGAIN.
+    let runtime = match tokio::runtime::Builder::new_multi_thread()
+        .worker_threads(kigi_tty_utils::runtime::capped_worker_threads().get())
        .enable_all()
        .build()
-        .unwrap_or_else(|e| panic!("failed to start tokio runtime: {e}"));
+    {
+        Ok(runtime) => runtime,
+        Err(e) => {
+            // A host out of threads (RLIMIT_NPROC / pids.max) is an
+            // environment problem, not a bug — report it, don't panic.
+            kigi_tty_utils::restore_native_stderr();
+            kigi_tui::best_effort_stderr::eprint_line(&format!(
+                "Error: failed to start the async runtime: {e}"
+            ));
+            kigi_tui::best_effort_stderr::eprint_line(
+                "The host may be out of threads (RLIMIT_NPROC or the cgroup pids ceiling); \
+                 close other processes or raise the limit, then retry.",
+            );
+            std::process::exit(1);
+        }
+    };
    let result = run_and_shutdown(runtime, async_main(), RUNTIME_SHUTDOWN_GRACE);
    kigi_log::debug_log::flush();
    if let Err(e) = result {
        kigi_tty_utils::restore_native_stderr();
-        eprintln!("Error: {e:#}");
+        // fd 2 may be a closed pane: one attempt, never a panic under panic = "abort".
+        kigi_tui::best_effort_stderr::eprint_line(&format!("Error: {e:#}"));
        std::process::exit(1);
    }
 }
@@ -1580,16 +1611,25 @@ async fn async_main() -> Result<()> {
    let result = kigi_tui::app::run(args, bg_update_rx).await;
    kigi_sandbox::flush();
    match result {
-        Ok(true) => {
+        Ok(exit) if exit.quit_for_update => {
            let adopted = bg_update_wait.lock().await.take();
-            if finish_update_on_exit(adopted, &update_config).await {
-                eprintln!("Update installed. Run `kigi` to start.");
-            } else {
-                eprintln!("Update did not complete. Run `kigi update` to retry.");
+            let installed =
+                finish_update_on_exit(adopted, &update_config, exit.terminal_reading).await;
+            // A terminal that stopped reading at teardown would block these prints.
+            if exit.terminal_reading {
+                if installed {
+                    kigi_tui::best_effort_stderr::eprint_line(
+                        "Update installed. Run `kigi` to start.",
+                    );
+                } else {
+                    kigi_tui::best_effort_stderr::eprint_line(
+                        "Update did not complete. Run `kigi update` to retry.",
+                    );
+                }
            }
            Ok(())
        }
-        Ok(false) => Ok(()),
+        Ok(_) => Ok(()),
        Err(e) => Err(e),
    }
 }
@@ -1605,10 +1645,15 @@ async fn async_main() -> Result<()> {
 async fn finish_update_on_exit(
    adopted: Option<tokio::task::JoinHandle<std::io::Result<std::process::ExitStatus>>>,
    update_config: &UpdateConfig,
+    terminal_reading: bool,
 ) -> bool {
    let run_blocking = |reason: Option<String>| async move {
+        // The blocking updater prints and its child inherits stderr: not on a stalled terminal.
+        if !terminal_reading {
+            return false;
+        }
        if let Some(reason) = reason {
-            eprintln!("{reason}");
+            kigi_tui::best_effort_stderr::eprint_line(&reason);
        }
        auto_update::run_update_if_available(
            auto_update::UpdateRunMode::Blocking,
@@ -1620,7 +1665,11 @@ async fn finish_update_on_exit(
    };
    match adopted {
        Some(handle) => {
-            eprintln!("Waiting for the update download to finish...");
+            if terminal_reading {
+                kigi_tui::best_effort_stderr::eprint_line(
+                    "Waiting for the update download to finish...",
+                );
+            }
            match handle.await {
                Ok(Ok(status)) if status.success() => true,
                Ok(Ok(status)) => {
@@ -1793,7 +1842,9 @@ async fn signal_leaders_to_relaunch(installed_version: &str) {
                to_version,
                ..
            })) => {
-                eprintln!("  ↻ Relaunching shared session (leader {from_version} → {to_version})…");
+                kigi_tui::best_effort_stderr::eprint_line(&format!(
+                    "  ↻ Relaunching shared session (leader {from_version} → {to_version})…"
+                ));
            }
            Ok(Ok(kigi_shell::leader::ControlPayload::RelaunchDeclined { reason })) => {
                tracing::debug!(% reason, "Leader declined relaunch");
@@ -390,14 +390,24 @@ impl ChatStateActor {
        });
    }

-    /// Replace the entire conversation, persist, re-estimate `total_tokens`,
-    /// and emit reset + token-update events.
-    ///
-    /// Compaction replaces carry the provider-side overhead forward as a
-    /// *ratio* (`base_estimate × provider_total ÷ estimate_at_last_response`,
-    /// capped at the pre-compaction total; `base_estimate` when that estimate is
-    /// 0) so the reseed neither springs back nor over-counts (see
-    /// `COMPACTION.md`).
+    /// Reseed after a rewrite: the raw estimate scaled by the provider-confirmed ratio, capped at the pre-rewrite total (see COMPACTION.md).
+    pub(super) fn reseed_total_tokens(&self, base_estimate: u64) -> u64 {
+        let pre_replace_total = self.state.total_tokens;
+        let mut estimated_tokens =
+            if pre_replace_total > 0 && self.state.estimate_at_last_response > 0 {
+                let ratio = pre_replace_total as f64 / self.state.estimate_at_last_response as f64;
+                (base_estimate as f64 * ratio).round() as u64
+            } else {
+                base_estimate
+            };
+        // Over-counting fires premature compaction; an under-count self-heals on the next response.
+        if pre_replace_total > 0 {
+            estimated_tokens = estimated_tokens.min(pre_replace_total);
+        }
+        estimated_tokens
+    }
+
+    /// Replace the conversation, persist, reseed `total_tokens` and emit reset + token events; `is_compaction` only marks the turn capture.
    pub(super) fn replace_conversation(
        &mut self,
        items: Vec<ConversationItem>,
@@ -407,28 +417,16 @@ impl ChatStateActor {
        if is_compaction && let Some(cap) = &mut self.state.turn_capture {
            cap.compaction_occurred = true;
        }
-        let pre_replace_total = self.state.total_tokens;
        // `harness_trace_buffer` / `harness_trace_turns` intentionally untouched:
        // the planner/verifier subagents ran, so their sealed trace turns survive
        // a conversation replace (same intent as the `TruncateToPromptIndex` arm).
        self.persistence.replace_history(&items);
        let base_estimate = super::state::estimate_conversation_tokens(&items);
-        let mut estimated_tokens =
-            if is_compaction && pre_replace_total > 0 && self.state.estimate_at_last_response > 0 {
-                let ratio = pre_replace_total as f64 / self.state.estimate_at_last_response as f64;
-                (base_estimate as f64 * ratio).round() as u64
-            } else {
-                base_estimate
-            };
-        // Compaction must never appear to increase usage.
-        if is_compaction && pre_replace_total > 0 {
-            estimated_tokens = estimated_tokens.min(pre_replace_total);
-        }
+        let estimated_tokens = self.reseed_total_tokens(base_estimate);
        self.state.conversation = items;
        self.state.estimated_tokens_since_model = 0;
        self.state.total_tokens = estimated_tokens;
-        self.state.estimate_at_last_response =
-            super::state::estimate_conversation_tokens(&self.state.conversation);
+        self.state.estimate_at_last_response = base_estimate;
        self.rebase_turn_capture_offset();
        self.send_event(ChatStateEvent::ConversationReset {
            new_len: self.state.conversation.len(),
@@ -68,10 +68,10 @@ impl ChatStateActor {
        self.state.conversation.truncate(truncate_at);
        self.state.prompt_texts.truncate(target_prompt_index);
        self.state.prompt_index = target_prompt_index;
-        self.state.total_tokens =
-            super::state::estimate_conversation_tokens(&self.state.conversation);
+        let base_estimate = super::state::estimate_conversation_tokens(&self.state.conversation);
+        self.state.total_tokens = self.reseed_total_tokens(base_estimate);
        self.state.estimated_tokens_since_model = 0;
-        self.state.estimate_at_last_response = self.state.total_tokens;
+        self.state.estimate_at_last_response = base_estimate;

        self.persistence.replace_history(&self.state.conversation);

@@ -68,12 +68,10 @@ pub fn estimate_item_tokens(item: &ConversationItem) -> u64 {
            kigi_token_estimation::estimate_tokens(&b.text_summary())
        }
        ConversationItem::Reasoning(r) => {
-            // Summary + content text follow the standard bytes-per-token
-            // estimate; encrypted blobs are base64 and don't survive
-            // tokenization 1:1, so estimate at len/4 as well.
+            // Text and ciphertext are the same reasoning twice: the larger wins (base64 is 4/3 over raw).
            let text_bytes = kigi_sampling_types::reasoning_item_text(r).len();
            let enc_bytes = r.encrypted_content.as_deref().map(str::len).unwrap_or(0);
-            ((text_bytes + enc_bytes) as u64) / kigi_token_estimation::BYTES_PER_TOKEN
+            (text_bytes.max(enc_bytes * 3 / 4) as u64) / kigi_token_estimation::BYTES_PER_TOKEN
        }
    }
 }
@@ -658,22 +658,66 @@ async fn compaction_reseed_without_provider_count_matches_plain_estimate() {
 }

 #[tokio::test]
-async fn non_compaction_replace_does_not_carry_overhead() {
+async fn non_compaction_replace_carries_confirmed_total() {
+    // A rewind/mode-switch replace scales from the provider-confirmed count, not the raw estimate.
    let h = TestHarness::new();
    h.handle
        .push_user_message(ConversationItem::user("x".repeat(4000)));
-    h.handle.record_token_usage(51_000);
+    h.handle.record_token_usage(500);
+    // Estimate at last response = 1_000, confirmed = 500 → ratio 0.5.

    h.handle
        .replace_conversation(vec![ConversationItem::user("q".repeat(4000))]);

    let total = h.handle.get_total_tokens().await;
    assert_eq!(
-        total, 1_000,
-        "non-compaction replace (e.g. rewind) keeps the plain estimate"
+        total, 500,
+        "same-size replace carries the provider-confirmed count"
    );
 }

+#[tokio::test]
+async fn replace_never_increases_total_tokens() {
+    // A growing replace is capped at the confirmed total; the under-count self-heals on the next usage.
+    let h = TestHarness::new();
+    h.handle
+        .push_user_message(ConversationItem::user("x".repeat(4000)));
+    h.handle.record_token_usage(1_500);
+
+    h.handle
+        .replace_conversation(vec![ConversationItem::user("q".repeat(40_000))]);
+
+    assert_eq!(h.handle.get_total_tokens().await, 1_500);
+}
+
+#[tokio::test]
+async fn truncate_scales_from_confirmed_total() {
+    // Rewind path: TruncateToPromptIndex must use the same carry as replace_conversation.
+    let h = TestHarness::new();
+    h.handle
+        .push_user_message(ConversationItem::user("x".repeat(4000)));
+    h.handle.increment_prompt_index();
+    h.handle
+        .push_user_message(ConversationItem::user("y".repeat(4000)));
+    h.handle.increment_prompt_index();
+    h.handle.record_token_usage(1_000);
+    // Estimate at last response = 2_000, confirmed = 1_000 → ratio 0.5.
+
+    h.handle.truncate_to_prompt_index(1).await;
+
+    // Keeps the first user item (raw estimate 1_000) → scaled to 500.
+    assert_eq!(h.handle.get_total_tokens().await, 500);
+}
+
+#[test]
+fn reasoning_estimate_takes_the_larger_of_text_and_ciphertext() {
+    // "thinking for r1" is 15 bytes; 800 base64 bytes are ~600 raw bytes.
+    let encrypted = reasoning_sibling("r1", Some(&"A".repeat(800)));
+    assert_eq!(crate::estimate_conversation_tokens(&[encrypted]), 150);
+    let plain = reasoning_sibling("r1", None);
+    assert_eq!(crate::estimate_conversation_tokens(&[plain]), 3);
+}
+
 #[tokio::test]
 async fn flush_calls_persistence_flush() {
    let mut h = TestHarness::new();
@@ -668,6 +668,11 @@ pub struct RemoteSettings {
    /// remote settings verbatim-input flag; env (`KIGI_COMPACTION_VERBATIM_INPUT`) and config override it. `None` = default (true).
    #[serde(default)]
    pub compaction_verbatim_input: Option<bool>,
+    /// Compaction summarizer `tool_choice` (`auto` | `none`) from remote
+    /// settings. Env (`KIGI_COMPACTION_TOOL_CHOICE`) and config override it.
+    /// Some backends reject `tool_choice: "none"` with tools attached.
+    #[serde(default)]
+    pub compaction_tool_choice: Option<String>,
    /// Master switch for jemalloc heap sampling + threshold dumps.
    /// `Some(true)` enables, `Some(false)` kill-switch, `None` = client default off.
    #[serde(default)]
@@ -78,11 +78,10 @@ pub fn detect_windows_shell() -> &'static WindowsShell {
            kigi_tty_utils::detach_std_command(&mut cmd);
            cmd.arg("pwsh.exe").stdin(std::process::Stdio::null());
            cmd.output()
-        } {
-            if output.status.success() {
-                tracing::info!("Windows shell: pwsh");
-                return WindowsShell::Pwsh;
-            }
+        } && output.status.success()
+        {
+            tracing::info!("Windows shell: pwsh");
+            return WindowsShell::Pwsh;
        }

        // powershell.exe (Windows PowerShell 5.1).
@@ -131,14 +130,13 @@ fn find_git_bash() -> Option<String> {
        kigi_tty_utils::detach_std_command(&mut cmd);
        cmd.arg("bash.exe").stdin(std::process::Stdio::null());
        cmd.output()
-    } {
-        if output.status.success() {
-            let stdout = String::from_utf8_lossy(&output.stdout);
-            for line in stdout.lines() {
-                let line = line.trim();
-                if line.to_ascii_lowercase().contains("git") {
-                    return Some(line.to_string());
-                }
+    } && output.status.success()
+    {
+        let stdout = String::from_utf8_lossy(&output.stdout);
+        for line in stdout.lines() {
+            let line = line.trim();
+            if line.to_ascii_lowercase().contains("git") {
+                return Some(line.to_string());
            }
        }
    }
@@ -616,7 +616,7 @@ mod win {
            }

            let signal = exception_to_signal(exception_code);
-            let si_code = exception_code as i32;
+            let si_code = exception_code;

            // ExceptionInformation[1] holds the faulting address for ACCESS_VIOLATION.
            let si_addr = if exception_code == EXCEPTION_ACCESS_VIOLATION
@@ -21,7 +21,7 @@ pub struct KigiEndpoints {
 pub const PRODUCTION_ENDPOINTS: KigiEndpoints = KigiEndpoints {
    coding_api_base_url: "https://api.kimi.com/coding/v1",
    oauth_host: "https://auth.kimi.com",
-    update_base_url: "https://git.zacharyzhang.com/api/v1/repos/ZacharyZhang-NY/Kigi-CLI/releases",
+    update_base_url: "https://api.github.com/repos/ZacharyZhang-NY/Kigi-CLI/releases",
    upgrade_page_url: "https://www.kimi.com/code/",
 };

@@ -76,6 +76,7 @@ mod tests {
        assert_eq!(std::fs::read(&dest).unwrap(), data);
    }

+    #[cfg(unix)]
    #[test]
    fn test_clone_preserves_permissions() {
        use std::os::unix::fs::PermissionsExt;
@@ -341,12 +341,14 @@ pub fn resolve_kigi_home() -> Result<PathBuf> {
    if let Ok(v) = std::env::var("KIGI_SHARE_DIR") {
        return Ok(PathBuf::from(v));
    }
-    let home =
-        PathBuf::from(std::env::var("HOME").context("neither $KIGI_SHARE_DIR nor $HOME is set")?);
-    // Canonicalize the home dir so worktree paths share the same physical .kigi
-    // tree as trust/hooks even when it is symlinked. The dunce canonicalization
-    // must stay in sync with kigi_config::default_kigi_home();
-    // home resolution deliberately differs ($HOME here vs std::env::home_dir()).
+    // Standard home-dir lookup (`USERPROFILE`/Known Folders on Windows, where
+    // `HOME` is usually unset outside Git Bash) — the same resolution as
+    // kigi_config::default_kigi_home(), so worktree paths land in the same
+    // physical .kigi tree as trust/hooks.
+    let home = std::env::home_dir()
+        .context("neither $KIGI_SHARE_DIR nor a home directory could be resolved")?;
+    // Canonicalize so a symlinked home still maps to one physical tree; the
+    // dunce canonicalization must stay in sync with default_kigi_home().
    Ok(dunce::canonicalize(&home).unwrap_or(home).join(".kigi"))
 }

@@ -37,6 +37,7 @@ pub(crate) fn execute_plan(plan: WorktreePlan) -> Result<CreateWorktreeResult> {

 #[cfg(test)]
 mod tests {
+    #[cfg(unix)]
    use super::*;
    use crate::{IgnoredFilesMode, WorkingTreeMode, WorktreeBuilder};
    use kigi_test_utils::git::{git_commit_all, init_git_repo};
@@ -89,10 +89,10 @@ fn is_platform_system_dir(cwd: &Path) -> bool {

 #[cfg(target_os = "windows")]
 fn is_platform_system_dir(cwd: &Path) -> bool {
-    if let Ok(temp) = std::env::var("TEMP").or_else(|_| std::env::var("TMP")) {
-        if cwd.starts_with(&temp) {
-            return true;
-        }
+    if let Ok(temp) = std::env::var("TEMP").or_else(|_| std::env::var("TMP"))
+        && cwd.starts_with(&temp)
+    {
+        return true;
    }

    let path_lower = cwd.to_string_lossy().to_lowercase();
@@ -103,7 +103,7 @@ fn is_platform_system_dir(cwd: &Path) -> bool {
        return true;
    }

-    if cwd.parent().map_or(false, |p| p.parent().is_none()) && cwd.to_string_lossy().len() <= 3 {
+    if cwd.parent().is_some_and(|p| p.parent().is_none()) && cwd.to_string_lossy().len() <= 3 {
        return true;
    }

@@ -971,7 +971,7 @@ pub(crate) fn start_with_timeout(

    let progress_for_thread = progress.clone();

-    let thread = std::thread::spawn(move || {
+    let watcher_loop = move || {
        let update_stage = |stage: &'static str| {
            if let Ok(mut p) = progress_for_thread.lock() {
                p.set_stage(stage);
@@ -1351,7 +1351,13 @@ pub(crate) fn start_with_timeout(
                let _ = ready_tx.send(Err(Box::new(e)));
            }
        }
-    });
+    };
+    // Recoverable spawn: a host out of threads must surface as an error the
+    // caller can degrade on, not a `std::thread::spawn` panic.
+    let thread = std::thread::Builder::new()
+        .name("fsnotify-watcher".into())
+        .spawn(watcher_loop)
+        .map_err(|e| crate::FsNotifyError::WatcherStart(Box::new(e)))?;

    // Wait for watcher to be ready (with timeout)
    if let Ok(mut p) = progress.lock() {
@@ -318,6 +318,7 @@ mod tests {
        assert!(matches!(state, FileContentState::TooLarge { .. }));
    }

+    #[cfg(unix)]
    #[tokio::test]
    async fn test_read_file_bounded_symlink() {
        let dir = tempfile::tempdir().unwrap();
@@ -150,6 +150,7 @@ fn write_private(path: &Path, contents: &str) -> std::io::Result<()> {
 #[cfg(test)]
 mod tests {
    use super::*;
+    #[cfg(unix)]
    use std::time::Instant;

    /// Write an executable `#!/bin/sh` fake `mmdc` and return (dir-guard, path).
@@ -179,9 +179,12 @@ fn reap_process_group(_child: &Child) {
 #[cfg(test)]
 mod tests {
    use super::*;
+    #[cfg(unix)]
    use std::process::Stdio;
+    #[cfg(unix)]
    use std::time::Instant;

+    #[cfg(unix)]
    fn detached(mut cmd: Command) -> Command {
        cmd.stdin(Stdio::null())
            .stdout(Stdio::null())
@@ -214,6 +214,7 @@ pub fn open_url_if_safe(url: &str, filter: SchemeFilter) -> bool {
 mod tests {
    use super::*;

+    #[cfg(not(target_os = "windows"))]
    #[test]
    fn open_path_command_passes_path_as_a_single_arg() {
        // Path with spaces must be one argument, never shell-interpolated.
@@ -146,6 +146,10 @@ impl TermWriter {
    pub fn writer_sync(&self) -> &WriterSync {
        &self.sync
    }
+    /// A cloneable handle onto this writer's queue for out-of-band escapes.
+    pub fn escape_writer(&self) -> EscapeWriter {
+        EscapeWriter::new(self.tx.clone(), self.sync.clone())
+    }
 }
 impl Write for TermWriter {
    fn write(&mut self, data: &[u8]) -> std::io::Result<usize> {
@@ -166,6 +170,56 @@ impl Drop for TermWriter {
        let _ = self.flush();
    }
 }
+/// Out-of-band escapes ride the frame queue (an inline stderr write deadlocks on a stalled tty); drop every clone before `join_within`.
+#[derive(Clone)]
+pub struct EscapeWriter {
+    tx: mpsc::Sender<Vec<u8>>,
+    sync: WriterSync,
+}
+impl EscapeWriter {
+    pub fn new(tx: mpsc::Sender<Vec<u8>>, sync: WriterSync) -> Self {
+        Self { tx, sync }
+    }
+    /// No writer thread behind it: sends are dropped. For tests and views built before the terminal exists.
+    pub fn disconnected() -> Self {
+        let (tx, _rx) = mpsc::channel::<Vec<u8>>();
+        Self {
+            tx,
+            sync: WriterSync::new(),
+        }
+    }
+    /// Never blocks; covered by [`WriterSync::wait_drained`].
+    pub fn emit(&self, bytes: impl Into<Vec<u8>>) {
+        let data: Vec<u8> = bytes.into();
+        if data.is_empty() {
+            return;
+        }
+        self.sync.mark_queued();
+        let _ = self.tx.send(data);
+    }
+    /// Winapi-only commands run synchronously: a console API call, not a tty write.
+    pub fn emit_command(&self, command: impl crossterm::Command) {
+        #[cfg(windows)]
+        if !command.is_ansi_code_supported() {
+            let _ = command.execute_winapi();
+            return;
+        }
+        let mut ansi = String::new();
+        if command.write_ansi(&mut ansi).is_ok() {
+            self.emit(ansi);
+        }
+    }
+}
+/// Outcome of [`WriterThread::join_within`].
+#[derive(Debug, Clone, Copy, PartialEq, Eq)]
+pub enum WriterJoin {
+    /// The thread drained its queue and exited.
+    Joined,
+    /// Still running at the deadline (tty blocked, or a sender alive); detached.
+    TimedOut,
+}
+/// Teardown waits this long for the writer thread before detaching it, so a stalled tty cannot hang `/quit`.
+pub const WRITER_JOIN_GRACE: Duration = Duration::from_secs(2);
 /// Handle for the background writer thread.
 ///
 /// Joining ensures all queued frames have been written to the terminal
@@ -174,19 +228,45 @@ pub struct WriterThread {
    handle: Option<std::thread::JoinHandle<()>>,
 }
 impl WriterThread {
-    /// Block until the writer thread has processed all pending frames and
-    /// exited. The [`mpsc::Sender`] must be dropped *before* calling this,
-    /// otherwise the thread will never see the channel close.
-    pub fn join(mut self) {
-        if let Some(h) = self.handle.take() {
-            let _ = h.join();
+    /// Wait up to `grace` for the thread to drain and exit, else detach it; every sender (`EscapeWriter` clones too) must be dropped first.
+    pub fn join_within(mut self, grace: Duration) -> WriterJoin {
+        let Some(handle) = self.handle.take() else {
+            return WriterJoin::Joined;
+        };
+        if wait_finished(&handle, grace) {
+            let _ = handle.join();
+            WriterJoin::Joined
+        } else {
+            tracing::warn!(
+                grace_ms = grace.as_millis() as u64,
+                "term-writer thread still running at teardown; detaching"
+            );
+            WriterJoin::TimedOut
+        }
+    }
+    #[cfg(test)]
+    fn for_test(handle: std::thread::JoinHandle<()>) -> Self {
+        Self {
+            handle: Some(handle),
        }
    }
 }
+fn wait_finished(handle: &std::thread::JoinHandle<()>, grace: Duration) -> bool {
+    let deadline = Instant::now() + grace;
+    while !handle.is_finished() {
+        if Instant::now() >= deadline {
+            return false;
+        }
+        std::thread::sleep(Duration::from_millis(1));
+    }
+    true
+}
 impl Drop for WriterThread {
    fn drop(&mut self) {
-        if let Some(h) = self.handle.take() {
-            let _ = h.join();
+        if let Some(handle) = self.handle.take()
+            && wait_finished(&handle, WRITER_JOIN_GRACE)
+        {
+            let _ = handle.join();
        }
    }
 }
@@ -197,7 +277,7 @@ impl Drop for WriterThread {
 /// `BufWriter`. The [`WriterSync`] must be shared with every [`TermWriter`]
 /// built on the sender so [`WriterSync::wait_drained`] tracks the queue.
 /// Drop the sender to signal the thread to exit, then call
-/// [`WriterThread::join`] to wait for it.
+/// [`WriterThread::join_within`] to wait for it.
 pub fn spawn_writer_thread() -> (mpsc::Sender<Vec<u8>>, WriterSync, WriterThread) {
    let (tx, rx) = mpsc::channel::<Vec<u8>>();
    let sync = WriterSync::new();
@@ -570,4 +650,47 @@ mod tests {
            "Round-tripped bytes do not decode to original UTF-8 string"
        );
    }
+    /// Frames and escapes share one queue, so the writer sees them in call order.
+    #[test]
+    #[cfg(not(windows))]
+    fn escapes_share_the_frame_queue_in_call_order() {
+        let (tx, rx) = mpsc::channel::<Vec<u8>>();
+        let sync = WriterSync::new();
+        let mut frames = TermWriter::new(tx, sync.clone());
+        let escapes = frames.escape_writer();
+        frames.write_all(b"frame-1").unwrap();
+        frames.flush().unwrap();
+        escapes.emit("\x1b[?1000h");
+        frames.write_all(b"frame-2").unwrap();
+        frames.flush().unwrap();
+        escapes.emit_command(crossterm::event::DisableMouseCapture);
+        let payloads: Vec<Vec<u8>> = std::iter::from_fn(|| rx.try_recv().ok()).collect();
+        assert_eq!(payloads.len(), 4);
+        assert_eq!(payloads[0], b"frame-1");
+        assert_eq!(payloads[1], b"\x1b[?1000h");
+        assert_eq!(payloads[2], b"frame-2");
+        assert!(payloads[3].starts_with(b"\x1b["), "{:?}", payloads[3]);
+        assert!(!sync.is_drained(), "four payloads queued, none written");
+    }
+    /// A writer thread parked in a blocked tty write must not hang teardown.
+    #[test]
+    fn join_within_times_out_on_a_stuck_writer_thread() {
+        let (release_tx, release_rx) = mpsc::channel::<()>();
+        let handle = std::thread::spawn(move || {
+            let _ = release_rx.recv();
+        });
+        let started = Instant::now();
+        let outcome = WriterThread::for_test(handle).join_within(Duration::from_millis(50));
+        assert_eq!(outcome, WriterJoin::TimedOut);
+        assert!(started.elapsed() < Duration::from_secs(2));
+        let _ = release_tx.send(());
+    }
+    #[test]
+    fn join_within_joins_a_finished_writer_thread() {
+        let handle = std::thread::spawn(|| {});
+        assert_eq!(
+            WriterThread::for_test(handle).join_within(Duration::from_secs(5)),
+            WriterJoin::Joined
+        );
+    }
 }
@@ -10,6 +10,7 @@
 //!   portable re-injection exists (TIOCSTI is blocked); accepted loss.

 use std::io::Write;
+#[cfg(unix)]
 use std::time::Duration;

 /// Bounds the reply buffer against terminals that stream without a terminator.
@@ -18,6 +18,7 @@ eventsource-stream = { workspace = true }
 futures-util = { workspace = true }
 indexmap = { workspace = true, features = ["serde"] }
 reqwest = { workspace = true, features = ["stream"] }
+kigi-secrets = { path = "../kigi-secrets" }
 serde = { workspace = true, features = ["derive"] }
 serde_json = { workspace = true }
 tokio = { workspace = true, features = ["rt", "macros", "time", "sync"] }
@@ -193,6 +193,16 @@ fn extract_context_total(value: &serde_json::Value) -> Option<u32> {
    Some(i.saturating_add(o))
 }

+/// The endpoint as the HTTP client parses it, minus userinfo, sensitive query values and fragment; an unparseable one is not echoed.
+fn redacted_endpoint(endpoint: &str) -> String {
+    match reqwest::Url::parse(endpoint) {
+        Ok(mut url) => {
+            kigi_secrets::redact_url(&mut url);
+            url.to_string()
+        }
+        Err(_) => "<unparseable endpoint>".to_owned(),
+    }
+}
 /// Record `success=false` + `error` on the active inference span when a stream
 /// request fails before any response (transport/connect/TLS errors). Without
 /// this the `#[instrument]` span closes with both fields Empty, so an outage
@@ -806,7 +816,7 @@ impl SamplingClient {
        let server_message_lower = server_message.to_lowercase();

        let mut context_parts = vec![server_message.to_string()];
-        context_parts.push(format!("\nRequest URL: {}", endpoint));
+        context_parts.push(format!("\nRequest URL: {}", redacted_endpoint(endpoint)));

        // Show headers if error mentions headers
        if server_message_lower.contains("header") {
@@ -1006,7 +1016,8 @@ impl SamplingClient {
                let endpoint = self.endpoint("chat/completions");
                let server_message = response.text().await.unwrap_or_default();
                return Err(SamplingError::Auth(format!(
-                    "Unauthorized (401) from {endpoint}: {server_message}"
+                    "Unauthorized (401) from {}: {server_message}",
+                    redacted_endpoint(&endpoint)
                )));
            }

@@ -1196,7 +1207,8 @@ impl SamplingClient {
                let endpoint = self.endpoint("responses");
                let server_message = parse_error_bytes(bytes.as_ref());
                return Err(SamplingError::Auth(format!(
-                    "Unauthorized (401) from {endpoint}: {server_message}"
+                    "Unauthorized (401) from {}: {server_message}",
+                    redacted_endpoint(&endpoint)
                )));
            }

@@ -1363,7 +1375,8 @@ impl SamplingClient {
                let endpoint = self.endpoint("responses");
                let server_message = response.text().await.unwrap_or_default();
                return Err(SamplingError::Auth(format!(
-                    "Unauthorized (401) from {endpoint}: {server_message}"
+                    "Unauthorized (401) from {}: {server_message}",
+                    redacted_endpoint(&endpoint)
                )));
            }
            let model_metadata = extract_model_metadata(response.headers());
@@ -1542,7 +1555,8 @@ impl SamplingClient {
                let endpoint = self.endpoint("messages");
                let server_message = parse_error_bytes(bytes.as_ref());
                return Err(SamplingError::Auth(format!(
-                    "Unauthorized (401) from {endpoint}: {server_message}"
+                    "Unauthorized (401) from {}: {server_message}",
+                    redacted_endpoint(&endpoint)
                )));
            }

@@ -1657,7 +1671,8 @@ impl SamplingClient {
                let endpoint = self.endpoint("messages");
                let server_message = response.text().await.unwrap_or_default();
                return Err(SamplingError::Auth(format!(
-                    "Unauthorized (401) from {endpoint}: {server_message}"
+                    "Unauthorized (401) from {}: {server_message}",
+                    redacted_endpoint(&endpoint)
                )));
            }
            let model_metadata = extract_model_metadata(response.headers());
@@ -3048,3 +3063,31 @@ mod tests {
        ));
    }
 }
+
+#[cfg(test)]
+mod redacted_endpoint_tests {
+    /// Every spelling the client accepts (LF inside, one slash, no slash) loses its userinfo.
+    #[test]
+    fn redacted_endpoint_drops_userinfo_in_every_accepted_spelling() {
+        for spelling in [
+            "http://u:TEST_PASSWORD@127.0.0.1:1/v1/chat/completions",
+            "http://u:TEST PASSWORD@127.0.0.1:1/v1/chat/completions",
+            "http://u:TEST_PASSWORD\n@127.0.0.1:1/v1/chat/completions",
+            "http://u:TEST_PASSWORD\r\n@127.0.0.1:1/v1/chat/completions",
+            "http:/u:TEST_PASSWORD@127.0.0.1:1/v1/chat/completions",
+            "http:u:TEST_PASSWORD@127.0.0.1:1/v1/chat/completions",
+            "http:\\u:TEST_PASSWORD@127.0.0.1:1/v1/chat/completions",
+        ] {
+            let redacted = super::redacted_endpoint(spelling);
+            assert_eq!(
+                redacted, "http://127.0.0.1:1/v1/chat/completions",
+                "{spelling:?}"
+            );
+        }
+        let gateway = super::redacted_endpoint(
+            "http://127.0.0.1:1/?access_token=TEST_QUERY_TOKEN&route=x#frag",
+        );
+        assert!(!gateway.contains("TEST_QUERY_TOKEN"), "{gateway}");
+        assert!(gateway.contains("route=x"), "{gateway}");
+    }
+}
@@ -5,8 +5,11 @@
 //!
 //! # Retry behavior summary
 //!
-//! **Retried** (up to [`DEFAULT_MAX_RETRIES`] = 15, ~6 min with 30s backoff cap):
-//! - 500, 502, 503, 504, 520 (server errors)
+//! **Retried** (up to [`DEFAULT_MAX_RETRIES`] = 15, ~6 min with the
+//! [`MAX_RETRY_BACKOFF`] cap — a server `Retry-After` is clamped to it too):
+//! - Any 5xx except Cloudflare 525/526 — covers the Cloudflare edge pages
+//!   (520-524 origin unreachable/timed out, 530 edge 1xxx) and upstream
+//!   overload (529); see `is_retryable_api_status`
 //! - Connection errors (timeout, refused, reset)
 //! - `EventStreamError` / `StreamError` (mid-stream failures)
 //! - `EmptyResponse` (model returned no content/tool calls)
@@ -77,16 +80,35 @@ pub fn doom_loop_backoff(retry_count: u32) -> Duration {
    Duration::from_millis(hasher.finish() % 251)
 }

-/// Exponential backoff (2s, 4s, 8s, ..., capped 30s) with +/-20% jitter
-/// to prevent thundering-herd retry storms.
+/// Longest single wait on the generic retry path — the exponential-backoff
+/// ceiling, and the clamp for a server `Retry-After`. Cloudflare answers 52x
+/// with `Retry-After: 60`-`120`; honoring that verbatim across the retry
+/// budget would stall a turn ~28 min instead of the ~6 min budget. The 429
+/// path deliberately waits the full `Retry-After` instead, bounded by
+/// [`RATE_LIMIT_RETRY_THRESHOLD`] attempts (and by the parse-level 120s cap
+/// on the header).
+pub const MAX_RETRY_BACKOFF: Duration = Duration::from_secs(30);
+
+/// Exponential backoff (2s, 4s, 8s, ..., capped at [`MAX_RETRY_BACKOFF`])
+/// with +/-20% jitter to prevent thundering-herd retry storms.
 pub fn retry_backoff_with_jitter(retry_count: u32) -> Duration {
+    let shift = retry_count.saturating_sub(1);
+    let base_ms = 2000u64
+        .checked_shl(shift)
+        .unwrap_or(u64::MAX)
+        .min(MAX_RETRY_BACKOFF.as_millis() as u64);
+    jittered(Duration::from_millis(base_ms))
+}
+
+/// +/-20% jitter around `base`, de-syncing clients that failed at the
+/// same instant (e.g. a mass Cloudflare 52x event during an origin outage).
+fn jittered(base: Duration) -> Duration {
    use std::hash::{Hash, Hasher};
    use std::sync::atomic::{AtomicU64, Ordering};

    static JITTER_SEQ: AtomicU64 = AtomicU64::new(0);

-    let shift = retry_count.saturating_sub(1);
-    let base_ms = 2000u64.checked_shl(shift).unwrap_or(u64::MAX).min(30_000);
+    let base_ms = base.as_millis() as u64;
    let jitter_range = base_ms / 5;
    let mut hasher = std::hash::DefaultHasher::new();
    JITTER_SEQ.fetch_add(1, Ordering::Relaxed).hash(&mut hasher);
@@ -210,9 +232,11 @@ pub fn classify_error(
        if next_attempt >= max_retries {
            return RetryDecision::Fatal(clone_error(err));
        }
+        // A server `Retry-After` is clamped: Cloudflare 52x pages ask for
+        // 60-120s, which across the budget would stall a turn ~28 min.
        let backoff = err
            .retry_after()
-            .map(Duration::from_secs)
+            .map(|secs| jittered(Duration::from_secs(secs).min(MAX_RETRY_BACKOFF)))
            .unwrap_or_else(|| retry_backoff_with_jitter(next_attempt));
        if next_attempt == 1 {
            return RetryDecision::RetryWithClientRebuild { backoff };
@@ -595,6 +619,22 @@ mod tests {
        }
    }

+    #[test]
+    fn classify_5xx_retry_after_is_clamped_to_backoff_cap() {
+        // Cloudflare answers 52x with Retry-After: 60-120; honoring it
+        // verbatim across the retry budget would stall a turn ~28 min.
+        let err = api_err_with_retry_after(StatusCode::from_u16(522).unwrap(), 120);
+        match classify_error(&err, 1, 15, RATE_LIMIT_RETRY_THRESHOLD) {
+            RetryDecision::Retry { backoff } => {
+                assert!(
+                    backoff <= MAX_RETRY_BACKOFF.mul_f64(1.2),
+                    "52x Retry-After must clamp to ~MAX_RETRY_BACKOFF, got {backoff:?}"
+                );
+            }
+            other => panic!("expected Retry, got {other:?}"),
+        }
+    }
+
    #[test]
    fn classify_5xx_first_retry_rebuilds_client() {
        let err = api_err(StatusCode::INTERNAL_SERVER_ERROR, "boom");
@@ -55,6 +55,15 @@ enum BlockType {
    Thinking,
 }

+/// Executable after a `max_tokens` stop: every call's arguments parse or are empty (zero-arg convention).
+fn tool_calls_complete(tool_calls: &[ToolCall]) -> bool {
+    !tool_calls.is_empty()
+        && tool_calls.iter().all(|tc| {
+            tc.arguments.trim().is_empty()
+                || serde_json::from_str::<serde::de::IgnoredAny>(&tc.arguments).is_ok()
+        })
+}
+
 /// Transform a raw Anthropic Messages API stream into a stream of
 /// [`SamplingEvent`]s.
 ///
@@ -454,11 +463,18 @@ pub fn stream_messages<'a>(
        }

        if final_stop_reason == Some(StopReason::Length) {
-            yield SamplingEvent::Failed {
-                request_id: request_id.clone(),
-                error: SamplingErrorInfo::from(&SamplingError::MaxTokensTruncation),
-            };
-            return;
+            if !tool_calls_complete(&assistant_tool_calls) {
+                yield SamplingEvent::Failed {
+                    request_id: request_id.clone(),
+                    error: SamplingErrorInfo::from(&SamplingError::MaxTokensTruncation),
+                };
+                return;
+            }
+            tracing::info!(
+                request_id = %request_id,
+                tool_calls = assistant_tool_calls.len(),
+                "max_tokens stop after complete tool calls; completing the turn so they run"
+            );
        }

        // Build the final response
@@ -653,3 +653,90 @@ async fn pure_cache_hit_with_zero_uncached_still_emits_usage() {
    assert_eq!(usage.cached_prompt_tokens, 2500);
    assert_eq!(usage.total_tokens, 2501);
 }
+
+fn tool_use_start(index: u32, id: &str) -> MessageStreamEvent {
+    MessageStreamEvent::ContentBlockStart {
+        index,
+        content_block: ContentBlock::ToolUse {
+            id: id.into(),
+            name: "do_thing".into(),
+            input: serde_json::json!({}),
+        },
+    }
+}
+
+fn arg_delta(index: u32, partial_json: &str) -> MessageStreamEvent {
+    MessageStreamEvent::ContentBlockDelta {
+        index,
+        delta: StreamDelta::InputJsonDelta {
+            partial_json: partial_json.into(),
+        },
+    }
+}
+
+async fn max_tokens_after(events: Vec<MessageStreamEvent>) -> Vec<SamplingEvent> {
+    let mut all = vec![Ok(message_start())];
+    all.extend(events.into_iter().map(Ok));
+    all.push(Ok(message_delta_with_stop(messages::StopReason::MaxTokens)));
+    all.push(Ok(MessageStreamEvent::MessageStop));
+    let raw = stream::iter(all).boxed();
+    collect(stream_messages(raw, None, rid(), Duration::from_secs(60))).await
+}
+
+/// Complete tool calls before a max_tokens stop complete the turn as ToolCalls.
+#[tokio::test]
+async fn max_tokens_after_complete_tool_calls_completes_as_tool_calls() {
+    let evs = max_tokens_after(vec![
+        tool_use_start(0, "call_done"),
+        arg_delta(0, "{\"x\": 1}"),
+        block_stop(0),
+    ])
+    .await;
+    match evs.last().unwrap() {
+        SamplingEvent::Completed { response, .. } => {
+            assert_eq!(response.stop_reason, Some(StopReason::ToolCalls));
+            assert_eq!(response.tool_calls().len(), 1);
+            assert_eq!(response.tool_calls()[0].arguments.as_ref(), "{\"x\": 1}");
+        }
+        other => panic!("expected Completed(ToolCalls), got {other:?}"),
+    }
+}
+
+/// A tool_use block with no argument deltas is the zero-argument convention, not truncation.
+#[tokio::test]
+async fn max_tokens_after_tool_call_without_arguments_completes() {
+    let evs = max_tokens_after(vec![tool_use_start(0, "call_no_args"), block_stop(0)]).await;
+    match evs.last().unwrap() {
+        SamplingEvent::Completed { response, .. } => {
+            assert_eq!(response.stop_reason, Some(StopReason::ToolCalls));
+            assert_eq!(response.tool_calls()[0].arguments.as_ref(), "");
+        }
+        other => panic!("expected Completed(ToolCalls), got {other:?}"),
+    }
+}
+
+/// A call cut mid-arguments is truncation: the turn fails so the caller can compact.
+#[tokio::test]
+async fn max_tokens_with_truncated_tool_arguments_fails() {
+    let evs = max_tokens_after(vec![
+        tool_use_start(0, "call_ok"),
+        arg_delta(0, "{}"),
+        block_stop(0),
+        tool_use_start(1, "call_cut"),
+        arg_delta(1, "{\"x\": \"trunc"),
+        block_stop(1),
+    ])
+    .await;
+    assert!(
+        !evs.iter()
+            .any(|e| matches!(e, SamplingEvent::Completed { .. })),
+        "a cut tool call must not complete: {evs:?}"
+    );
+    match evs.last().unwrap() {
+        SamplingEvent::Failed { error, .. } => assert_eq!(
+            error.kind,
+            crate::events::SamplingErrorKind::MaxTokensTruncation
+        ),
+        other => panic!("expected Failed(MaxTokensTruncation), got {other:?}"),
+    }
+}
@@ -239,9 +239,7 @@ impl SamplingError {
            SamplingError::InvalidConfiguration(_) => false,
            SamplingError::Http(err) => is_retryable_reqwest(err),
            SamplingError::Serialization(_) => false,
-            SamplingError::Api { status, .. } => {
-                matches!(status.as_u16(), 429 | 500 | 502 | 503 | 504 | 520)
-            }
+            SamplingError::Api { status, .. } => is_retryable_api_status(status.as_u16()),
            SamplingError::EventStreamError(_) => true,
            SamplingError::StreamError { .. } => true,
            SamplingError::IdleTimeout { .. } => false,
@@ -342,6 +340,15 @@ pub fn try_parse_stream_error(data: &str) -> Option<SamplingError> {
    })
 }

+/// 429 and any 5xx except Cloudflare 525/526 (origin TLS handshake /
+/// invalid certificate — a broken origin certificate never clears on its
+/// own). The 5xx sweep covers the Cloudflare edge pages (520-524 origin
+/// unreachable/timed out, 530 edge 1xxx) and upstream overload (529),
+/// which an outage serves in place of the API's own statuses.
+pub fn is_retryable_api_status(status: u16) -> bool {
+    status == 429 || ((500..=599).contains(&status) && !matches!(status, 525 | 526))
+}
+
 /// True when an error message indicates a context-window overflow. Backends report
 /// this inconsistently with no stable error code, so we match the message text; it's
 /// deterministic (re-sending the same payload always fails), so callers must not retry.
@@ -352,6 +359,7 @@ pub fn is_context_length_error(message: &str) -> bool {
        || m.contains("maximum prompt length")
        || m.contains("maximum context length")
        || m.contains("context_length_exceeded")
+        || (m.contains("current message") && m.contains("exceeds budget"))
 }

 /// Decide whether a [`reqwest::Error`] is worth retrying.
@@ -378,6 +386,27 @@ pub fn is_retryable_reqwest(err: &reqwest::Error) -> bool {
 mod tests {
    use super::*;

+    #[test]
+    fn retryable_api_status_covers_edge_5xx_and_excludes_broken_tls() {
+        for status in [
+            429u16, 500, 502, 503, 504, 520, 521, 522, 523, 524, 529, 530, 599,
+        ] {
+            assert!(is_retryable_api_status(status), "should retry {status}");
+        }
+        // Cloudflare 525/526: origin TLS handshake / invalid certificate —
+        // a broken origin certificate never clears on its own.
+        for status in [525u16, 526, 400, 401, 403, 404, 408, 422, 200] {
+            assert!(!is_retryable_api_status(status), "must not retry {status}");
+        }
+        let api = SamplingError::Api {
+            status: StatusCode::from_u16(522).unwrap(),
+            message: "cloudflare: connection timed out".into(),
+            model_metadata: None,
+            retry_after_secs: None,
+        };
+        assert!(api.is_retryable(), "522 edge page must be retryable");
+    }
+
    #[test]
    fn context_length_error_matches_backend_variants() {
        for msg in [
@@ -387,6 +416,7 @@ mod tests {
            "This model's maximum context length is 200000 tokens",
            "invalid_request_error: prompt is too long: 300000 tokens > 200000 maximum",
            "error type: context_length_exceeded",
+            "the current message exceeds budget: 120000 < 250000",
        ] {
            assert!(is_context_length_error(msg), "should match: {msg}");
        }
@@ -51,6 +51,7 @@ enforce = []

 [dev-dependencies]
 serial_test = { workspace = true }
+tempfile = { workspace = true }
 # Compiles the hand-rolled macOS deny-glob regex in tests to assert it matches
 # EXACTLY the set globset (the Linux backend) matches — the cross-platform parity
 # guard for the glob dialect.
@@ -14,9 +14,19 @@
 //! cargo run -p kigi-sandbox --example sandbox_smoke_test -- read-only
 //! ```

+// Kernel enforcement is unix-only (Seatbelt / Landlock); Windows builds a
+// stub so `--all-targets` stays green everywhere.
+#[cfg(unix)]
 use kigi_sandbox::{ProfileName, SandboxManager};
+#[cfg(unix)]
 use std::path::Path;

+#[cfg(not(unix))]
+fn main() {
+    eprintln!("sandbox_smoke_test exercises kernel enforcement and only runs on Unix.");
+}
+
+#[cfg(unix)]
 fn main() {
    let profile_name = std::env::args()
        .nth(1)
@@ -121,6 +131,7 @@ fn main() {
    println!("\n✅ Smoke test complete");
 }

+#[cfg(unix)]
 fn test_read(label: &str, path: &Path) {
    if path.is_file() {
        match std::fs::read(path) {
@@ -150,6 +161,7 @@ fn test_read(label: &str, path: &Path) {
    }
 }

+#[cfg(unix)]
 fn test_write(label: &str, path: &Path) {
    match std::fs::write(path, b"sandbox-test") {
        Ok(()) => {
@@ -1,91 +1,130 @@
 //! Per-child seccomp network filter. No-op on non-Linux.

-/// Install seccomp BPF filter blocking network syscalls.
+#[cfg(target_os = "linux")]
+mod bpf {
+    use libc::sock_filter;
+
+    pub(super) const SECCOMP_RET_ALLOW: u32 = 0x7fff_0000;
+    pub(super) const SECCOMP_RET_ERRNO: u32 = 0x0005_0000;
+    pub(super) const EPERM_VAL: u32 = 1;
+    #[cfg(target_arch = "x86_64")]
+    pub(super) const X32_SYSCALL_BIT: u32 = 0x4000_0000;
+    /// `seccomp_data` field offsets.
+    pub(super) const OFF_NR: u32 = 0;
+    pub(super) const OFF_ARCH: u32 = 4;
+    #[cfg(target_arch = "x86_64")]
+    pub(super) const EXPECTED_ARCH: u32 = 0xc000_003e; // AUDIT_ARCH_X86_64
+    #[cfg(target_arch = "aarch64")]
+    pub(super) const EXPECTED_ARCH: u32 = 0xc000_00b7; // AUDIT_ARCH_AARCH64
+    #[cfg(not(any(target_arch = "x86_64", target_arch = "aarch64")))]
+    pub(super) const EXPECTED_ARCH: u32 = 0;
+
+    pub(super) fn stmt(code: u32, k: u32) -> sock_filter {
+        sock_filter {
+            code: code as u16,
+            jt: 0,
+            jf: 0,
+            k,
+        }
+    }
+
+    pub(super) fn jump(code: u32, k: u32, jt: u8, jf: u8) -> sock_filter {
+        sock_filter {
+            code: code as u16,
+            jt,
+            jf,
+            k,
+        }
+    }
+
+    /// Reject a foreign-arch (and, on x86_64, x32) syscall number before
+    /// matching `nr`, so a syscall reached through another table cannot
+    /// alias past the deny list.
+    pub(super) fn push_arch_nr_gate(f: &mut Vec<sock_filter>) {
+        use libc::{BPF_ABS, BPF_JEQ, BPF_JMP, BPF_K, BPF_LD, BPF_RET, BPF_W};
+        f.push(stmt(BPF_LD | BPF_W | BPF_ABS, OFF_ARCH));
+        f.push(jump(BPF_JMP | BPF_JEQ | BPF_K, EXPECTED_ARCH, 1, 0));
+        f.push(stmt(BPF_RET | BPF_K, SECCOMP_RET_ERRNO | EPERM_VAL));
+        f.push(stmt(BPF_LD | BPF_W | BPF_ABS, OFF_NR));
+        #[cfg(target_arch = "x86_64")]
+        {
+            f.push(jump(
+                BPF_JMP | libc::BPF_JSET | BPF_K,
+                X32_SYSCALL_BIT,
+                0,
+                1,
+            ));
+            f.push(stmt(BPF_RET | BPF_K, SECCOMP_RET_ERRNO | EPERM_VAL));
+        }
+    }
+}
+
+/// Syscalls that open or send on a socket, including the `io_uring` and
+/// `sendmmsg` paths that bypass the classic entry points.
+#[cfg(target_os = "linux")]
+fn child_network_blocked_syscalls() -> [u32; 11] {
+    use libc::{
+        SYS_accept, SYS_accept4, SYS_bind, SYS_connect, SYS_io_uring_enter, SYS_io_uring_register,
+        SYS_io_uring_setup, SYS_listen, SYS_sendmmsg, SYS_sendmsg, SYS_sendto,
+    };
+    [
+        SYS_connect as u32,
+        SYS_bind as u32,
+        SYS_sendto as u32,
+        SYS_sendmsg as u32,
+        SYS_sendmmsg as u32,
+        SYS_listen as u32,
+        SYS_accept as u32,
+        SYS_accept4 as u32,
+        SYS_io_uring_setup as u32,
+        SYS_io_uring_enter as u32,
+        SYS_io_uring_register as u32,
+    ]
+}
+
+#[cfg(target_os = "linux")]
+fn build_child_network_filter() -> Vec<libc::sock_filter> {
+    use bpf::{EPERM_VAL, SECCOMP_RET_ALLOW, SECCOMP_RET_ERRNO, jump, stmt};
+    use libc::{BPF_JEQ, BPF_JMP, BPF_K, BPF_RET};
+
+    let blocked = child_network_blocked_syscalls();
+    let mut f = Vec::with_capacity(blocked.len() + 10);
+    bpf::push_arch_nr_gate(&mut f);
+    let n = blocked.len();
+    for (i, &sys) in blocked.iter().enumerate() {
+        let remaining = n - i - 1;
+        f.push(jump(BPF_JMP | BPF_JEQ | BPF_K, sys, remaining as u8 + 1, 0));
+    }
+    f.push(stmt(BPF_RET | BPF_K, SECCOMP_RET_ALLOW));
+    f.push(stmt(BPF_RET | BPF_K, SECCOMP_RET_ERRNO | EPERM_VAL));
+    f
+}
+
+/// Built once: `pre_exec` runs after `fork`, where allocating is unsafe.
+#[cfg(target_os = "linux")]
+pub fn prebuilt_child_network_filter() -> &'static [libc::sock_filter] {
+    static FILTER: std::sync::OnceLock<Vec<libc::sock_filter>> = std::sync::OnceLock::new();
+    FILTER.get_or_init(build_child_network_filter)
+}
+
+/// Install the seccomp BPF filter blocking network syscalls.
 ///
 /// # Safety
 ///
 /// Must be called in a `pre_exec` context (after `fork`, before `exec`).
 #[cfg(target_os = "linux")]
-pub unsafe fn install_child_network_filter() -> std::io::Result<()> {
-    use libc::{
-        BPF_ABS, BPF_JEQ, BPF_JMP, BPF_K, BPF_LD, BPF_RET, BPF_W, PR_SET_NO_NEW_PRIVS,
-        PR_SET_SECCOMP, SECCOMP_MODE_FILTER, SYS_accept, SYS_accept4, SYS_bind, SYS_connect,
-        SYS_listen, SYS_sendmsg, SYS_sendto, prctl, sock_filter, sock_fprog,
-    };
-
-    const SECCOMP_RET_ALLOW: u32 = 0x7fff_0000;
-    const SECCOMP_RET_ERRNO: u32 = 0x0005_0000;
-    // libc::EPERM
-    const EPERM_VAL: u32 = 1;
-
-    macro_rules! bpf_stmt {
-        ($code:expr, $k:expr) => {
-            sock_filter {
-                code: $code as u16,
-                jt: 0,
-                jf: 0,
-                k: $k as u32,
-            }
-        };
-    }
-
-    macro_rules! bpf_jump {
-        ($code:expr, $k:expr, $jt:expr, $jf:expr) => {
-            sock_filter {
-                code: $code as u16,
-                jt: $jt,
-                jf: $jf,
-                k: $k as u32,
-            }
-        };
-    }
-
-    // seccomp_data.nr offset
-    const NR_OFFSET: u32 = 0;
-
-    let blocked_syscalls: &[i64] = &[
-        SYS_connect,
-        SYS_bind,
-        SYS_sendto,
-        SYS_sendmsg,
-        SYS_listen,
-        SYS_accept,
-        SYS_accept4,
-    ];
-
-    let mut filter: Vec<sock_filter> = Vec::new();
-    let total_checks = blocked_syscalls.len();
-
-    filter.push(bpf_stmt!(BPF_LD | BPF_W | BPF_ABS, NR_OFFSET));
-
-    for (i, &syscall) in blocked_syscalls.iter().enumerate() {
-        let remaining = total_checks - i - 1;
-        filter.push(bpf_jump!(
-            BPF_JMP | BPF_JEQ | BPF_K,
-            syscall,
-            // match: jump to ERRNO
-            remaining as u8 + 1,
-            // no match: check next
-            0
-        ));
-    }
-
-    filter.push(bpf_stmt!(BPF_RET | BPF_K, SECCOMP_RET_ALLOW));
-
-    filter.push(bpf_stmt!(BPF_RET | BPF_K, SECCOMP_RET_ERRNO | EPERM_VAL));
+pub unsafe fn install_child_network_filter(filter: &[libc::sock_filter]) -> std::io::Result<()> {
+    use libc::{PR_SET_NO_NEW_PRIVS, PR_SET_SECCOMP, SECCOMP_MODE_FILTER, prctl, sock_fprog};

    let prog = sock_fprog {
        len: filter.len() as u16,
-        filter: filter.as_mut_ptr(),
+        filter: filter.as_ptr().cast_mut(),
    };
-
-    // Must set PR_SET_NO_NEW_PRIVS before applying seccomp filter
    // SAFETY: prctl with PR_SET_NO_NEW_PRIVS is safe in pre_exec context.
    if unsafe { prctl(PR_SET_NO_NEW_PRIVS, 1, 0, 0, 0) } != 0 {
        return Err(std::io::Error::last_os_error());
    }
-
-    // SAFETY: prog is a valid sock_fprog pointing to our filter array.
+    // SAFETY: prog is a valid sock_fprog pointing to a filter the kernel copies.
    if unsafe {
        prctl(
            PR_SET_SECCOMP,
@@ -98,7 +137,6 @@ pub unsafe fn install_child_network_filter() -> std::io::Result<()> {
    {
        return Err(std::io::Error::last_os_error());
    }
-
    Ok(())
 }

@@ -109,3 +147,101 @@ pub unsafe fn install_child_network_filter() -> std::io::Result<()> {
 pub unsafe fn install_child_network_filter() -> std::io::Result<()> {
    Ok(())
 }
+
+#[cfg(all(test, target_os = "linux"))]
+mod tests {
+    use super::bpf::*;
+    use super::{build_child_network_filter, child_network_blocked_syscalls};
+    use libc::sock_filter;
+
+    /// Minimal classic-BPF interpreter over synthetic `seccomp_data` fields.
+    fn eval(filter: &[sock_filter], arch: u32, nr: u32) -> u32 {
+        use libc::{BPF_ABS, BPF_JEQ, BPF_JMP, BPF_JSET, BPF_K, BPF_LD, BPF_RET, BPF_W};
+        let mut pc = 0usize;
+        let mut a = 0u32;
+        for _ in 0..filter.len().saturating_mul(2) {
+            let insn = &filter[pc];
+            let op = insn.code as u32;
+            if op == (BPF_LD | BPF_W | BPF_ABS) {
+                a = match insn.k {
+                    OFF_NR => nr,
+                    OFF_ARCH => arch,
+                    _ => 0,
+                };
+                pc += 1;
+            } else if op == (BPF_JMP | BPF_JEQ | BPF_K) {
+                pc = if a == insn.k {
+                    pc + 1 + insn.jt as usize
+                } else {
+                    pc + 1 + insn.jf as usize
+                };
+            } else if op == (BPF_JMP | BPF_JSET | BPF_K) {
+                pc = if a & insn.k != 0 {
+                    pc + 1 + insn.jt as usize
+                } else {
+                    pc + 1 + insn.jf as usize
+                };
+            } else if op == (BPF_RET | BPF_K) {
+                return insn.k;
+            } else {
+                panic!("unsupported opcode {:#x} at {pc}", insn.code);
+            }
+            assert!(pc < filter.len(), "pc out of range");
+        }
+        panic!("filter did not RET");
+    }
+
+    fn is_allow(r: u32) -> bool {
+        r == SECCOMP_RET_ALLOW
+    }
+
+    fn is_eperm(r: u32) -> bool {
+        r == (SECCOMP_RET_ERRNO | EPERM_VAL)
+    }
+
+    #[test]
+    fn blocks_every_network_syscall_and_allows_read_and_socket() {
+        let f = build_child_network_filter();
+        for sys in child_network_blocked_syscalls() {
+            assert!(
+                is_eperm(eval(&f, EXPECTED_ARCH, sys)),
+                "syscall {sys} must be EPERM"
+            );
+        }
+        assert!(is_allow(eval(&f, EXPECTED_ARCH, libc::SYS_read as u32)));
+        assert!(is_allow(eval(&f, EXPECTED_ARCH, libc::SYS_socket as u32)));
+    }
+
+    #[test]
+    fn covers_io_uring_and_sendmmsg() {
+        let blocked = child_network_blocked_syscalls();
+        for sys in [
+            libc::SYS_sendmmsg,
+            libc::SYS_io_uring_setup,
+            libc::SYS_io_uring_enter,
+            libc::SYS_io_uring_register,
+        ] {
+            assert!(blocked.contains(&(sys as u32)), "{sys} missing");
+        }
+    }
+
+    #[test]
+    fn wrong_arch_and_x32_are_denied_before_the_nr_match() {
+        let f = build_child_network_filter();
+        assert!(is_eperm(eval(&f, 0xdead_beef, libc::SYS_read as u32)));
+        #[cfg(target_arch = "x86_64")]
+        assert!(is_eperm(eval(
+            &f,
+            EXPECTED_ARCH,
+            (libc::SYS_read as u32) | X32_SYSCALL_BIT
+        )));
+    }
+
+    #[test]
+    fn prebuilt_filter_is_shared() {
+        let a = super::prebuilt_child_network_filter();
+        let b = super::prebuilt_child_network_filter();
+        assert!(std::ptr::eq(a, b));
+        assert_eq!(a.len(), build_child_network_filter().len());
+    }
+}
@@ -333,6 +333,27 @@ fn is_devbox_based(profile: &ProfileName, config: &SandboxConfig) -> bool {
        _ => false,
    }
 }
+/// Whether an existing `/data` needs the devbox bwrap read-only bind; a filesystem error other than NotFound conservatively requires it.
+#[cfg(target_os = "linux")]
+pub fn requires_data_write_deny(profile: &ProfileName, workspace: &Path) -> bool {
+    requires_data_write_deny_for(
+        profile,
+        &profiles::load_sandbox_config(workspace),
+        data_path_requires_bind(Path::new("/data")),
+    )
+}
+#[cfg(target_os = "linux")]
+fn requires_data_write_deny_for(
+    profile: &ProfileName,
+    config: &SandboxConfig,
+    is_data_present: bool,
+) -> bool {
+    is_devbox_based(profile, config) && is_data_present
+}
+#[cfg(target_os = "linux")]
+fn data_path_requires_bind(path: &Path) -> bool {
+    path.try_exists().unwrap_or(true)
+}
 /// Whether kernel read-deny enforcement is required. The single source of truth
 /// for this classification so callers (e.g. the shell's fail-closed startup path)
 /// cannot drift and silently fail open.
@@ -360,6 +381,23 @@ pub fn requires_read_deny(profile: &ProfileName, workspace: &Path) -> bool {
 pub fn requires_read_deny(_profile: &ProfileName, _workspace: &Path) -> bool {
    false
 }
+/// Whether the resolved profile leaves `~/.kigi` read-only outside its
+/// state dirs (strict and every profile extending it), so files first use
+/// would write there must be minted before enforcement.
+#[cfg(all(feature = "enforce", unix))]
+pub fn restricts_home_writes(profile: &ProfileName, workspace: &Path) -> bool {
+    if *profile == ProfileName::Off {
+        return false;
+    }
+    let config = profiles::load_sandbox_config(workspace);
+    profile
+        .resolve_profile(workspace, &config)
+        .is_ok_and(|resolved| !resolved.read_write.contains(&paths::kigi_home()))
+}
+#[cfg(not(all(feature = "enforce", unix)))]
+pub fn restricts_home_writes(_profile: &ProfileName, _workspace: &Path) -> bool {
+    false
+}
 /// A profile's resolved bwrap deny plan: read-only mounts (`deny_write`),
 /// bound-over unreadable placeholders (`deny_read`), and whether the profile
 /// carries deny globs (`has_globs`, so the re-exec proceeds even with zero
@@ -382,7 +420,11 @@ struct BwrapDenyPlan {
 #[cfg(all(feature = "enforce", target_os = "linux"))]
 fn bwrap_deny_plan(profile: &ProfileName, workspace: &Path) -> Option<BwrapDenyPlan> {
    let config = profiles::load_sandbox_config(workspace);
-    let deny_write: Vec<String> = if is_devbox_based(profile, &config) {
+    let deny_write: Vec<String> = if requires_data_write_deny_for(
+        profile,
+        &config,
+        data_path_requires_bind(Path::new("/data")),
+    ) {
        vec!["/data".to_string()]
    } else {
        Vec::new()
@@ -424,7 +466,11 @@ fn bwrap_deny_plan(profile: &ProfileName, workspace: &Path) -> Option<BwrapDenyP
 #[cfg(all(not(feature = "enforce"), target_os = "linux"))]
 fn bwrap_deny_plan(profile: &ProfileName, workspace: &Path) -> Option<BwrapDenyPlan> {
    let config = profiles::load_sandbox_config(workspace);
-    let deny_write: Vec<String> = if is_devbox_based(profile, &config) {
+    let deny_write: Vec<String> = if requires_data_write_deny_for(
+        profile,
+        &config,
+        data_path_requires_bind(Path::new("/data")),
+    ) {
        vec!["/data".to_string()]
    } else {
        Vec::new()
@@ -595,6 +641,43 @@ mod tests {
        );
    }
    #[test]
+    #[cfg(target_os = "linux")]
+    fn data_write_deny_requires_devbox_and_a_present_data_dir() {
+        let tmp = tempfile::tempdir().unwrap();
+        std::fs::create_dir_all(tmp.path().join(".kigi")).unwrap();
+        std::fs::write(
+            tmp.path().join(".kigi/sandbox.toml"),
+            "[profiles.devext]\nextends = \"devbox\"\n\n[profiles.plain]\nextends = \"workspace\"\n",
+        )
+        .unwrap();
+        let config = profiles::load_sandbox_config(tmp.path());
+        assert!(requires_data_write_deny_for(
+            &ProfileName::Devbox,
+            &config,
+            true
+        ));
+        assert!(!requires_data_write_deny_for(
+            &ProfileName::Devbox,
+            &config,
+            false
+        ));
+        assert!(requires_data_write_deny_for(
+            &ProfileName::Custom("devext".into()),
+            &config,
+            true
+        ));
+        assert!(!requires_data_write_deny_for(
+            &ProfileName::Custom("plain".into()),
+            &config,
+            true
+        ));
+        assert!(!requires_data_write_deny_for(
+            &ProfileName::Workspace,
+            &config,
+            true
+        ));
+    }
+    #[test]
    fn configured_profile_is_recorded() {
        set_configured_profile("read-only");
        assert_eq!(configured_profile_name(), Some("read-only"));
@@ -700,10 +783,11 @@ mod tests {
            "devbox-empty",
            "[profiles.devempty]\nextends = \"devbox\"\n",
        );
-        assert!(
+        assert_eq!(
            bwrap_reexec_for_profile(&ProfileName::Custom("devempty".to_string()), &ws_empty)
                .is_some(),
-            "devbox-extending custom must compose the /data write-deny re-exec"
+            Path::new("/data").exists(),
+            "devbox-extending custom composes the /data write-deny re-exec exactly when /data exists"
        );
        let _ = std::fs::remove_dir_all(&ws_empty);
        let ws_ws = temp_workspace_with_sandbox_toml(
@@ -717,3 +801,31 @@ mod tests {
        let _ = std::fs::remove_dir_all(&ws_ws);
    }
 }
+
+#[cfg(all(test, feature = "enforce", unix))]
+mod home_write_tests {
+    use super::{ProfileName, restricts_home_writes};
+
+    #[test]
+    fn strict_restricts_home_writes_and_workspace_does_not() {
+        let workspace = std::env::temp_dir();
+        assert!(restricts_home_writes(&ProfileName::Strict, &workspace));
+        assert!(!restricts_home_writes(&ProfileName::Workspace, &workspace));
+        assert!(!restricts_home_writes(&ProfileName::ReadOnly, &workspace));
+        assert!(!restricts_home_writes(&ProfileName::Off, &workspace));
+    }
+
+    #[test]
+    fn custom_profile_extending_strict_inherits_the_restriction() {
+        let tmp = tempfile::tempdir().unwrap();
+        let project = tmp.path().join(".kigi");
+        std::fs::create_dir_all(&project).unwrap();
+        std::fs::write(
+            project.join("sandbox.toml"),
+            "[profiles.audit-home-write-test]\nextends = \"strict\"\n",
+        )
+        .unwrap();
+        let custom = ProfileName::Custom("audit-home-write-test".to_string());
+        assert!(restricts_home_writes(&custom, tmp.path()));
+    }
+}
@@ -1,7 +1,7 @@
 //! Sandbox event logger.
 //!
 //! Events (profile applied, violations, bypasses) are buffered in memory and
-//! flushed as JSONL to `~/.kigi/sandbox-events.jsonl`.
+//! flushed as JSONL to `~/.kigi/sessions/sandbox-events.jsonl`.

 use std::path::PathBuf;
 use std::sync::Mutex;
@@ -87,8 +87,12 @@ impl SandboxLogger {
        Ok(())
    }

+    /// Under `sessions/`: the only home subtree a strict sandbox leaves
+    /// writable, and this flush runs after enforcement.
    fn log_file_path() -> PathBuf {
-        kigi_config::kigi_home().join("sandbox-events.jsonl")
+        kigi_config::kigi_home()
+            .join("sessions")
+            .join("sandbox-events.jsonl")
    }
 }

@@ -96,6 +96,37 @@ pub(crate) fn essential_writable_paths(workspace: &Path) -> Vec<PathBuf> {
    paths
 }

+/// Home subtrees the process keeps writing after enforcement: session and
+/// worktree state, memory, diagnostics. Landlock cannot carve a child out of
+/// a write grant, so the rest of `~/.kigi` (`config.toml`, `auth.json`,
+/// `sandbox.toml`, `hooks/`, `bin/`, `vendor/`, `plugin-data/`) stays
+/// read-only and a sandboxed child cannot arm the next launch.
+#[cfg(all(feature = "enforce", unix))]
+const HOME_STATE_DIRS: &[&str] = &[
+    "run",
+    "sessions",
+    "worktrees",
+    "worktree_pool",
+    "tasks",
+    "memory",
+    "logs",
+    "debug",
+    "crash",
+    "memtrace",
+    "trace-exports",
+];
+
+/// Writable paths for the strict profile: the workspace, the home state
+/// subtrees, and temp.
+#[cfg(all(feature = "enforce", unix))]
+pub(crate) fn essential_writable_paths_strict(workspace: &Path) -> Vec<PathBuf> {
+    let home = kigi_home();
+    let mut paths = vec![workspace.to_path_buf()];
+    paths.extend(HOME_STATE_DIRS.iter().map(|dir| home.join(dir)));
+    paths.extend(temp_writable_paths());
+    paths
+}
+
 /// Writable directory paths for the read-only profile (minimal: just ~/.kigi + temp).
 /// Device files are handled separately via `allow_file` in `to_capability_set_with_config`.
 #[cfg(all(feature = "enforce", unix))]
@@ -18,6 +18,7 @@ use crate::paths::kigi_home;
 #[cfg(all(feature = "enforce", unix))]
 use crate::paths::{
    DEVICE_DIRS, DEVICE_FILES, essential_writable_paths, essential_writable_paths_minimal,
+    essential_writable_paths_strict,
 };

 /// A resolved sandbox profile ready to be converted to a `CapabilitySet`.
@@ -186,6 +187,32 @@ fn load_config_file(path: &Path) -> Option<SandboxConfig> {
 #[cfg(all(feature = "enforce", unix))]
 impl ProfileName {
    /// Convert this profile into a nono `CapabilitySet` for the given workspace.
+    /// The directory a `read_write` entry really grants. A symlink under
+    /// `~/.kigi` is followed only to the default `~/.kigi/sessions` (the
+    /// `KIGI_SHARE_DIR` override case); any other target would widen the
+    /// grant to wherever the link points. Missing paths are created.
+    fn read_write_grant_path(path: &Path, home: &Path) -> Option<PathBuf> {
+        match std::fs::symlink_metadata(path) {
+            Ok(meta) if meta.file_type().is_symlink() => {
+                if path == home || !path.starts_with(home) {
+                    return Some(path.to_path_buf());
+                }
+                let real = dunce::canonicalize(path).ok()?;
+                let default_sessions = kigi_config::default_kigi_home().join("sessions");
+                (real.is_dir()
+                    && path.file_name() == Some(std::ffi::OsStr::new("sessions"))
+                    && real == default_sessions)
+                    .then_some(real)
+            }
+            Ok(_) => Some(path.to_path_buf()),
+            Err(err) if err.kind() == std::io::ErrorKind::NotFound => {
+                std::fs::create_dir_all(path).ok()?;
+                Some(path.to_path_buf())
+            }
+            Err(_) => None,
+        }
+    }
+
    pub fn to_capability_set(&self, workspace: &Path) -> anyhow::Result<CapabilitySet> {
        let config = load_sandbox_config(workspace);
        self.to_capability_set_with_config(workspace, &config)
@@ -231,13 +258,14 @@ impl ProfileName {
        // apply time (it opens an O_PATH fd), but new files within it can
        // be created freely after the sandbox is applied. Pre-create
        // directories like ~/.kigi/ that may not exist on first run.
+        let home = kigi_home();
        for path in &profile.read_write {
-            if !path.exists() && std::fs::create_dir_all(path).is_err() {
-                tracing::warn!(path = ?path, "read_write path does not exist and could not be created, skipping");
+            let Some(grant) = Self::read_write_grant_path(path, &home) else {
+                tracing::warn!(path = ?path, "read_write grant skipped");
                continue;
-            }
-            let Some(path_str) = path.to_str() else {
-                tracing::warn!(path = ?path, "Skipping non-UTF8 read_write path");
+            };
+            let Some(path_str) = grant.to_str() else {
+                tracing::warn!(path = ?grant, "Skipping non-UTF8 read_write path");
                continue;
            };
            caps = caps.allow_path(path_str, AccessMode::ReadWrite)?;
@@ -380,12 +408,15 @@ impl ProfileName {
                .chain(std::iter::once(home.join("Library")))
                .filter(|p| p.exists())
                .chain(std::iter::once(workspace.to_path_buf()))
+                // Read-only: config, auth and global hooks are read after
+                // enforcement; only sessions/ below stays writable.
+                .chain(std::iter::once(kigi_home()))
                .collect();

                Ok(SandboxProfile {
                    name: "strict".to_string(),
                    read_only: system_read,
-                    read_write: essential_writable_paths(workspace),
+                    read_write: essential_writable_paths_strict(workspace),
                    deny: vec![],
                    default_read: false,
                    restrict_network: true,
@@ -544,6 +575,82 @@ mod tests {
        }
    }

+    #[test]
+    #[cfg(all(feature = "enforce", unix))]
+    fn strict_read_write_is_workspace_home_state_and_temp() {
+        let workspace = std::env::temp_dir();
+        let profile = ProfileName::Strict
+            .resolve_profile(&workspace, &SandboxConfig::default())
+            .expect("strict resolves");
+        let home = kigi_home();
+        assert_eq!(profile.read_write[0], workspace);
+        for dir in ["sessions", "worktrees", "memory", "logs", "debug"] {
+            assert!(
+                profile.read_write.contains(&home.join(dir)),
+                "{dir} must stay writable: {:?}",
+                profile.read_write
+            );
+        }
+        for denied in [
+            home.clone(),
+            home.join("hooks"),
+            home.join("vendor"),
+            home.join("bin"),
+        ] {
+            assert!(
+                !profile.read_write.contains(&denied),
+                "strict must not grant {}: {:?}",
+                denied.display(),
+                profile.read_write
+            );
+        }
+        assert!(
+            profile.read_only.contains(&home),
+            "config/auth/hooks are read after enforcement: {:?}",
+            profile.read_only
+        );
+    }
+
+    #[test]
+    #[cfg(all(feature = "enforce", unix))]
+    fn read_write_grant_follows_real_dirs_and_creates_missing_ones() {
+        let tmp = tempfile::tempdir().unwrap();
+        let home = tmp.path().join("home");
+        std::fs::create_dir_all(home.join("sessions")).unwrap();
+        assert_eq!(
+            ProfileName::read_write_grant_path(&home.join("sessions"), &home),
+            Some(home.join("sessions"))
+        );
+        let missing = home.join("fresh");
+        assert_eq!(
+            ProfileName::read_write_grant_path(&missing, &home),
+            Some(missing.clone())
+        );
+        assert!(missing.is_dir());
+    }
+
+    #[test]
+    #[cfg(all(feature = "enforce", unix))]
+    fn read_write_grant_refuses_a_home_symlink_pointing_elsewhere() {
+        let tmp = tempfile::tempdir().unwrap();
+        let home = tmp.path().join("home");
+        let outside = tmp.path().join("outside");
+        std::fs::create_dir_all(&home).unwrap();
+        std::fs::create_dir_all(&outside).unwrap();
+        std::os::unix::fs::symlink(&outside, home.join("sessions")).unwrap();
+        assert_eq!(
+            ProfileName::read_write_grant_path(&home.join("sessions"), &home),
+            None
+        );
+        // A symlinked workspace outside the home is the user's own choice.
+        let ws_link = tmp.path().join("ws");
+        std::os::unix::fs::symlink(&outside, &ws_link).unwrap();
+        assert_eq!(
+            ProfileName::read_write_grant_path(&ws_link, &home),
+            Some(ws_link.clone())
+        );
+    }
+
    #[test]
    #[cfg(all(feature = "enforce", unix))]
    fn base_profile_capability_set_builds() {
@@ -1164,6 +1164,7 @@ mod platform {
 #[cfg(not(target_os = "macos"))]
 mod platform {
    use super::ImageData;
+    #[cfg(target_os = "linux")]
    use std::process::{Command, Stdio};

    /// No subprocess-free pasteboard probe exists off-macOS.
@@ -1926,19 +1927,18 @@ mod platform {
    // Public API

    pub fn get_text() -> anyhow::Result<Option<String>> {
-        let mut arboard_error = None;
-        match arboard_get_text() {
+        let arboard_error = match arboard_get_text() {
            Ok(Some(text)) => return Ok(Some(text)),
            // Wayland-only: arboard's empty answer is not authoritative,
            // fall through to wl-paste (see `wayland_tool_selected`).
            #[cfg(target_os = "linux")]
-            Ok(None) if wayland_tool_selected(linux_tool_spec()) => {}
+            Ok(None) if wayland_tool_selected(linux_tool_spec()) => None,
            Ok(None) => return Ok(None),
            Err(e) => {
                tracing::debug!("arboard get_text failed: {e}");
-                arboard_error = Some(e);
+                Some(e)
            }
-        }
+        };
        #[cfg(target_os = "linux")]
        if let Some(spec) = linux_tool_spec() {
            let bytes = run_capture_out_checked(spec.read_text, CLI_READ_WAIT)?;
@@ -2039,19 +2039,18 @@ mod platform {
    }

    pub fn get_image() -> anyhow::Result<Option<ImageData>> {
-        let mut arboard_error = None;
-        match arboard_get_image() {
+        let arboard_error = match arboard_get_image() {
            Ok(Some(image)) => return Ok(Some(image)),
            // Wayland-only: arboard's empty answer is not authoritative,
            // fall through to wl-paste (see `wayland_tool_selected`).
            #[cfg(target_os = "linux")]
-            Ok(None) if wayland_tool_selected(linux_tool_spec()) => {}
+            Ok(None) if wayland_tool_selected(linux_tool_spec()) => None,
            Ok(None) => return Ok(None),
            Err(e) => {
                tracing::debug!("arboard get_image failed: {e}");
-                arboard_error = Some(e);
+                Some(e)
            }
-        }
+        };
        #[cfg(target_os = "linux")]
        if let Some(spec) = linux_tool_spec()
            && let Some(argv) = spec.read_png
@@ -26,7 +26,7 @@ use std::io::{self, Write};
 use std::path::Path;

 #[cfg(unix)]
-use std::os::unix::fs::OpenOptionsExt;
+use std::os::unix::fs::{OpenOptionsExt, PermissionsExt};

 /// Creates or opens a file with secure permissions (owner read/write only).
 ///
@@ -58,17 +58,20 @@ pub fn write_secure_file(path: &Path, contents: &[u8]) -> io::Result<()> {
    file.write_all(contents)?;
    file.flush()?;

-    // On Windows, we need to set permissions after file creation
-    #[cfg(windows)]
-    {
-        set_windows_secure_permissions(path)?;
-    }
+    // Re-assert owner-only bits: `OpenOptions::mode` only applies on create,
+    // so an existing world-readable file would otherwise keep open perms.
+    ensure_owner_only_permissions(path)?;

    Ok(())
 }

 /// Opens a file for writing with secure permissions set during creation (Unix)
 /// or prepares it for permission setting after creation (Windows).
+///
+/// Callers that write secret material should also call
+/// [`ensure_owner_only_permissions`] after the write (or use
+/// [`write_secure_file`]), because `mode(0o600)` only applies when the file
+/// is newly created — not when truncating an existing path.
 pub fn open_secure_file(path: &Path) -> io::Result<File> {
    let mut options = OpenOptions::new();
    options.truncate(true).write(true).create(true);
@@ -82,6 +85,44 @@ pub fn open_secure_file(path: &Path) -> io::Result<File> {
    options.open(path)
 }

+/// Ensure `path` is owner-read/write only (Unix `0o600` / Windows user ACL).
+///
+/// Best-effort on missing files (`NotFound` is ignored). Other errors
+/// propagate so callers can fail closed when tightening a secret store.
+///
+/// Use on **load** of credential files so a hand-copied or restored
+/// world-readable `auth.json` is tightened before the process continues.
+pub fn ensure_owner_only_permissions(path: &Path) -> io::Result<()> {
+    match ensure_owner_only_permissions_inner(path) {
+        Ok(()) => Ok(()),
+        Err(e) if e.kind() == io::ErrorKind::NotFound => Ok(()),
+        Err(e) => Err(e),
+    }
+}
+
+fn ensure_owner_only_permissions_inner(path: &Path) -> io::Result<()> {
+    #[cfg(unix)]
+    {
+        let metadata = std::fs::metadata(path)?;
+        let mode = metadata.permissions().mode();
+        if mode & 0o777 != 0o600 {
+            let mut perms = metadata.permissions();
+            perms.set_mode(0o600);
+            std::fs::set_permissions(path, perms)?;
+        }
+        Ok(())
+    }
+    #[cfg(windows)]
+    {
+        set_windows_secure_permissions(path)
+    }
+    #[cfg(not(any(unix, windows)))]
+    {
+        let _ = path;
+        Ok(())
+    }
+}
+
 /// Sets Windows-specific secure permissions on a file.
 ///
 /// This function modifies the file's ACL to:
@@ -219,8 +260,6 @@ mod tests {
    #[cfg(unix)]
    #[test]
    fn test_unix_permissions() {
-        use std::os::unix::fs::PermissionsExt;
-
        let temp_dir = tempfile::tempdir().unwrap();
        let file_path = temp_dir.path().join("test_perms.txt");

@@ -231,4 +270,43 @@ mod tests {
        // Check that only owner has read/write (0o600), ignoring file type bits
        assert_eq!(mode & 0o777, 0o600);
    }
+
+    #[cfg(unix)]
+    #[test]
+    fn ensure_owner_only_tightens_world_readable_file() {
+        let temp_dir = tempfile::tempdir().unwrap();
+        let file_path = temp_dir.path().join("loose.txt");
+        fs::write(&file_path, b"secret").unwrap();
+        let mut loose = fs::metadata(&file_path).unwrap().permissions();
+        loose.set_mode(0o644);
+        fs::set_permissions(&file_path, loose).unwrap();
+        assert_eq!(
+            fs::metadata(&file_path).unwrap().permissions().mode() & 0o777,
+            0o644
+        );
+
+        ensure_owner_only_permissions(&file_path).unwrap();
+        assert_eq!(
+            fs::metadata(&file_path).unwrap().permissions().mode() & 0o777,
+            0o600
+        );
+    }
+
+    #[cfg(unix)]
+    #[test]
+    fn write_secure_file_tightens_existing_world_readable_file() {
+        let temp_dir = tempfile::tempdir().unwrap();
+        let file_path = temp_dir.path().join("existing.txt");
+        fs::write(&file_path, b"old").unwrap();
+        let mut loose = fs::metadata(&file_path).unwrap().permissions();
+        loose.set_mode(0o666);
+        fs::set_permissions(&file_path, loose).unwrap();
+
+        write_secure_file(&file_path, b"new secret").unwrap();
+        assert_eq!(
+            fs::metadata(&file_path).unwrap().permissions().mode() & 0o777,
+            0o600
+        );
+        assert_eq!(fs::read_to_string(&file_path).unwrap(), "new secret");
+    }
 }
@@ -150,6 +150,7 @@ fs2 = { workspace = true }
 zstd = { workspace = true }
 tar = { workspace = true }
 url = { workspace = true }
+kigi-secrets = { path = "../kigi-secrets" }
 walkdir = { workspace = true }
 tdigests = "1.0"
 semver = { workspace = true }
@@ -2101,7 +2101,7 @@ kigi --sandbox strict
 | `off` (default) | Unrestricted       | Unrestricted              | Unrestricted  | No sandbox               |
 | `workspace`     | Everywhere         | CWD + `/tmp` + `~/.kigi/` | Allowed       | Normal development       |
 | `read-only`     | Everywhere         | `~/.kigi/` only           | Blocked       | Exploration, code review |
-| `strict`        | CWD + system paths | CWD + `/tmp` + `~/.kigi/` | Blocked       | Untrusted code           |
+| `strict`        | CWD + system paths | CWD + `/tmp` + `~/.kigi/` state dirs | Blocked       | Untrusted code           |

 Sensitive paths (`~/.ssh/`, `~/.aws/`, `~/.gnupg/`, `~/.kigi/auth/`) are always
 write-protected regardless of profile.
@@ -2160,7 +2160,7 @@ model cannot convince the agent to relax restrictions at runtime.

 ### Event Logging

-Sandbox events (profile applied, violations) are logged to `~/.kigi/sandbox-events.jsonl`
+Sandbox events (profile applied, violations) are logged to `~/.kigi/sessions/sandbox-events.jsonl`
 for telemetry and debugging.

 ---
@@ -2171,6 +2171,19 @@ impl Config {
        )
        .with_segment_detail(self.resolve_compaction_detail())
    }
+    /// Resolve summarizer `tool_choice`: env `KIGI_COMPACTION_TOOL_CHOICE` >
+    /// config > remote settings > default (`auto`).
+    pub(crate) fn resolve_compaction_tool_choice(
+        &self,
+    ) -> crate::util::config::CompactionToolChoice {
+        crate::util::config::resolve_compaction_tool_choice_from(
+            env_string(crate::util::config::ENV_COMPACTION_TOOL_CHOICE).as_deref(),
+            self.features.compaction_tool_choice.as_deref(),
+            self.remote_settings
+                .as_ref()
+                .and_then(|r| r.compaction_tool_choice.as_deref()),
+        )
+    }
    /// Resolve verbatim-input flag: env `KIGI_COMPACTION_VERBATIM_INPUT` > config > remote settings > default `true`.
    pub(crate) fn resolve_compaction_verbatim_input(&self) -> bool {
        BoolFlag::env("KIGI_COMPACTION_VERBATIM_INPUT")
@@ -2588,6 +2601,7 @@ pub(crate) fn resolve_model_list(
        }
        resolved = prefetched;
    }
+    let mut explicit_api_backend_keys = std::collections::HashSet::new();
    for (key, model_override) in &cfg.config_models {
        let had_base = resolved.contains_key(key);
        let base = resolved.shift_remove(key);
@@ -2604,6 +2618,9 @@ pub(crate) fn resolve_model_list(
                );
            }
        }
+        if model_override.api_backend.is_some() {
+            explicit_api_backend_keys.insert(key.as_str());
+        }
        let entry = model_override.apply(key, base, &cfg.endpoints);
        tracing::debug!(
            model_key = % key, base_url = % entry.info.base_url, has_api_key = entry
@@ -2625,7 +2642,7 @@ pub(crate) fn resolve_model_list(
                    )
                })
                .collect();
-        for entry in resolved.values_mut() {
+        for (key, entry) in resolved.iter_mut() {
            if let Some((donor_cw, donor_backend)) = donors.get(&entry.info.model) {
                if entry.info.context_window.get() == default_cw {
                    tracing::debug!(
@@ -2635,7 +2652,9 @@ pub(crate) fn resolve_model_list(
                    );
                    entry.info.context_window = *donor_cw;
                }
-                if entry.info.api_backend == ApiBackend::default()
+                // An explicit `[model.X] api_backend` wins even when it is the default.
+                if !explicit_api_backend_keys.contains(key.as_str())
+                    && entry.info.api_backend == ApiBackend::default()
                    && *donor_backend != ApiBackend::default()
                {
                    entry.info.api_backend.clone_from(donor_backend);
@@ -3696,6 +3715,9 @@ pub struct Features {
    /// Feed the summarizer the verbatim conversation instead of the lossy rewrite; `None` = defer to env/remote settings/default (true).
    #[serde(default, skip_serializing_if = "Option::is_none")]
    pub compaction_verbatim_input: Option<bool>,
+    /// Summarizer `tool_choice` (`auto` | `none`); `None` = defer to env/remote settings/default (`auto`).
+    #[serde(default, skip_serializing_if = "Option::is_none")]
+    pub compaction_tool_choice: Option<String>,
    /// Snapshot a completed subagent's isolated worktree into a durable git ref
    /// and delete its directory (resume rehydrates from the ref). This is the
    /// per-deployment rollout lever (set in managed_config.toml `[features]`).
@@ -9051,6 +9073,47 @@ default = "kigi-4.5"
             sibling 'kigi-build' (same model slug), not stay at {default_cw}"
        );
    }
+    /// An explicit `[model.X] api_backend` survives a same-slug sibling even when it is the default.
+    #[test]
+    fn slug_propagation_keeps_an_explicit_default_api_backend() {
+        // Only kigi-messages has a non-default context window: the slug's single donor.
+        let default_cw = DEFAULT_CONTEXT_WINDOW;
+        let raw: toml::Value = toml::from_str(&format!(
+            r#"
+            [model.kigi-messages]
+            model = "kigi-4.5"
+            context_window = 500000
+            base_url = "https://test.example.com/v1"
+            api_backend = "messages"
+
+            [model.kigi-chat]
+            model = "kigi-4.5"
+            context_window = {default_cw}
+            base_url = "https://test.example.com/v1"
+            api_backend = "chat_completions"
+
+            [model.kigi-plain]
+            model = "kigi-4.5"
+            context_window = {default_cw}
+            base_url = "https://test.example.com/v1"
+            "#
+        ))
+        .unwrap();
+        let cfg = Config::new_from_toml_cfg(&raw).expect("config should parse");
+        let resolved = resolve_model_list(&cfg, None, &Default::default());
+        let backend = |key: &str| resolved.get(key).expect(key).info.api_backend.clone();
+        assert_eq!(
+            backend("kigi-chat"),
+            ApiBackend::ChatCompletions,
+            "explicit default backend must not be overwritten by the sibling"
+        );
+        assert_eq!(
+            backend("kigi-plain"),
+            ApiBackend::Messages,
+            "unset backend still inherits"
+        );
+        assert_eq!(backend("kigi-messages"), ApiBackend::Messages);
+    }
    /// Slug propagation should carry over api_backend but NOT agent_type.
    #[test]
    fn slug_propagation_inherits_api_backend_but_not_agent_type() {
@@ -12,8 +12,7 @@ use std::sync::Arc;
 use serde::Serialize;

 /// Where non-subscription users (no OAuth session) submit feedback instead.
-pub const FEEDBACK_ISSUES_URL: &str =
-    "https://git.zacharyzhang.com/ZacharyZhang-NY/Kigi-CLI/issues";
+pub const FEEDBACK_ISSUES_URL: &str = "https://github.com/ZacharyZhang-NY/Kigi-CLI/issues";

 /// HTTP error from the feedback endpoint with a preserved status code, so
 /// callers can distinguish auth failures (401) without string matching.
@@ -210,9 +210,9 @@ pub(crate) async fn apply(
        auto_compact_threshold_percent: new_threshold,
        responds_to: tx,
    });
-    let updated_model = rx
+    let (updated_model, context_window) = rx
        .await
-        .map_err(|_| acp::Error::internal_error().data("failed to set session model"))?;
+        .map_err(|_| acp::Error::internal_error().data("failed to set session model"))??;
    if let Some(handle) = agent.sessions.borrow_mut().get_mut(&session_id) {
        handle.model_id = model_id.clone();
        handle.reasoning_effort = applied_effort;
@@ -224,6 +224,7 @@ pub(crate) async fn apply(
        &session_id,
        model_id.0.as_ref(),
        applied_effort.map(|eff| eff.to_string()),
+        context_window.get(),
    );
    if agent.cfg.borrow().mode != config::AgentMode::Leader {
        agent.models_manager.set_current_model_id(model_id);
@@ -232,9 +233,12 @@ pub(crate) async fn apply(
            .set_current_reasoning_effort(applied_effort);
    }
    Ok(acp::SetSessionModelResponse::new().meta(
-        serde_json::json!({ "model" : updated_model, })
-            .as_object()
-            .cloned(),
+        serde_json::json!({
+            "model": updated_model,
+            "totalContextTokens": context_window.get(),
+        })
+        .as_object()
+        .cloned(),
    ))
 }
 /// Broadcast a `ModelChanged` to every client subscribed to this session so
@@ -245,12 +249,14 @@ fn broadcast_model_changed(
    session_id: &acp::SessionId,
    model_id: &str,
    reasoning_effort: Option<String>,
+    context_window: u64,
 ) {
    let notification = crate::extensions::notification::SessionNotification {
        session_id: session_id.clone(),
        update: crate::extensions::notification::SessionUpdate::ModelChanged {
            model_id: model_id.to_owned(),
            reasoning_effort,
+            context_window: Some(context_window),
        },
        meta: None,
    };
@@ -1938,7 +1938,7 @@ fn resolve_prefetch_env(kimi_code_config: Option<KimiCodeConfig>) -> Option<Pref
 /// credentials from disk.
 pub fn start_early_prefetch_with_auth(auth: Option<KimiAuth>) -> Option<EarlyPrefetchHandle> {
    let env = resolve_prefetch_env_with_auth(auth)?;
-    Some(spawn_prefetch_thread(env))
+    spawn_prefetch_thread(env)
 }

 /// Start the model-catalog prefetch on a background thread.
@@ -1949,51 +1949,64 @@ pub fn start_early_prefetch(
    kimi_code_config: Option<KimiCodeConfig>,
 ) -> Option<EarlyPrefetchHandle> {
    let env = resolve_prefetch_env(kimi_code_config)?;
-    Some(spawn_prefetch_thread(env))
+    spawn_prefetch_thread(env)
 }

-fn spawn_prefetch_thread(env: PrefetchEnv) -> EarlyPrefetchHandle {
-    std::thread::spawn(move || {
-        let mut timer = crate::instrumentation_timer!("startup.early_prefetch");
-        let proxy_endpoint = env.endpoints.proxy_url();
-        timer.with_field("endpoint", proxy_endpoint.as_str());
-        // Resolve each stored subscription-OAuth session's bearer (refreshed
-        // on expiry) so those platforms are part of the STARTUP fetch plan —
-        // otherwise a claude-pro-max-only user boots onto the bundled Kimi
-        // table until some later async refresh happens to run, and the cache
-        // origin (which encodes enabled platforms) never matches the
-        // async-path's claude-inclusive origin.
-        let oauth_tokens = tokio::runtime::Builder::new_current_thread()
-            .enable_all()
-            .build()
-            .map(|rt| {
-                rt.block_on(crate::agent::models_fetch::resolve_generic_oauth_tokens(
-                    &crate::auth::oauth_registry::pool_home(),
-                ))
-            })
-            .unwrap_or_default();
-        let models = prefetch_models_blocking(
-            &env.endpoints,
-            env.auth.as_ref(),
-            &oauth_tokens,
-            env.model_fetch_auth,
-            &env.platform_keys,
-        );
-        if (env.endpoints.deployment_key.is_some() || crate::managed_config::has_active_team_auth())
-            && crate::config::is_managed_config_stale_for(
-                &crate::managed_config::current_serving_identity(),
-            )
-            && crate::managed_config::is_fetch_enabled()
-            && let Ok(rt) = tokio::runtime::Builder::new_current_thread()
+/// `None` when the OS refuses a thread: the prefetch is an optimization and
+/// runs before the runtime builds, so a spawn panic here would preempt the
+/// runtime's own out-of-threads diagnostic.
+fn spawn_prefetch_thread(env: PrefetchEnv) -> Option<EarlyPrefetchHandle> {
+    let spawned = std::thread::Builder::new()
+        .name("early-prefetch".into())
+        .spawn(move || {
+            let mut timer = crate::instrumentation_timer!("startup.early_prefetch");
+            let proxy_endpoint = env.endpoints.proxy_url();
+            timer.with_field("endpoint", proxy_endpoint.as_str());
+            // Resolve each stored subscription-OAuth session's bearer (refreshed
+            // on expiry) so those platforms are part of the STARTUP fetch plan —
+            // otherwise a claude-pro-max-only user boots onto the bundled Kimi
+            // table until some later async refresh happens to run, and the cache
+            // origin (which encodes enabled platforms) never matches the
+            // async-path's claude-inclusive origin.
+            let oauth_tokens = tokio::runtime::Builder::new_current_thread()
                .enable_all()
                .build()
-        {
-            crate::managed_config::clear_orphan();
-            let _ = rt.block_on(crate::managed_config::sync());
-        }
+                .map(|rt| {
+                    rt.block_on(crate::agent::models_fetch::resolve_generic_oauth_tokens(
+                        &crate::auth::oauth_registry::pool_home(),
+                    ))
+                })
+                .unwrap_or_default();
+            let models = prefetch_models_blocking(
+                &env.endpoints,
+                env.auth.as_ref(),
+                &oauth_tokens,
+                env.model_fetch_auth,
+                &env.platform_keys,
+            );
+            if (env.endpoints.deployment_key.is_some()
+                || crate::managed_config::has_active_team_auth())
+                && crate::config::is_managed_config_stale_for(
+                    &crate::managed_config::current_serving_identity(),
+                )
+                && crate::managed_config::is_fetch_enabled()
+                && let Ok(rt) = tokio::runtime::Builder::new_current_thread()
+                    .enable_all()
+                    .build()
+            {
+                crate::managed_config::clear_orphan();
+                let _ = rt.block_on(crate::managed_config::sync());
+            }

-        EarlyPrefetchResult { models }
-    })
+            EarlyPrefetchResult { models }
+        });
+    match spawned {
+        Ok(handle) => Some(handle),
+        Err(e) => {
+            tracing::error!(error = %e, "early prefetch thread spawn failed; skipping prefetch");
+            None
+        }
+    }
 }

 /// Map a model id (catalog key or routing slug) to its catalog key.
@@ -1725,8 +1725,8 @@ impl acp::Agent for MvpAgent {
                return Ok(acp::PromptResponse::new(acp::StopReason::EndTurn));
            }
        }
-        let intake_lock = self.prompt_intake_lock(&arguments.session_id);
-        let intake_guard = intake_lock.lock().await;
+        let dispatch_lock = self.dispatch_lock(&arguments.session_id);
+        let dispatch_guard = dispatch_lock.lock().await;
        let meta_prompt_mode = arguments
            .meta
            .as_ref()
@@ -1830,7 +1830,7 @@ impl acp::Agent for MvpAgent {
                acp::Error::internal_error()
                    .data(format!("failed to dispatch prompt to session: {e}"))
            })?;
-        drop(intake_guard);
+        drop(dispatch_guard);
        self.push_roster_activity_delta(
            &arguments.session_id,
            crate::agent::roster::RosterActivity::Working,
@@ -2025,6 +2025,11 @@ impl acp::Agent for MvpAgent {
                .and_then(|m| m.get("rewindIfPristine"))
                .and_then(|v| v.as_bool())
                .unwrap_or(false);
+            // Ordered behind any in-flight prompt intake: a cancel that
+            // overtakes the prompt it targets lands on an idle actor, and the
+            // prompt then starts a turn nothing will ever cancel.
+            let dispatch_lock = self.dispatch_lock(&args.session_id);
+            let _dispatch_guard = dispatch_lock.lock().await;
            let _ = handle
                .cmd_tx
                .send(SessionCommand::Cancel {
@@ -978,7 +978,10 @@ impl MvpAgent {
            sessions: RefCell::new(HashMap::new()),
            activity,
            loading_sessions: RefCell::new(HashMap::new()),
-            prompt_intake_locks: RefCell::new(HashMap::new()),
+            dispatch_locks: RefCell::new(HashMap::new()),
+            subagent_limits:
+                kigi_tools::implementations::kigi::task::admission::SubagentLimits::from_env(),
+            subagent_spawn_permits: RefCell::new(HashMap::new()),
            session_threads: RefCell::new(HashMap::new()),
            resident_roster_titles: RefCell::new(HashMap::new()),
            initialize_request: OnceLock::new(),
@@ -1362,7 +1365,8 @@ impl MvpAgent {
    /// `FlushAndAck` on the persistence actor — a true sync barrier that only
    /// resolves after all queued writes (chat messages, updates) hit disk.
    ///
-    /// Returns `Ok(())` on success, `Err(reason)` on timeout or channel failure.
+    /// Returns `Ok(())` once the writes are on disk, `Err(reason)` on a failed
+    /// flush, timeout or channel failure.
    pub(crate) async fn flush_session(
        &self,
        session_id: &acp::SessionId,
@@ -1381,7 +1385,11 @@ impl MvpAgent {
            return Err("send failed");
        }
        match tokio::time::timeout(std::time::Duration::from_secs(5), rx).await {
-            Ok(Ok(_)) => Ok(()),
+            Ok(Ok(Ok(()))) => Ok(()),
+            Ok(Ok(Err(error))) => {
+                tracing::error!(session_id = %session_id.0, %error, "persistence flush failed");
+                Err("persistence flush failed")
+            }
            Ok(Err(_)) => Err("channel closed"),
            Err(_) => Err("timeout"),
        }
@@ -2183,6 +2191,7 @@ impl MvpAgent {
            .cfg
            .borrow()
            .resolve_compaction_verbatim_input();
+        let compaction_tool_choice = self.cfg.borrow().resolve_compaction_tool_choice();
        let two_pass_enabled = self.cfg.borrow().is_two_pass_compaction_enabled();
        let auto_update = self.cfg.borrow().cli.auto_update;
        let client_type = *self.client_type.borrow();
@@ -2568,6 +2577,7 @@ impl MvpAgent {
                    system_prompt_label,
                    compaction_mode,
                    compaction_verbatim_input,
+                    compaction_tool_choice,
                    two_pass_enabled,
                    buffering_settings,
                    origin_client.clone(),
@@ -462,21 +462,30 @@ pub struct MvpAgent {
    loading_sessions: RefCell<
        HashMap<acp::SessionId, tokio::sync::watch::Receiver<bool>>,
    >,
-    /// Per-session prompt-intake serialization lock. LEADER-SAFE(per-session):
-    /// keyed by SessionId, mirrors `sessions` lifecycle.
-    ///
-    /// Each incoming `session/prompt` RPC is dispatched as its own task by the
-    /// ACP message loop, and [`Self::prompt`] runs an async preamble (prompt-mode
-    /// query, trace context, model lookup) BEFORE it enqueues
-    /// `SessionCommand::Prompt` onto the actor's FIFO mailbox. Without
-    /// serialization those preambles interleave across tasks, so the mailbox —
-    /// and therefore the authoritative prompt queue — receives prompts out of
-    /// submission order. `prompt()` holds this lock across the preamble and
-    /// releases it immediately after the enqueue (the turn itself runs unlocked),
-    /// which makes intake order match arrival order.
-    prompt_intake_locks: RefCell<
+    /// Per-session lock ordering dispatch onto the actor's mailbox:
+    /// [`Self::prompt`] holds it across its intake preamble (prompt-mode
+    /// query, trace context, model lookup) and releases it right after the
+    /// enqueue, so prompts land in submission order; `cancel` holds it
+    /// around its `Cancel` send, so a cancel cannot overtake the prompt it
+    /// targets and wedge the turn slot (see
+    /// `cancel_never_overtakes_in_flight_prompt_intake`). Cancels wait out
+    /// preambles held ahead of them — keep preambles lean.
+    /// LEADER-SAFE(per-session): mirrors `sessions` lifecycle.
+    dispatch_locks: RefCell<
        HashMap<acp::SessionId, std::rc::Rc<tokio::sync::Mutex<()>>>,
    >,
+    /// Session-scoped subagent spawn limits, read once from env at
+    /// construction (composition root).
+    subagent_limits: kigi_tools::implementations::kigi::task::admission::SubagentLimits,
+    /// Per-parent-session spawn semaphores sized to `subagent_limits`; a
+    /// wide fan-out queues on the semaphore instead of exhausting file
+    /// descriptors. Keyed by session id string (the coordinator's
+    /// `parent_session_id`). Entries deliberately live for the whole
+    /// process: a removed-then-resumed session keeps its id, so sharing one
+    /// semaphore across generations keeps the limit correct while old
+    /// children still run. Growth is bounded by distinct sessions per
+    /// process and each entry is tiny. LEADER-SAFE(per-session).
+    subagent_spawn_permits: RefCell<HashMap<String, std::sync::Arc<tokio::sync::Semaphore>>>,
    /// LEADER-SAFE(per-session): keyed by SessionId. Mirrors `sessions` lifecycle.
    session_threads: RefCell<HashMap<acp::SessionId, SessionThread>>,
    /// Title per resident session id, refreshed each `build_roster`. Lets the
@@ -717,10 +726,18 @@ pub struct MvpAgent {
 /// sampling client is discarded, but the TLS root certificates it loads
 /// are cached at the process level by `rustls-native-certs`.
 pub fn warm_async_http_client() {
-    std::thread::spawn(|| {
-        let _timer = crate::instrumentation_timer!("startup.async_http_warmup");
-        let _ = crate::http::shared_client();
-    });
+    // Recoverable spawn: the warmup is an optimization that runs before the
+    // runtime builds — an out-of-threads host must reach the runtime's own
+    // diagnostic, not panic here.
+    let spawned = std::thread::Builder::new()
+        .name("http-warmup".into())
+        .spawn(|| {
+            let _timer = crate::instrumentation_timer!("startup.async_http_warmup");
+            let _ = crate::http::shared_client();
+        });
+    if let Err(e) = spawned {
+        tracing::warn!(error = %e, "async HTTP warmup thread spawn failed; first request warms instead");
+    }
 }
 pub(crate) fn resolve_required_agent_type(
    model_agent_type: Option<&str>,
@@ -38,19 +38,16 @@ impl MvpAgent {
    /// drops the `session_live_state` entry so that map stays bounded.
    pub(crate) fn remove_session(&self, id: &acp::SessionId) {
        self.sessions.borrow_mut().remove(id);
-        self.prompt_intake_locks.borrow_mut().remove(id);
+        self.dispatch_locks.borrow_mut().remove(id);
        self.session_threads.borrow_mut().remove(id);
        self.session_index_claims.borrow_mut().remove(id);
        self.require_gateway_sessions.borrow_mut().remove(id);
        self.session_live_state.borrow_mut().remove(id);
    }
-    /// Get-or-create the per-session prompt-intake lock (see
-    /// [`Self::prompt_intake_locks`]).
-    pub(super) fn prompt_intake_lock(
-        &self,
-        id: &acp::SessionId,
-    ) -> std::rc::Rc<tokio::sync::Mutex<()>> {
-        self.prompt_intake_locks
+    /// Get-or-create the per-session dispatch lock (see
+    /// [`Self::dispatch_locks`]). Cheap clone of the shared `Rc`.
+    pub(super) fn dispatch_lock(&self, id: &acp::SessionId) -> std::rc::Rc<tokio::sync::Mutex<()>> {
+        self.dispatch_locks
            .borrow_mut()
            .entry(id.clone())
            .or_default()
@@ -272,6 +272,19 @@ impl MvpAgent {
        self.try_build_subagent_spawn_context(parent_session_id)
            .expect("parent session must exist when spawning subagents")
    }
+    /// Get-or-create the parent session's spawn semaphore (see
+    /// [`Self::subagent_spawn_permits`]). Cheap clone of the shared `Arc`.
+    fn subagent_spawn_permit(&self, parent_session_id: &str) -> std::sync::Arc<tokio::sync::Semaphore> {
+        self.subagent_spawn_permits
+            .borrow_mut()
+            .entry(parent_session_id.to_string())
+            .or_insert_with(|| {
+                std::sync::Arc::new(tokio::sync::Semaphore::new(
+                    self.subagent_limits.effective_max_concurrent(),
+                ))
+            })
+            .clone()
+    }
    /// Fallible variant of [`Self::build_subagent_spawn_context`]: returns
    /// `None` when the parent `SessionHandle` is absent (evicted / torn down)
    /// instead of panicking, so read-only paths that can race a teardown can
@@ -387,6 +400,8 @@ impl MvpAgent {
        };
        Some(crate::agent::subagent::SubagentSpawnContext {
            lsp: parent_lsp,
+            spawn_limits: self.subagent_limits,
+            spawn_permits: self.subagent_spawn_permit(parent_session_id),
            gateway: self.gateway.clone(),
            client_hooks: Default::default(),
            sampling_config: self.sampling_config.borrow().clone(),
@@ -2152,6 +2152,59 @@ fn cancel_does_not_forward_to_bridge_in_local_mode() {
        );
    });
 }
+/// Regression (post-cancel slot hang): a cancel racing the prompt it targets
+/// must land on the actor mailbox AFTER that prompt, or it hits an idle
+/// actor as a no-op and the prompt then starts a turn nothing cancels.
+/// The dispatch lock (`dispatch_locks`) provides that ordering.
+#[test]
+fn cancel_never_overtakes_in_flight_prompt_intake() {
+    use crate::session::SessionCommand;
+    use acp::Agent as _;
+    run_local_for_bridge_test(|| async {
+        let agent = build_minimal_agent_for_tests();
+        let sid = acp::SessionId::new("sess-cancel-intake-race");
+        let (handle, _tx, mut cmd_rx) = make_live_session_handle(&sid, None);
+        agent.sessions.borrow_mut().insert(sid.clone(), handle);
+        let order: std::rc::Rc<std::cell::RefCell<Vec<&'static str>>> =
+            std::rc::Rc::new(std::cell::RefCell::new(Vec::new()));
+        let (intake_parked_tx, intake_parked_rx) = tokio::sync::oneshot::channel::<()>();
+        let driver_order = order.clone();
+        tokio::task::spawn_local(async move {
+            let mut intake_parked_tx = Some(intake_parked_tx);
+            while let Some(cmd) = cmd_rx.recv().await {
+                match cmd {
+                    SessionCommand::GetCurrentPromptMode { .. } => {
+                        if let Some(tx) = intake_parked_tx.take() {
+                            let _ = tx.send(());
+                            tokio::time::sleep(std::time::Duration::from_millis(10)).await;
+                        }
+                    }
+                    SessionCommand::Prompt { .. } => driver_order.borrow_mut().push("prompt"),
+                    SessionCommand::Cancel { .. } => driver_order.borrow_mut().push("cancel"),
+                    _ => {}
+                }
+            }
+        });
+        let prompt_fut = agent.prompt(acp::PromptRequest::new(
+            sid.clone(),
+            vec![acp::ContentBlock::from("hi")],
+        ));
+        let cancel_fut = async {
+            intake_parked_rx
+                .await
+                .expect("prompt intake reaches the fake actor");
+            let _ = agent
+                .cancel(acp::CancelNotification::new(sid.clone()))
+                .await;
+        };
+        let _ = futures::join!(prompt_fut, cancel_fut);
+        assert_eq!(
+            order.borrow().as_slice(),
+            ["prompt", "cancel"],
+            "cancel must land on the actor mailbox after the prompt it targets"
+        );
+    });
+}
 use crate::session::SessionCommand as TestSessionCommand;
 /// Build a session handle wired to a *live* command channel. Returns the
 /// handle (move into `sessions`) plus a probe `cmd_tx`/`cmd_rx` so a test
@@ -27,7 +27,7 @@ async fn subagent_spawn_context_inherits_parent_permission_handle() {
                kigi_workspace::permission::spawn_permission_manager(
                    sid.clone(),
                    gateway,
-                    cwd,
+                    cwd.clone(),
                    kigi_workspace::permission::types::ClientType::Generic,
                    Some(PermissionConfig::new(vec![PermissionRule {
                        action: RuleAction::Deny,
@@ -64,6 +64,8 @@ async fn subagent_spawn_context_inherits_parent_permission_handle() {
                        Some("child-session".to_owned()),
                        Some("general-purpose".to_owned()),
                        Some("permission inheritance regression".to_owned()),
+                        cwd.clone(),
+                        None,
                    )
                    .await;
                assert!(
@@ -19,6 +19,99 @@ use kigi_tools::implementations::kigi::task::types::*;
 use kigi_workspace::file_system::AsyncFileSystem;
 use kigi_hunk_tracker::HunkTrackerHandle;
 use super::*;
+
+/// Bounds each parent-side await in the child completion path so a starved parent `cmd_rx` cannot park a completed child forever.
+pub(super) const PARENT_ACK_TIMEOUT: std::time::Duration = std::time::Duration::from_secs(30);
+/// Bounds reads of the child's own actors, which freeze with a tool that blocks the child thread; teardown must still reach `Shutdown`.
+pub(super) const CHILD_ACTOR_ACK_TIMEOUT: std::time::Duration =
+    std::time::Duration::from_secs(10);
+
+/// One bounded child-actor query, degrading to `fallback` when the actor never answers.
+pub(super) async fn child_actor_query<T>(
+    what: &'static str,
+    query: impl std::future::Future<Output = T>,
+    fallback: T,
+) -> T {
+    match tokio::time::timeout(CHILD_ACTOR_ACK_TIMEOUT, query).await {
+        Ok(value) => value,
+        Err(_) => {
+            tracing::warn!(query = what, "child actor did not answer in time; using fallback");
+            fallback
+        }
+    }
+}
+
+/// Bounds the post-Shutdown worktree snapshot (a `git add` behind a stuck clean filter must not park completion).
+const WORKTREE_SNAPSHOT_TIMEOUT: std::time::Duration = std::time::Duration::from_secs(120);
+
+/// Snapshot the child worktree to `ref_name`, persist the ref, then remove the worktree; returns `(snapshot_ref, removed)`.
+async fn dispose_subagent_worktree(
+    worktree: &Path,
+    source_repo: &Path,
+    ref_name: &str,
+    meta_dir: &Path,
+    final_status: &str,
+    subagent_id: &str,
+) -> (Option<String>, bool) {
+    let snapshot = tokio::time::timeout(
+        WORKTREE_SNAPSHOT_TIMEOUT,
+        crate::session::worktree::snapshot_subagent_worktree(worktree, source_repo, ref_name),
+    )
+    .await;
+    let snapshot_ref = match snapshot {
+        Ok(Ok(snapshot_ref)) => snapshot_ref,
+        Ok(Err(e)) => {
+            tracing::warn!(
+                subagent_id = % subagent_id, worktree_path = % worktree.display(), error = % e,
+                "Failed to snapshot subagent worktree; preserving for review"
+            );
+            return (None, false);
+        }
+        Err(_) => {
+            tracing::warn!(
+                subagent_id = % subagent_id, worktree_path = % worktree.display(),
+                "worktree snapshot did not finish in time; preserving worktree for resume"
+            );
+            return (None, false);
+        }
+    };
+    if !update_subagent_meta_snapshot_ref(meta_dir, &snapshot_ref, final_status) {
+        tracing::warn!(
+            subagent_id = % subagent_id, worktree_path = % worktree.display(),
+            "snapshot_ref not persisted; preserving worktree for resume"
+        );
+        return (None, false);
+    }
+    // Removal is awaited in full: a detached deletion could race a resume that rehydrates into this path.
+    match crate::session::worktree::remove_subagent_worktree(worktree).await {
+        Ok(()) => {
+            tracing::info!(
+                subagent_id = % subagent_id, worktree_path = % worktree.display(),
+                "snapshotted and removed subagent worktree"
+            );
+            (Some(snapshot_ref), true)
+        }
+        Err(e) => {
+            tracing::warn!(
+                subagent_id = % subagent_id, worktree_path = % worktree.display(), error = % e,
+                "snapshotted subagent worktree but removal failed; ref persisted for resume"
+            );
+            (Some(snapshot_ref), false)
+        }
+    }
+}
+
+/// Await a parent ack for at most `PARENT_ACK_TIMEOUT`; `false` when it never answers or drops the sender.
+pub(super) async fn parent_ack(what: &'static str, ack: oneshot::Receiver<()>) -> bool {
+    match tokio::time::timeout(PARENT_ACK_TIMEOUT, ack).await {
+        Ok(Ok(())) => true,
+        Ok(Err(_)) => false,
+        Err(_) => {
+            tracing::warn!(ack = what, "parent did not ack in time; proceeding to teardown");
+            false
+        }
+    }
+}
 pub(super) fn task_model_override_error(
    requested: Option<&str>,
    provenance: ModelOverrideProvenance,
@@ -81,6 +174,64 @@ pub(crate) async fn handle_subagent_request(
        }
        _ => {}
    }
+    // Session-scoped concurrency admission: hold one permit for the whole
+    // child run so a wide fan-out queues (default) or fails (configured)
+    // instead of exhausting file descriptors. Queued spawns are not yet
+    // visible in the coordinator; they admit in FIFO order.
+    use kigi_tools::implementations::kigi::task::admission::LimitBehavior;
+    let _spawn_permit = match ctx.spawn_limits.behavior {
+        LimitBehavior::Fail => match ctx.spawn_permits.clone().try_acquire_owned() {
+            Ok(permit) => permit,
+            Err(_) => {
+                let limit = ctx.spawn_limits.effective_max_concurrent();
+                kigi_log::unified_log::warn(
+                    "subagent.spawn.rejected_at_concurrent_limit",
+                    Some(&ctx.parent_session_id),
+                    Some(serde_json::json!({ "limit": limit })),
+                );
+                let msg = format!(
+                    "Session subagent limit reached ({limit} running); \
+                     retry after one finishes, or raise KIGI_MAX_CONCURRENT_SUBAGENTS"
+                );
+                send_pre_spawn_failure(request, &msg, coordinator, &ctx, gateway);
+                return;
+            }
+        },
+        LimitBehavior::Queue => {
+            if ctx.spawn_permits.available_permits() == 0 {
+                kigi_log::unified_log::warn(
+                    "subagent.spawn.queued_at_concurrent_limit",
+                    Some(&ctx.parent_session_id),
+                    Some(serde_json::json!({
+                        "limit": ctx.spawn_limits.effective_max_concurrent(),
+                        "subagent_id": request.id,
+                    })),
+                );
+            }
+            ctx.spawn_permits
+                .clone()
+                .acquire_owned()
+                .await
+                .expect("subagent spawn semaphore is never closed")
+        }
+    };
+    // The parent can end while a spawn is queued on the semaphore (a queued
+    // request is not yet in the coordinator, so the teardown cancel sweep
+    // cannot see it). A closed parent command channel is the teardown
+    // signal. The channel closes when the actor finishes shutting down, so a
+    // permit granted during that brief drain can still slip through — the
+    // same in-flight window an ordinary spawn racing teardown already has;
+    // closing it fully would require queueing inside the coordinator.
+    if ctx.parent_cmd_tx.as_ref().is_some_and(|tx| tx.is_closed()) {
+        send_pre_spawn_failure(
+            request,
+            "Parent session ended while the spawn was queued at the subagent limit.",
+            coordinator,
+            &ctx,
+            gateway,
+        );
+        return;
+    }
    let run_in_background = request.run_in_background
        || definition.background.unwrap_or(false);
    let cancel_token = CancellationToken::new();
@@ -837,13 +988,22 @@ pub(crate) async fn handle_subagent_request(
            }
        }
    }
-    let agent_mcp_servers: Vec<_> = if is_plugin_agent {
-        if !definition.mcp_servers.is_empty() {
-            tracing::warn!(
-                agent = % definition.name, plugin = ? definition.plugin_name,
-                "ignoring mcpServers on plugin agent (not supported for security)"
-            );
-        }
+    let agent_mcp_servers: Vec<_> = if definition.mcp_servers.is_empty() {
+        vec![]
+    } else if is_plugin_agent {
+        tracing::warn!(
+            agent = % definition.name, plugin = ? definition.plugin_name,
+            "ignoring mcpServers on plugin agent (not supported for security)"
+        );
+        vec![]
+    } else if !crate::agent::folder_trust::agent_inline_hooks_allowed(definition.scope, || {
+        crate::agent::folder_trust::project_scope_allowed(&ctx.parent_cwd)
+    }) {
+        // An untrusted repo's inline stdio MCP command would run like an inline hook.
+        tracing::warn!(
+            agent = % definition.name,
+            "ignoring mcpServers on untrusted project agent (folder not trusted; re-run with --trust)"
+        );
        vec![]
    } else {
        definition
@@ -1006,6 +1166,7 @@ pub(crate) async fn handle_subagent_request(
            kigi_agent::DEFAULT_SYSTEM_PROMPT_LABEL.to_string(),
            kigi_chat_state::CompactionMode::Summary,
            ctx.resolve_compaction_verbatim_input(),
+            ctx.resolve_compaction_tool_choice(),
            false,
            None,
            None,
@@ -1274,7 +1435,12 @@ pub(crate) async fn handle_subagent_request(
    let duration_ms = start.elapsed().as_millis() as u64;
    let mut result = match wait_outcome {
        SubagentWaitOutcome::Cancelled => {
-            let (tool_calls, turns) = signals_snapshot_counts(&child_handle).await;
+            let (tool_calls, turns) = child_actor_query(
+                "signals_snapshot_counts",
+                signals_snapshot_counts(&child_handle),
+                (0, 0),
+            )
+            .await;
            SubagentResult {
                success: false,
                cancelled: true,
@@ -1301,14 +1467,28 @@ pub(crate) async fn handle_subagent_request(
                        },
                    ),
                ) => (snapshot.current.tool_call_count, snapshot.current.turn_count),
-                _ => signals_snapshot_counts(&child_handle).await,
+                _ => {
+                    child_actor_query(
+                        "signals_snapshot_counts",
+                        signals_snapshot_counts(&child_handle),
+                        (0, 0),
+                    )
+                    .await
+                }
            };
-            let final_text = child_handle
-                .chat_state_handle
-                .get_last_assistant_text()
-                .await
-                .unwrap_or_default();
-            let result_tokens = child_handle.chat_state_handle.get_total_tokens().await;
+            let final_text = child_actor_query(
+                "get_last_assistant_text",
+                child_handle.chat_state_handle.get_last_assistant_text(),
+                None,
+            )
+            .await
+            .unwrap_or_default();
+            let result_tokens = child_actor_query(
+                "get_total_tokens",
+                child_handle.chat_state_handle.get_total_tokens(),
+                0,
+            )
+            .await;
            match *turn_result {
                Ok(
                    Ok(
@@ -1468,14 +1648,21 @@ pub(crate) async fn handle_subagent_request(
    let final_status = result.status().to_string();
    let snapshot_dispose_enabled = ctx.resolve_subagent_worktree_snapshot_enabled();
    let telemetry_tokens = if result.tool_calls > 0 || result.success {
-        child_handle.chat_state_handle.get_total_tokens().await
+        child_actor_query(
+            "get_total_tokens",
+            child_handle.chat_state_handle.get_total_tokens(),
+            0,
+        )
+        .await
    } else {
        0
    };
-    let (subagent_usage_by_model, subagent_usage_incomplete) = match child_handle
-        .chat_state_handle
-        .try_get_session_usage()
-        .await
+    let (subagent_usage_by_model, subagent_usage_incomplete) = match child_actor_query(
+        "try_get_session_usage",
+        child_handle.chat_state_handle.try_get_session_usage(),
+        Err(()),
+    )
+    .await
    {
        Ok(u) => (Some(u.by_model.into_iter().collect::<Vec<_>>()), u.incomplete),
        Err(()) => (None, true),
@@ -1494,7 +1681,7 @@ pub(crate) async fn handle_subagent_request(
                        respond_to,
                    })
                {
-                    Ok(()) => ack.await.is_ok(),
+                    Ok(()) => parent_ack("RecordSubagentUsage", ack).await,
                    Err(_) => false,
                }
            } else {
@@ -1511,53 +1698,71 @@ pub(crate) async fn handle_subagent_request(
            .parent_prompt_id
            .clone()
            .or_else(|| coordinator.borrow().parent_prompt_id_for(&request.id));
-        if let Some(cmd_tx) = ctx.parent_cmd_tx.as_ref() {
+        // A slow parent may apply both the fold and this mark; both are idempotent.
+        let marked_by_parent = if let Some(cmd_tx) = ctx.parent_cmd_tx.as_ref() {
            let (respond_to, ack) = tokio::sync::oneshot::channel();
-            if cmd_tx
+            cmd_tx
                .send(crate::session::commands::SessionCommand::MarkSubagentUsageNotApplied {
-                    parent_prompt_id: sticky_prompt,
+                    parent_prompt_id: sticky_prompt.clone(),
                    respond_to,
                })
                .is_ok()
-            {
-                let _ = ack.await;
-            }
-        } else if let Some(ref pid) = sticky_prompt {
+                && parent_ack("MarkSubagentUsageNotApplied", ack).await
+        } else {
+            false
+        };
+        if !marked_by_parent && let Some(ref pid) = sticky_prompt {
            coordinator.borrow_mut().mark_subagent_usage_not_applied(pid);
        }
    }
    match (&ctx.parent_terminal_backend, &ctx.parent_notification_handle) {
        (Some(parent_tb), Some(parent_notif_handle)) => {
            if !request.surface_completion {
-                let reparented_task_ids: Vec<String> = parent_tb
-                    .list_tasks()
-                    .await
-                    .into_iter()
-                    .filter(|t| {
-                        !t.completed
-                            && t.owner_session_id.as_deref()
-                                == Some(&*child_session_id.0)
-                    })
-                    .map(|t| t.task_id)
-                    .collect();
-                if !reparented_task_ids.is_empty()
-                    && let Some(cmd_tx) = ctx.parent_cmd_tx.as_ref()
-                {
-                    let _ = cmd_tx
-                        .send(SessionCommand::RecordGoalTurnTaskIds {
-                            task_ids: reparented_task_ids,
-                        });
+                match tokio::time::timeout(PARENT_ACK_TIMEOUT, parent_tb.list_tasks()).await {
+                    Ok(tasks) => {
+                        let reparented_task_ids: Vec<String> = tasks
+                            .into_iter()
+                            .filter(|t| {
+                                !t.completed
+                                    && t.owner_session_id.as_deref()
+                                        == Some(&*child_session_id.0)
+                            })
+                            .map(|t| t.task_id)
+                            .collect();
+                        if !reparented_task_ids.is_empty()
+                            && let Some(cmd_tx) = ctx.parent_cmd_tx.as_ref()
+                        {
+                            let _ = cmd_tx.send(SessionCommand::RecordGoalTurnTaskIds {
+                                task_ids: reparented_task_ids,
+                            });
+                        }
+                    }
+                    Err(_) => tracing::warn!(
+                        child_session_id = % child_session_id.0,
+                        parent_session_id = % ctx.parent_session_id,
+                        "list_tasks not answered in time; skipping goal-turn task-id recording"
+                    ),
                }
            }
            let parent_backend_weak = std::sync::Arc::downgrade(parent_tb);
-            parent_tb
-                .reparent_notifications(
+            if tokio::time::timeout(
+                PARENT_ACK_TIMEOUT,
+                parent_tb.reparent_notifications(
                    &child_session_id.0,
                    &ctx.parent_session_id,
                    parent_notif_handle.clone(),
                    parent_backend_weak,
-                )
-                .await;
+                ),
+            )
+            .await
+            .is_err()
+            {
+                tracing::warn!(
+                    child_session_id = % child_session_id.0,
+                    parent_session_id = % ctx.parent_session_id,
+                    "reparent_notifications not acked in time; proceeding to child shutdown"
+                );
+            }
        }
        (Some(_), None) | (None, Some(_)) => {
            tracing::warn!(
@@ -1578,55 +1783,17 @@ pub(crate) async fn handle_subagent_request(
        if snapshot_dispose_enabled {
            let ref_name = format!("refs/kigi/subagents/{}", request.id);
            let source_repo = resolve_subagent_source_repo(&ctx);
-            match crate::session::worktree::snapshot_subagent_worktree(
-                    wt_path,
-                    &source_repo,
-                    &ref_name,
-                )
-                .await
-            {
-                Ok(snapshot_ref) => {
-                    let persisted = update_subagent_meta_snapshot_ref(
-                        &subagent_meta_dir,
-                        &snapshot_ref,
-                        &final_status,
-                    );
-                    if persisted {
-                        disposed_snapshot_ref = Some(snapshot_ref);
-                        match crate::session::worktree::remove_subagent_worktree(wt_path)
-                            .await
-                        {
-                            Ok(()) => {
-                                worktree_removed = true;
-                                tracing::info!(
-                                    subagent_id = % request.id, worktree_path = % wt_path
-                                    .display(), "snapshotted and removed subagent worktree"
-                                );
-                            }
-                            Err(e) => {
-                                tracing::warn!(
-                                    subagent_id = % request.id, worktree_path = % wt_path
-                                    .display(), error = % e,
-                                    "snapshotted subagent worktree but removal failed; ref persisted for resume"
-                                )
-                            }
-                        }
-                    } else {
-                        tracing::warn!(
-                            subagent_id = % request.id, worktree_path = % wt_path
-                            .display(),
-                            "snapshot_ref not persisted; preserving worktree for resume"
-                        );
-                    }
-                }
-                Err(e) => {
-                    tracing::warn!(
-                        subagent_id = % request.id, worktree_path = % wt_path.display(),
-                        error = % e,
-                        "Failed to snapshot subagent worktree; preserving for review"
-                    );
-                }
-            }
+            let (snapshot_ref, removed) = dispose_subagent_worktree(
+                wt_path,
+                &source_repo,
+                &ref_name,
+                &subagent_meta_dir,
+                &final_status,
+                &request.id,
+            )
+            .await;
+            disposed_snapshot_ref = snapshot_ref;
+            worktree_removed = removed;
        } else {
            tracing::info!(
                subagent_id = % request.id, worktree_path = % wt_path.display(),
@@ -133,6 +133,13 @@ impl AutoCompactThresholdTiers {
 pub(crate) struct SubagentSpawnContext {
    /// Parent's LSP runtime — inherited via ToolContext, same as fs/terminal.
    pub lsp: Option<std::sync::Arc<dyn kigi_tools::implementations::lsp::LspBackend>>,
+    /// Session-scoped spawn limits (env-resolved once at the composition root).
+    pub spawn_limits: kigi_tools::implementations::kigi::task::admission::SubagentLimits,
+    /// Parent session's spawn semaphore, sized to
+    /// `spawn_limits.effective_max_concurrent()`. A permit is held for the
+    /// whole child run, so a wide fan-out queues (or fails, per behavior)
+    /// instead of exhausting file descriptors.
+    pub spawn_permits: std::sync::Arc<tokio::sync::Semaphore>,
    #[expect(
        dead_code,
        reason = "unused in production; remove expect when wired or delete the item"
@@ -397,6 +404,21 @@ impl SubagentSpawnContext {
            .resolve()
            .value
    }
+    /// Subagent summarizer `tool_choice`, mirroring
+    /// `Config::resolve_compaction_tool_choice` (env > config > remote
+    /// settings > default `auto`).
+    pub fn resolve_compaction_tool_choice(&self) -> crate::util::config::CompactionToolChoice {
+        crate::util::config::resolve_compaction_tool_choice_from(
+            crate::agent::config::env_string(crate::util::config::ENV_COMPACTION_TOOL_CHOICE)
+                .as_deref(),
+            self.agent_config
+                .as_ref()
+                .and_then(|c| c.features.compaction_tool_choice.as_deref()),
+            self.remote_settings
+                .as_ref()
+                .and_then(|r| r.compaction_tool_choice.as_deref()),
+        )
+    }
    /// Whether a completed subagent's worktree is snapshotted into a durable ref
    /// and its directory deleted. Resolution mirrors the other subagent gates
    /// (env > config > remote settings > default). Default `false` so it ships dark;
@@ -3298,3 +3298,27 @@ fn spawn_test_parent_chat_state(model_slug: &str) -> kigi_chat_state::ChatStateH
    )
 }
 mod rest;
+
+/// A blocked child actor must not park teardown: the read degrades to the fallback in bounded time.
+#[tokio::test(start_paused = true)]
+async fn child_actor_query_is_bounded_when_the_actor_never_answers() {
+    use super::handle_request::{CHILD_ACTOR_ACK_TIMEOUT, child_actor_query};
+    let tokens = tokio::time::timeout(
+        20 * CHILD_ACTOR_ACK_TIMEOUT,
+        child_actor_query("test_query", std::future::pending::<u64>(), 7),
+    )
+    .await
+    .expect("child-actor queries must complete in bounded time");
+    assert_eq!(tokens, 7, "a starved query must degrade to the fallback");
+}
+
+/// A parent that never services `cmd_rx` must not park a completed child.
+#[tokio::test(start_paused = true)]
+async fn parent_ack_is_bounded_when_the_parent_never_answers() {
+    use super::handle_request::{PARENT_ACK_TIMEOUT, parent_ack};
+    let (_never_answered, ack) = tokio::sync::oneshot::channel::<()>();
+    let acked = tokio::time::timeout(20 * PARENT_ACK_TIMEOUT, parent_ack("test_ack", ack))
+        .await
+        .expect("parent acks must complete in bounded time");
+    assert!(!acked, "an unanswered ack must report a miss");
+}
@@ -224,6 +224,18 @@ impl AuthFileLock {

 pub fn read_auth_json(auth_file: &Path) -> std::io::Result<AuthStore> {
    let mut file = File::open(auth_file)?;
+
+    // Tighten world-readable copies (hand-restored, umask edge cases, etc.)
+    // right after open, so even an unreadable/invalid file gets fixed.
+    // Best-effort: a chmod failure must not block login/read paths.
+    if let Err(e) = crate::util::secure_file::ensure_owner_only_permissions(auth_file) {
+        tracing::warn!(
+            path = %auth_file.display(),
+            error = %e,
+            "auth: failed to enforce owner-only permissions on auth.json"
+        );
+    }
+
    let mut contents = String::new();
    file.read_to_string(&mut contents)?;

@@ -289,6 +301,14 @@ pub(crate) fn backup_corrupt_auth_file(path: &Path) -> Option<PathBuf> {

    match std::fs::rename(path, &backup) {
        Ok(()) => {
+            // Corrupt backups still hold token material — keep them owner-only.
+            if let Err(e) = crate::util::secure_file::ensure_owner_only_permissions(&backup) {
+                tracing::warn!(
+                    path = %backup.display(),
+                    error = %e,
+                    "auth: failed to tighten permissions on corrupt-auth backup"
+                );
+            }
            tracing::warn!(
                original = %path.display(),
                backup = %backup.display(),
@@ -421,18 +441,59 @@ fn write_store_to(path: &Path, auth_store: &AuthStore) -> std::io::Result<()> {
        .into_inner()
        .map_err(|e| e.into_error())?
        .sync_all()?;
-    #[cfg(windows)]
-    {
-        crate::util::secure_file::set_windows_secure_permissions(path)?;
+    // `open_secure_file` mode bits apply only on create; tighten existing
+    // paths. Best-effort after durable content: a chmod-only failure must not
+    // look like a failed write — the in-place fallback restores the prior
+    // snapshot on any `write_store_to` Err, which would discard freshly
+    // written tokens. The load path re-tightens on the next read.
+    if let Err(e) = crate::util::secure_file::ensure_owner_only_permissions(path) {
+        tracing::warn!(
+            error = %e,
+            path = %path.display(),
+            "auth: failed to ensure owner-only permissions after write"
+        );
    }
    Ok(())
 }

 /// Atomic write: tmp + Windows-safe replace (see `util::fs::replace_file`).
 fn write_auth_json_atomic(auth_file: &Path, auth_store: &AuthStore) -> std::io::Result<()> {
-    let tmp = auth_file.with_extension(format!("json.{}.tmp", std::process::id()));
-    write_store_to(&tmp, auth_store)?;
-    crate::util::fs::replace_file(&tmp, auth_file)
+    write_auth_json_atomic_with(auth_file, auth_store, write_store_to)
+}
+
+/// Inner of [`write_auth_json_atomic`] with the tmp-file writer injectable so
+/// the torn-tmp-write path is unit-testable without real I/O faults.
+fn write_auth_json_atomic_with(
+    auth_file: &Path,
+    auth_store: &AuthStore,
+    write: fn(&Path, &AuthStore) -> std::io::Result<()>,
+) -> std::io::Result<()> {
+    // Unique per write (pid + monotonic seq): two concurrent in-process
+    // writers (e.g. background mint + proactive refresher) must not share
+    // one tmp path, or each can tear the other's half-written bytes.
+    static TMP_SEQ: std::sync::atomic::AtomicU64 = std::sync::atomic::AtomicU64::new(0);
+    let tmp = auth_file.with_extension(format!(
+        "json.{}.{}.tmp",
+        std::process::id(),
+        TMP_SEQ.fetch_add(1, std::sync::atomic::Ordering::Relaxed)
+    ));
+    if let Err(e) = write(&tmp, auth_store) {
+        // Reclaim the partial tmp; the unique name otherwise accumulates
+        // one orphan per failed write. `replace_file` cleans its own leg.
+        let _ = std::fs::remove_file(&tmp);
+        return Err(e);
+    }
+    crate::util::fs::replace_file(&tmp, auth_file)?;
+    // Re-assert on the final path (covers rename edge cases / FS quirks).
+    // Best-effort: the replace already published the new tokens.
+    if let Err(e) = crate::util::secure_file::ensure_owner_only_permissions(auth_file) {
+        tracing::warn!(
+            error = %e,
+            path = %auth_file.display(),
+            "auth: failed to ensure owner-only permissions after replace"
+        );
+    }
+    Ok(())
 }

 /// Non-atomic fallback: truncate and rewrite `auth.json` in place.
@@ -486,10 +547,7 @@ fn restore_prior_bytes(auth_file: &Path, bytes: &[u8]) -> std::io::Result<()> {
    let mut file = open_secure_file(auth_file)?;
    file.write_all(bytes)?;
    file.sync_all()?;
-    #[cfg(windows)]
-    {
-        crate::util::secure_file::set_windows_secure_permissions(auth_file)?;
-    }
+    crate::util::secure_file::ensure_owner_only_permissions(auth_file)?;
    Ok(())
 }

@@ -804,7 +862,87 @@ mod write_fallback_tests {
        );
    }

-    /// Rollback after a failed write must keep the file owner-only (0o600).
+    /// An in-place rewrite over a pre-existing world-readable auth.json must
+    /// tighten it to 0o600 — `open_secure_file` mode bits apply only on
+    /// create, so truncating an existing loose file keeps its old mode. The
+    /// in-place path is the one that reuses the inode (the atomic path
+    /// replaces it with a fresh 0o600 temp file).
+    #[cfg(unix)]
+    #[test]
+    fn in_place_write_tightens_preexisting_world_readable_auth_json() {
+        use std::os::unix::fs::PermissionsExt;
+        let dir = tempfile::tempdir().unwrap();
+        let path = dir.path().join("auth.json");
+        std::fs::write(&path, b"{}").unwrap();
+        let mut loose = std::fs::metadata(&path).unwrap().permissions();
+        loose.set_mode(0o644);
+        std::fs::set_permissions(&path, loose).unwrap();
+
+        write_auth_json_in_place(&path, &sample_store()).unwrap();
+        let mode = std::fs::metadata(&path).unwrap().permissions().mode();
+        assert_eq!(
+            mode & 0o777,
+            0o600,
+            "in-place rewrite must tighten preexisting open perms"
+        );
+    }
+
+    /// Simulates a torn tmp write: partial bytes land at the target path,
+    /// then the write fails — the temp file must not be left behind.
+    fn fake_partial_write_then_fail(path: &Path, _: &AuthStore) -> std::io::Result<()> {
+        std::fs::write(path, b"{ torn")?;
+        Err(std::io::Error::from(std::io::ErrorKind::StorageFull))
+    }
+
+    /// A tmp write that fails partway must reclaim the temp file; the unique
+    /// tmp name otherwise accumulates one orphan per failed write.
+    #[test]
+    fn atomic_write_reclaims_tmp_when_store_write_fails() {
+        let dir = tempfile::tempdir().unwrap();
+        let path = dir.path().join("auth.json");
+        let err = write_auth_json_atomic_with(&path, &sample_store(), fake_partial_write_then_fail)
+            .unwrap_err();
+        assert_eq!(err.kind(), std::io::ErrorKind::StorageFull);
+
+        let orphans: Vec<_> = std::fs::read_dir(dir.path())
+            .unwrap()
+            .filter_map(Result::ok)
+            .filter(|e| e.file_name().to_string_lossy().contains(".tmp"))
+            .collect();
+        assert!(
+            orphans.is_empty(),
+            "failed tmp write must not leave temp files behind: {orphans:?}"
+        );
+    }
+
+    /// The replace step failing (target path occupied by a directory) must
+    /// also leave no temp file — `replace_file` reclaims its tmp on failure.
+    #[test]
+    fn atomic_write_reclaims_tmp_when_replace_fails() {
+        let dir = tempfile::tempdir().unwrap();
+        let path = dir.path().join("auth.json");
+        std::fs::create_dir(&path).unwrap();
+
+        assert!(
+            write_auth_json_atomic(&path, &sample_store()).is_err(),
+            "replacing a directory must fail"
+        );
+
+        let orphans: Vec<_> = std::fs::read_dir(dir.path())
+            .unwrap()
+            .filter_map(Result::ok)
+            .filter(|e| e.file_name().to_string_lossy().contains(".tmp"))
+            .collect();
+        assert!(
+            orphans.is_empty(),
+            "failed replace must not leave temp files behind: {orphans:?}"
+        );
+    }
+
+    /// Rollback after a failed write must leave the file owner-only (0o600)
+    /// even when the file had loose permissions going in — the restore path
+    /// truncates the existing inode, which keeps its old mode unless
+    /// explicitly tightened.
    #[cfg(unix)]
    #[test]
    fn in_place_restore_is_owner_only() {
@@ -812,9 +950,13 @@ mod write_fallback_tests {
        let dir = tempfile::tempdir().unwrap();
        let path = dir.path().join("auth.json");
        write_auth_json_in_place(&path, &sample_store()).unwrap();
+        let mut loose = std::fs::metadata(&path).unwrap().permissions();
+        loose.set_mode(0o644);
+        std::fs::set_permissions(&path, loose).unwrap();
+
        let _ = write_auth_json_in_place_with(&path, &sample_store(), fake_truncate_then_fail);
        let mode = std::fs::metadata(&path).unwrap().permissions().mode();
-        assert_eq!(mode & 0o777, 0o600, "restored file must stay 0o600");
+        assert_eq!(mode & 0o777, 0o600, "restored file must be 0o600");
    }
 }

@@ -1034,16 +1034,27 @@ pub fn apply_sandbox(
        .and_then(|p| dunce::canonicalize(p).ok())
        .or_else(|| std::env::current_dir().ok())
        .unwrap_or_else(|| std::path::PathBuf::from("."));
+    if kigi_sandbox::restricts_home_writes(&sandbox_profile, &workspace) {
+        // The profile leaves ~/.kigi read-only outside its state dirs: mint
+        // what first use would otherwise write after enforcement.
+        if let Err(e) = crate::auth::device::device_id() {
+            tracing::warn!(error = %e, "sandbox: device id not created before enforcement");
+        }
+        kigi_tools::implementations::kigi::grep::ripgrep::rg_path();
+        crate::builtin::extract_bundled_files(&crate::util::kigi_home::kigi_home());
+    }
+    // The devbox `/data` write-deny is a bwrap mount; Landlock would grant the workspace under `/data`.
    #[cfg(target_os = "linux")]
-    let requires_read_deny = kigi_sandbox::requires_read_deny(&sandbox_profile, &workspace);
+    let bwrap_required = kigi_sandbox::requires_read_deny(&sandbox_profile, &workspace)
+        || kigi_sandbox::requires_data_write_deny(&sandbox_profile, &workspace);
    #[cfg(target_os = "linux")]
    {
        let refuse_unprotected = |detail: &str| {
            eprintln!(
-                "error: this sandbox could not enforce its read-deny set on Linux \
-                 (bubblewrap missing/unusable, or a deny glob exceeded its expansion \
-                 limit — see any message above). Install bubblewrap with \
-                 `apt install -y bubblewrap` if needed. Refusing to start with denied \
+                "error: this sandbox could not enforce its read-deny set or the devbox \
+                 `/data` write-deny on Linux (bubblewrap missing/unusable, or a deny glob \
+                 exceeded its expansion limit — see any message above). Install bubblewrap \
+                 with `apt install -y bubblewrap` if needed. Refusing to start with denied \
                 paths unprotected.{detail}"
            );
        };
@@ -1051,7 +1062,7 @@ pub fn apply_sandbox(
            Some(mut cmd) => {
                use std::os::unix::process::CommandExt;
                let err = cmd.exec();
-                if requires_read_deny {
+                if bwrap_required {
                    refuse_unprotected(&format!(" (bwrap exec failed: {err})"));
                    std::process::exit(1);
                }
@@ -1061,7 +1072,7 @@ pub fn apply_sandbox(
                     Install bubblewrap: apt install -y bubblewrap"
                );
            }
-            None if requires_read_deny && !kigi_sandbox::is_inside_bwrap() => {
+            None if bwrap_required && !kigi_sandbox::is_inside_bwrap() => {
                refuse_unprotected("");
                std::process::exit(1);
            }
@@ -37,9 +37,9 @@ async fn handle_compact(agent: &MvpAgent, args: &acp::ExtRequest) -> ExtResult {
            respond_to: tx,
        });
    }
+    // The session's own ACP error goes through unchanged: a Debug wrapper would escape its detail.
    rx.await
-        .map_err(|_| acp::Error::internal_error().data("session failed to respond"))?
-        .map_err(|e| acp::Error::internal_error().data(format!("Internal error: {:?}", e)))?;
+        .map_err(|_| acp::Error::internal_error().data("session failed to respond"))??;
    to_raw_response(&CompactConversationResponse {})
 }

@@ -691,6 +691,11 @@ pub enum SessionUpdate {
        /// does not support reasoning effort or no effort override was applied.
        #[serde(default, skip_serializing_if = "Option::is_none")]
        reasoning_effort: Option<String>,
+        /// The session's effective context window after the switch, so a
+        /// follower repaints its header without waiting for the next turn.
+        /// `None` from an agent that predates the field.
+        #[serde(default, skip_serializing_if = "Option::is_none")]
+        context_window: Option<u64>,
    },
    /// Streaming chunk of a tool call's arguments.
    ///
@@ -1970,15 +1975,18 @@ mod tests {
        let with_effort = SessionUpdate::ModelChanged {
            model_id: "kigi-4".into(),
            reasoning_effort: Some("high".into()),
+            context_window: Some(262_144),
        };
        let json = serde_json::to_value(&with_effort).unwrap();
        assert_eq!(json["sessionUpdate"], "model_changed");
        assert_eq!(json["model_id"], "kigi-4");
        assert_eq!(json["reasoning_effort"], "high");
+        assert_eq!(json["context_window"], 262_144);

        let without_effort = SessionUpdate::ModelChanged {
            model_id: "kigi-3".into(),
            reasoning_effort: None,
+            context_window: None,
        };
        let json = serde_json::to_value(&without_effort).unwrap();
        assert_eq!(json["sessionUpdate"], "model_changed");
@@ -1988,6 +1996,7 @@ mod tests {
            "reasoning_effort: None must be skipped on the wire so old pagers \
             and third-party ACP clients see a smaller, no-extra-keys payload"
        );
+        assert!(json.get("context_window").is_none());
    }

    /// `ModelChanged` round-trips through JSON: a follower client deserializes
@@ -2000,6 +2009,7 @@ mod tests {
        let original = SessionUpdate::ModelChanged {
            model_id: "kigi-4".into(),
            reasoning_effort: Some("medium".into()),
+            context_window: Some(1_000_000),
        };
        let json_str = serde_json::to_string(&original).unwrap();
        let parsed: SessionUpdate = serde_json::from_str(&json_str).unwrap();
@@ -2021,6 +2031,7 @@ mod tests {
            update: SessionUpdate::ModelChanged {
                model_id: "kigi-4".into(),
                reasoning_effort: None,
+                context_window: None,
            },
            meta: None,
        };
@@ -52,10 +52,16 @@ pub fn default_lock_path_in(root: &Path) -> PathBuf {
    root.join("leader.lock")
 }

+/// Runtime files live under `~/.kigi/run`: the one home subtree a strict
+/// sandbox leaves writable for them.
+pub(super) fn runtime_root() -> PathBuf {
+    kigi_home().join("run")
+}
+
 /// Effective leader lock path: the [`LEADER_SOCKET_ENV`] override's sibling
-/// `.lock` when set, else the default under kigi home.
+/// `.lock` when set, else the default under `~/.kigi/run`.
 pub fn default_lock_path() -> PathBuf {
-    resolve_lock_path(leader_socket_override(), &kigi_home())
+    resolve_lock_path(leader_socket_override(), &runtime_root())
 }

 pub fn default_socket_path_in(root: &Path) -> PathBuf {
@@ -63,9 +69,9 @@ pub fn default_socket_path_in(root: &Path) -> PathBuf {
 }

 /// Effective leader socket path: the [`LEADER_SOCKET_ENV`] override when set,
-/// else the default under kigi home.
+/// else the default under `~/.kigi/run`.
 pub fn default_socket_path() -> PathBuf {
-    resolve_socket_path(leader_socket_override(), &kigi_home())
+    resolve_socket_path(leader_socket_override(), &runtime_root())
 }

 /// The instance suffix encoded in a lock/socket file-name pair
@@ -145,6 +151,9 @@ impl LeaderLock {
    }

    fn open_lock_file(&self) -> Result<File, LockError> {
+        if let Some(parent) = self.lock_path.parent() {
+            std::fs::create_dir_all(parent)?;
+        }
        Ok(OpenOptions::new()
            .read(true)
            .write(true)
@@ -23,7 +23,7 @@
 //! │  │   - Tracks session ownership for routing                 ││
 //! │  └────────────────────────┬────────────────────────────────┘│
 //! └───────────────────────────┼──────────────────────────────────┘
-//!                             │ IPC (Unix socket at ~/.kigi/leader.sock)
+//!                             │ IPC (Unix socket at ~/.kigi/run/leader.sock)
 //!         ┌───────────────────┼───────────────────┐
 //!         ▼                   ▼                   ▼
 //! ┌───────────────┐   ┌───────────────┐   ┌───────────────┐
@@ -481,7 +481,10 @@ async fn discover_leaders_in(root: &Path) -> Vec<LeaderDescriptor> {
    entries
 }
 pub async fn discover_leaders() -> Vec<LeaderDescriptor> {
-    discover_leaders_in(&crate::util::kigi_home::kigi_home()).await
+    let mut entries = discover_leaders_in(&lock::runtime_root()).await;
+    // Pre-0.1.15 leaders and socket overrides keep their files at the home root.
+    entries.extend(discover_leaders_in(&crate::util::kigi_home::kigi_home()).await);
+    entries
 }
 /// (pid, leader_binary_version) of socket-verified (Reachable) leaders; a
 /// stale-lock-only descriptor is skipped (its `pid_from_lock` may be recycled).
@@ -610,13 +613,28 @@ fn resolve_matching_descriptors(
 fn resolve_target_from_descriptors(
    target: LeaderTarget,
    leaders: Vec<LeaderDescriptor>,
+    current_default_socket: &Path,
 ) -> Result<LeaderTargetSelection, LeaderTargetError> {
    match target {
        LeaderTarget::Default => {
-            let matching: Vec<_> = leaders
+            let defaults: Vec<_> = leaders
                .into_iter()
                .filter(|descriptor| descriptor.is_default)
                .collect();
+            // An upgrade can leave a pre-0.1.15 default at the home root; the
+            // current default socket wins while it exists.
+            let current: Vec<_> = defaults
+                .iter()
+                .filter(|descriptor| {
+                    descriptor.socket_path.as_deref() == Some(current_default_socket)
+                })
+                .cloned()
+                .collect();
+            let matching = if current.is_empty() {
+                defaults
+            } else {
+                current
+            };
            resolve_matching_descriptors("default", matching)
        }
        LeaderTarget::Socket(socket_path) => {
@@ -685,7 +703,7 @@ pub async fn resolve_leader_target(
    target: LeaderTarget,
 ) -> Result<LeaderTargetSelection, LeaderTargetError> {
    let leaders = discover_leaders().await;
-    resolve_target_from_descriptors(target, leaders)
+    resolve_target_from_descriptors(target, leaders, &lock::default_socket_path())
 }
 #[derive(Debug, thiserror::Error)]
 pub enum ConnectionError {
@@ -1315,7 +1333,12 @@ fn spawn_leader_subprocess() -> Result<u32, ConnectionError> {
    }
    cmd.stdin(std::process::Stdio::null())
        .stdout(std::process::Stdio::null());
-    let log_path = crate::util::kigi_home::kigi_home().join("leader.log");
+    let log_path = crate::util::kigi_home::kigi_home()
+        .join("logs")
+        .join("leader.log");
+    if let Some(parent) = log_path.parent() {
+        let _ = std::fs::create_dir_all(parent);
+    }
    match std::fs::File::create(&log_path) {
        Ok(log_file) => {
            info!("Leader stderr → log file");
@@ -1444,6 +1467,46 @@ mod tests {
            vec![(222, "0.2.52".to_string())]
        );
    }
+    fn reachable_default(socket: &str) -> LeaderDescriptor {
+        let socket_path = PathBuf::from(socket);
+        LeaderDescriptor {
+            pid_from_lock: Some(1),
+            lock_path: Some(socket_path.with_extension("lock")),
+            socket_path: Some(socket_path.clone()),
+            socket_suffix: String::new(),
+            classification: LeaderDiscoveryState::Reachable,
+            is_default: true,
+            live_info: Some(LiveLeaderInfo {
+                pid: 1,
+                socket_path,
+                lock_path: PathBuf::new(),
+                socket_suffix: String::new(),
+                leader_protocol_version: 1,
+                leader_binary_version: "0.1.15".to_string(),
+            }),
+            target_error: None,
+        }
+    }
+    #[test]
+    fn default_target_prefers_the_current_socket_over_a_legacy_default() {
+        let legacy = reachable_default("/h/.kigi/leader.sock");
+        let current = reachable_default("/h/.kigi/run/leader.sock");
+        let current_socket = PathBuf::from("/h/.kigi/run/leader.sock");
+        let picked = resolve_target_from_descriptors(
+            LeaderTarget::Default,
+            vec![current.clone(), legacy.clone()],
+            &current_socket,
+        )
+        .expect("current default resolves");
+        assert_eq!(picked.descriptor.socket_path, current.socket_path);
+        let fallback = resolve_target_from_descriptors(
+            LeaderTarget::Default,
+            vec![legacy.clone()],
+            &current_socket,
+        )
+        .expect("legacy default still resolves alone");
+        assert_eq!(fallback.descriptor.socket_path, legacy.socket_path);
+    }
    #[test]
    fn leader_is_older_than_directional() {
        assert!(leader_is_older_than("0.1.0", "0.2.0"));
@@ -1146,6 +1146,9 @@ pub async fn run_leader_server(
    control_state: LeaderServerControlState,
 ) -> Result<(), ServerError> {
    let _ = std::fs::remove_file(&socket_path);
+    if let Some(parent) = socket_path.parent() {
+        std::fs::create_dir_all(parent)?;
+    }
    let shutdown_reason_rx = shutdown_tx.subscribe();
    let listener = LeaderListener::bind(&socket_path)?;
    info!("Leader server listening");
@@ -171,6 +171,70 @@ pub fn error_detail_from_data(data: &serde_json::Value) -> Option<String> {
        .map(str::to_owned)
 }

+/// The human-readable text of an ACP error: its `data` detail when present (Display would JSON-escape it), else the message.
+pub fn acp_error_message(err: &acp::Error) -> String {
+    err.data
+        .as_ref()
+        .and_then(error_detail_from_data)
+        .unwrap_or_else(|| err.message.clone())
+}
+
+/// First line of a compaction error without internal `compact failed:` prefixes, capped for a banner; diagnostic lines (request URLs) are dropped.
+pub fn user_facing_compact_error(raw: &str) -> String {
+    const INTERNAL_PREFIX: &str = "compact failed:";
+    const MAX_CHARS: usize = 200;
+    // Redact before line selection: a line break inside a URL must not hide its `@`.
+    let raw = redact_urls_in_text(raw);
+    let mut rest = raw.trim();
+    while rest
+        .get(..INTERNAL_PREFIX.len())
+        .is_some_and(|head| head.eq_ignore_ascii_case(INTERNAL_PREFIX))
+    {
+        rest = rest[INTERNAL_PREFIX.len()..].trim_start();
+    }
+    rest.lines()
+        .next()
+        .unwrap_or_default()
+        .trim()
+        .chars()
+        .take(MAX_CHARS)
+        .collect()
+}
+
+/// Every canonical `scheme://` URL in `text` loses userinfo, sensitive query values and fragment; one that fails to parse is not echoed.
+fn redact_urls_in_text(text: &str) -> String {
+    let mut out = String::with_capacity(text.len());
+    let mut rest = text;
+    while let Some(sep) = rest.find("://") {
+        // The delimiter before the scheme may be multi-byte (`“`, `：`): step past its UTF-8 length.
+        let scheme_start = rest[..sep]
+            .char_indices()
+            .rev()
+            .find(|(_, c)| !(c.is_ascii_alphanumeric() || matches!(c, '+' | '-' | '.')))
+            .map_or(0, |(at, c)| at + c.len_utf8());
+        out.push_str(&rest[..scheme_start]);
+        // The authority may hold whitespace (a configured password); the token ends at whitespace after it.
+        let authority = &rest[sep + 3..];
+        let authority_end = authority.find(['/', '?', '#']).unwrap_or(authority.len());
+        let token_end = authority[authority_end..]
+            .find(char::is_whitespace)
+            .map_or(rest.len(), |at| sep + 3 + authority_end + at);
+        let token = &rest[scheme_start..token_end];
+        let url_text = token.trim_end_matches([')', ']', '}', ',', '.', ';', '\'', '"', '>']);
+        match url::Url::parse(url_text) {
+            Ok(mut url) => {
+                kigi_secrets::redact_url(&mut url);
+                out.push_str(url.as_str());
+            }
+            Err(_) => out.push_str("<unparseable url>"),
+        }
+        out.push_str(&token[url_text.len()..]);
+        rest = &rest[token_end..];
+    }
+    out.push_str(rest);
+    out
+}
+
 pub fn http_status_from_error(err: &acp::Error) -> Option<u16> {
    err.data
        .as_ref()?
@@ -254,6 +318,77 @@ pub fn prompt_complete_fields(

 #[cfg(test)]
 mod tests {
+    #[test]
+    fn compact_error_text_keeps_the_first_decoded_line_only() {
+        let err = super::acp::Error::internal_error().data(
+            "compact failed: API error (status 500): overloaded\n\nRequest URL: http://u:TEST_PASSWORD@127.0.0.1/v1",
+        );
+        let detail = super::user_facing_compact_error(&super::acp_error_message(&err));
+        assert_eq!(detail, "API error (status 500): overloaded");
+        assert!(!detail.contains("TEST_PASSWORD"));
+        assert_eq!(
+            super::user_facing_compact_error(&"x".repeat(500))
+                .chars()
+                .count(),
+            200
+        );
+        assert_eq!(
+            super::acp_error_message(&super::acp::Error::internal_error()),
+            "Internal error"
+        );
+    }
+
+    /// A sampler 401 names the endpoint on the retained line; its userinfo must not survive.
+    #[test]
+    fn compact_error_text_redacts_url_credentials_on_the_first_line() {
+        for password in [
+            "TEST_PASSWORD",
+            "TEST PASSWORD",
+            "TEST%20PASSWORD",
+            "TEST_PASSWORD\n",
+            "TEST_PASSWORD\r\n",
+        ] {
+            let raw = format!(
+                "compact failed: Unauthorized (401) from http://u:{password}@127.0.0.1:33263/401/v1/chat/completions: bad auth"
+            );
+            assert_eq!(
+                super::user_facing_compact_error(&raw),
+                "Unauthorized (401) from http://127.0.0.1:33263/401/v1/chat/completions: bad auth",
+                "{password:?}"
+            );
+        }
+        assert_eq!(
+            super::redact_urls_in_text("see https://a:b@h.example/x and http://h2/"),
+            "see https://h.example/x and http://h2/"
+        );
+    }
+
+    /// A credential carried as a query value survives no failure shape: 401 message or transport error.
+    #[test]
+    fn compact_error_text_redacts_query_credentials_in_every_failure_shape() {
+        let unauthorized = "compact failed: Unauthorized (401) from http://127.0.0.1:1/?access_token=TEST_QUERY_TOKEN&route=/v1/chat/completions: bad auth";
+        let transport = "compact failed: request error: error sending request for url (http://127.0.0.1:1/?access_token=TEST_QUERY_TOKEN&route=/500/v1/chat/completions)";
+        for raw in [unauthorized, transport] {
+            let detail = super::user_facing_compact_error(raw);
+            assert!(!detail.contains("TEST_QUERY_TOKEN"), "{detail}");
+            assert!(detail.contains("route="), "{detail}");
+        }
+        assert!(super::user_facing_compact_error("see http://:://").contains("<unparseable url>"));
+    }
+
+    /// A multi-byte delimiter right before the URL must not break the scan.
+    #[test]
+    fn compact_error_text_handles_unicode_before_a_url() {
+        for raw in [
+            "compact failed: Service unavailable. See “https://u:TEST_PASSWORD@status.example/”.",
+            "compact failed: 请参阅：https://u:TEST_PASSWORD@status.example/",
+        ] {
+            let detail = super::user_facing_compact_error(raw);
+            assert!(!detail.contains("TEST_PASSWORD"), "{detail}");
+            assert!(detail.contains("https://status.example/"), "{detail}");
+        }
+    }
+
    use super::*;
    use reqwest::StatusCode;

@@ -1518,6 +1518,9 @@ mod load_user_prompts_tests;
 #[path = "acp_session_tests/memory_config_tests.rs"]
 mod memory_config_tests;
 #[cfg(test)]
+#[path = "acp_session_tests/model_switch_window_tests.rs"]
+mod model_switch_window_tests;
+#[cfg(test)]
 #[path = "acp_session_tests/parallel_dispatch_tests.rs"]
 mod parallel_dispatch_tests;
 #[cfg(test)]
@@ -10,7 +10,7 @@ impl SessionActor {
        apply_prompt_override: bool,
        skip_prompt_rewrite: bool,
        auto_compact_threshold_percent: u8,
-    ) -> Result<acp::ModelId, acp::Error> {
+    ) -> Result<(acp::ModelId, std::num::NonZeroU64), acp::Error> {
        let model_id = acp::ModelId::new(sampling_config.model.clone());
        // H4: record the picker's catalog KEY as this SESSION's own selection.
        // `sampling_config.model` is the ambiguous bare slug; the key is what
@@ -120,7 +120,7 @@ impl SessionActor {
                agent_name: Some(agent_name),
                reasoning_effort: Some(sampling_config.reasoning_effort),
            });
-        Ok(model_id)
+        Ok((model_id, new_context_window))
    }
    /// Handle [`SessionCommand::RebuildAgentForDefinition`].
    ///
@@ -548,17 +548,10 @@ impl SessionActor {

        if report.changed() && !dry_run {
            // Flush barrier: success must mean the rewrite is on disk.
-            let (flush_tx, flush_rx) = tokio::sync::oneshot::channel();
-            if self
-                .notifications
-                .persistence_tx
-                .send(PersistenceMsg::FlushAndAck {
-                    respond_to: flush_tx,
-                })
-                .is_err()
-                || flush_rx.await.is_err()
-            {
-                anyhow::bail!("history repaired in memory but the persistence flush failed");
+            if let Err(error) = self.persistence_barrier().await {
+                anyhow::bail!(
+                    "history repaired in memory but the persistence flush failed: {error}"
+                );
            }
            tracing::warn!(
                session_id = %self.session_info.id.0,
@@ -112,6 +112,7 @@ pub(crate) async fn spawn_session_actor(
    system_prompt_label: String,
    compaction_mode: kigi_chat_state::CompactionMode,
    compaction_verbatim_input: bool,
+    compaction_tool_choice: crate::util::config::CompactionToolChoice,
    two_pass_enabled: bool,
    buffering_settings: Option<BufferingSettings>,
    origin_client: Option<crate::http::OriginClientInfo>,
@@ -1032,6 +1033,7 @@ pub(crate) async fn spawn_session_actor(
            previous_model: std::cell::Cell::new(None),
            compaction_mode,
            verbatim_input: compaction_verbatim_input,
+            tool_choice: compaction_tool_choice,
            prefire: crate::session::compaction_config::PrefireState::default(),
            prefix_released: std::sync::atomic::AtomicBool::new(false),
        },
@@ -1535,6 +1537,7 @@ pub(crate) async fn spawn_session_on_thread(
    system_prompt_label: String,
    compaction_mode: kigi_chat_state::CompactionMode,
    compaction_verbatim_input: bool,
+    compaction_tool_choice: crate::util::config::CompactionToolChoice,
    two_pass_enabled: bool,
    buffering_settings: Option<BufferingSettings>,
    origin_client: Option<crate::http::OriginClientInfo>,
@@ -1636,10 +1639,23 @@ pub(crate) async fn spawn_session_on_thread(
                };
                (initial_last_compaction, initial_prompt_texts)
            };
-            let rt = tokio::runtime::Builder::new_current_thread()
+            // Runtime construction acquires fds (epoll/kqueue, waker) and
+            // fails with EMFILE/EAGAIN under resource pressure — report
+            // through the init channel instead of panicking the thread.
+            let rt = match tokio::runtime::Builder::new_current_thread()
                .enable_all()
                .build()
-                .expect("session runtime");
+            {
+                Ok(rt) => rt,
+                Err(e) => {
+                    tracing::error!(
+                        error = %e,
+                        "failed to build session runtime (resource exhaustion?)"
+                    );
+                    let _ = init_tx.send(Err(kigi_agent::AgentBuildError::RuntimeBuild(e)));
+                    return;
+                }
+            };
            let local = tokio::task::LocalSet::new();
            local.block_on(&rt, async move {
                let _trace_span = parent_traceparent.as_ref().map(|tp| {
@@ -1679,6 +1695,7 @@ pub(crate) async fn spawn_session_on_thread(
                    system_prompt_label,
                    compaction_mode,
                    compaction_verbatim_input,
+                    compaction_tool_choice,
                    two_pass_enabled,
                    buffering_settings,
                    origin_client,
@@ -1759,8 +1776,19 @@ pub(crate) async fn spawn_session_on_thread(
                }));
                let _ = session_done_rx.await;
            });
-        })
-        .expect("spawn session thread");
+        });
+    let join_handle = match join_handle {
+        Ok(handle) => handle,
+        Err(e) => {
+            tracing::error!(
+                error = %e,
+                "failed to spawn session thread (thread/PID limit or memory pressure?)"
+            );
+            return Err(
+                acp::Error::internal_error().data(format!("failed to spawn session thread: {e}"))
+            );
+        }
+    };
    let init = init_rx
        .await
        .map_err(|_| {
@@ -957,6 +957,8 @@ impl SessionActor {
                        Some(self.session_info.id.0.to_string()),
                        None,
                        None,
+                        self.tool_context.cwd.clone(),
+                        self.display_cwd.get().map(std::path::PathBuf::from),
                    )
                    .await
            };
@@ -681,22 +681,14 @@ impl SessionActor {
                    .await
                    .is_some()
                {
-                    let (flush_tx, flush_rx) = oneshot::channel();
-                    if self
-                        .notifications
-                        .persistence_tx
-                        .send(PersistenceMsg::FlushAndAck {
-                            respond_to: flush_tx,
-                        })
-                        .is_ok()
-                        && flush_rx.await.is_ok()
-                    {
-                        let _ = ack.send(());
-                    } else {
-                        tracing::error!(
-                            session_id = % self.session_info.id.0, prompt_id = %
-                            prompt_id, "persist_ack flush barrier failed"
-                        );
+                    match self.persistence_barrier().await {
+                        Ok(()) => {
+                            let _ = ack.send(());
+                        }
+                        Err(error) => tracing::error!(
+                            session_id = % self.session_info.id.0, prompt_id = % prompt_id,
+                            %error, "persist_ack flush barrier failed"
+                        ),
                    }
                } else {
                    tracing::error!(
@@ -333,15 +333,24 @@ impl SessionActor {
        if let Err(e) = crate::session::replay_events::flush_replay_actor(&self.event_tx).await {
            tracing::warn!(?e, "flush_replay_actor failed");
        }
+        if let Err(error) = self.persistence_barrier().await {
+            tracing::error!(%error, "persistence flush failed");
+        }
+    }
+    /// The `FlushAndAck` durability barrier: `Ok` only once every prior
+    /// persistence write is on stable media.
+    pub(super) async fn persistence_barrier(&self) -> std::io::Result<()> {
        let (tx, rx) = tokio::sync::oneshot::channel();
        if self
            .notifications
            .persistence_tx
            .send(PersistenceMsg::FlushAndAck { respond_to: tx })
-            .is_ok()
+            .is_err()
        {
-            let _ = rx.await;
+            return Err(std::io::Error::other("persistence actor gone"));
        }
+        rx.await
+            .unwrap_or_else(|_| Err(std::io::Error::other("persistence actor gone")))
    }
    /// Extracts the update type name and relevant parameters for logging
    fn extract_update_info(
@@ -157,6 +157,7 @@ async fn persist_ack_waits_for_disk_flush_before_success() {
                    previous_model: std::cell::Cell::new(None),
                    compaction_mode: kigi_chat_state::CompactionMode::Transcript,
                    verbatim_input: true,
+                    tool_choice: Default::default(),
                    prefire: crate::session::compaction_config::PrefireState::default(),
                    prefix_released: std::sync::atomic::AtomicBool::new(false),
                },
@@ -434,7 +435,7 @@ async fn first_turn_memory_injection_persists_to_chat_history() {
                    respond_to: flush_tx,
                })
                .unwrap();
-            flush_rx.await.unwrap();
+            flush_rx.await.unwrap().unwrap();
            let loaded = storage
                .load_session_without_updates(&session_info)
                .await
@@ -611,6 +612,7 @@ async fn first_turn_memory_injection_disabled_does_not_persist_to_chat_history()
                    previous_model: std::cell::Cell::new(None),
                    compaction_mode: kigi_chat_state::CompactionMode::Transcript,
                    verbatim_input: true,
+                    tool_choice: Default::default(),
                    prefire: crate::session::compaction_config::PrefireState::default(),
                    prefix_released: std::sync::atomic::AtomicBool::new(false),
                },
@@ -755,7 +757,7 @@ async fn first_turn_memory_injection_disabled_does_not_persist_to_chat_history()
                    respond_to: flush_tx,
                })
                .unwrap();
-            flush_rx.await.unwrap();
+            flush_rx.await.unwrap().unwrap();
            let storage = crate::session::storage::JsonlStorageAdapter::with_explicit_session_dir(
                session_dir.path().to_path_buf(),
            );
@@ -871,6 +873,7 @@ async fn cancel_running_task_teardown_clears_running_and_pending_work() {
                    previous_model: std::cell::Cell::new(None),
                    compaction_mode: kigi_chat_state::CompactionMode::Transcript,
                    verbatim_input: true,
+                    tool_choice: Default::default(),
                    prefire: crate::session::compaction_config::PrefireState::default(),
                    prefix_released: std::sync::atomic::AtomicBool::new(false),
                },
@@ -1307,7 +1310,7 @@ async fn actor_with_persistence_drain() -> std::sync::Arc<SessionActor> {
    tokio::task::spawn_local(async move {
        while let Some(msg) = persistence_rx.recv().await {
            if let PersistenceMsg::FlushAndAck { respond_to } = msg {
-                let _ = respond_to.send(());
+                let _ = respond_to.send(Ok(()));
            }
        }
    });
@@ -1865,6 +1868,7 @@ async fn cancel_propagates_to_sampler_handle_so_no_further_emission() {
                    previous_model: std::cell::Cell::new(None),
                    compaction_mode: kigi_chat_state::CompactionMode::Transcript,
                    verbatim_input: true,
+                    tool_choice: Default::default(),
                    prefire: crate::session::compaction_config::PrefireState::default(),
                    prefix_released: std::sync::atomic::AtomicBool::new(false),
                },
@@ -176,6 +176,7 @@ async fn test_e2e_idle_resume_refreshes_model_metadata() {
                    previous_model: std::cell::Cell::new(None),
                    compaction_mode: kigi_chat_state::CompactionMode::Transcript,
                    verbatim_input: true,
+                    tool_choice: Default::default(),
                    prefire: crate::session::compaction_config::PrefireState::default(),
                    prefix_released: std::sync::atomic::AtomicBool::new(false),
                },
@@ -112,6 +112,7 @@ async fn create_test_actor(
            previous_model: std::cell::Cell::new(None),
            compaction_mode: kigi_chat_state::CompactionMode::Transcript,
            verbatim_input: true,
+            tool_choice: Default::default(),
            prefire: crate::session::compaction_config::PrefireState::default(),
            prefix_released: std::sync::atomic::AtomicBool::new(false),
        },
@@ -539,6 +540,7 @@ async fn create_test_actor_with_memory(
            previous_model: std::cell::Cell::new(None),
            compaction_mode: kigi_chat_state::CompactionMode::Transcript,
            verbatim_input: true,
+            tool_choice: Default::default(),
            prefire: crate::session::compaction_config::PrefireState::default(),
            prefix_released: std::sync::atomic::AtomicBool::new(false),
        },
@@ -1298,6 +1300,7 @@ async fn test_e2e_idle_resume_refreshes_model_metadata() {
                    previous_model: std::cell::Cell::new(None),
                    compaction_mode: kigi_chat_state::CompactionMode::Transcript,
                    verbatim_input: true,
+                    tool_choice: Default::default(),
                    prefire: crate::session::compaction_config::PrefireState::default(),
                    prefix_released: std::sync::atomic::AtomicBool::new(false),
                },
@@ -163,6 +163,7 @@ async fn create_test_actor_with_memory(
            previous_model: std::cell::Cell::new(None),
            compaction_mode: kigi_chat_state::CompactionMode::Transcript,
            verbatim_input: true,
+            tool_choice: Default::default(),
            prefire: crate::session::compaction_config::PrefireState::default(),
            prefix_released: std::sync::atomic::AtomicBool::new(false),
        },
@@ -0,0 +1,24 @@
+//! `handle_set_session_model` reports the session's effective context
+//! window so clients repaint their header without waiting for a turn.
+use super::support::*;
+
+#[tokio::test(flavor = "current_thread")]
+async fn set_session_model_returns_the_effective_context_window() {
+    let local = tokio::task::LocalSet::new();
+    local
+        .run_until(async {
+            let (actor, _rx) = build_actor().await;
+            let cfg = kigi_sampler::SamplerConfig {
+                model: "narrow".to_string(),
+                context_window: 128_000,
+                ..Default::default()
+            };
+            let (model_id, window) = actor
+                .handle_set_session_model(cfg, None, false, false, true, 85)
+                .await
+                .expect("model switch");
+            assert_eq!(model_id.0.as_ref(), "narrow");
+            assert_eq!(window.get(), 128_000);
+        })
+        .await;
+}
@@ -82,6 +82,8 @@ async fn set_auto_mode_path_wires_live_side_query_via_session_actor() {
                    None,
                    None,
                    None,
+                    session.tool_context.cwd.clone(),
+                    None,
                )
                .await;
            // cargo is heuristic-allow when sampling fails; must not be Prompt-only
@@ -99,6 +101,8 @@ async fn set_auto_mode_path_wires_live_side_query_via_session_actor() {
                    None,
                    None,
                    None,
+                    session.tool_context.cwd.clone(),
+                    None,
                )
                .await;
            assert!(
@@ -116,6 +116,7 @@ pub(super) async fn make_replay_send_update_fixture() -> ReplaySendUpdateFixture
            previous_model: std::cell::Cell::new(None),
            compaction_mode: kigi_chat_state::CompactionMode::Transcript,
            verbatim_input: true,
+            tool_choice: Default::default(),
            prefire: crate::session::compaction_config::PrefireState::default(),
            prefix_released: std::sync::atomic::AtomicBool::new(false),
        },
@@ -226,6 +226,7 @@ pub(crate) async fn create_test_actor_ex(
            previous_model: std::cell::Cell::new(None),
            compaction_mode: kigi_chat_state::CompactionMode::Transcript,
            verbatim_input: true,
+            tool_choice: Default::default(),
            prefire: crate::session::compaction_config::PrefireState::default(),
            prefix_released: std::sync::atomic::AtomicBool::new(false),
        },
@@ -173,7 +173,9 @@ pub enum SessionCommand {
        /// `compaction.threshold_percent` (which is `Cell<u8>` so it can
        /// update without `&mut self`).
        auto_compact_threshold_percent: u8,
-        responds_to: oneshot::Sender<Result<acp::ModelId, acp::Error>>,
+        /// Ok: the applied model id and the session's effective context
+        /// window, which clients repaint from without waiting for a turn.
+        responds_to: oneshot::Sender<Result<(acp::ModelId, std::num::NonZeroU64), acp::Error>>,
    },
    /// Zero-turn harness rebuild: build a brand-new `Agent` from the
    /// session's `AgentRebuildSpec` and the new `AgentDefinition`,
@@ -332,10 +334,10 @@ pub enum SessionCommand {
    CopyFile {
        respond_to: oneshot::Sender<anyhow::Result<crate::session::persistence::SessionStateCopy>>,
    },
-    /// Flush the replay buffer and persistence, then signal completion.
-    /// Used during reconnect to ensure all buffered content is persisted before replay.
+    /// Flush the replay buffer and persistence, then report the durability
+    /// barrier result. Used during reconnect so replay follows persisted content.
    FlushComplete {
-        respond_to: oneshot::Sender<()>,
+        respond_to: oneshot::Sender<std::io::Result<()>>,
    },
    /// Update MCP servers for an existing session (used during reconnect or
    /// mid-session via the `kigi/session/update_mcp_servers` extension method).
@@ -212,6 +212,7 @@ impl SessionActor {
            &sampling_config,
            self.inference_idle_timeout,
            wall_clock_budget_secs,
+            self.compaction.tool_choice,
        )
        .await
        {
@@ -688,6 +689,19 @@ impl SessionActor {
            .await;
        }
    }
+    /// Retry guidance headline, then the user-facing detail on its own line.
+    fn failure_with_retry_guidance(err: &acp::Error) -> String {
+        const RETRY_GUIDANCE: &str =
+            "it'll retry on the next turn, or start a new session using /new.";
+        let detail = crate::sampling::error::user_facing_compact_error(
+            &crate::sampling::error::acp_error_message(err),
+        );
+        if detail.is_empty() {
+            RETRY_GUIDANCE.to_string()
+        } else {
+            format!("{RETRY_GUIDANCE}\n{detail}")
+        }
+    }
    /// Map a deterministic failure's error text to a fixed, content-free
    /// [`SuppressReason`] (drives telemetry + sticky-vs-per-turn scope).
    fn classify_suppress_reason(error_msg: &str) -> SuppressReason {
@@ -983,6 +997,7 @@ impl SessionActor {
            sampling_config.clone(),
            self.inference_idle_timeout,
            wall_clock_budget_secs,
+            self.compaction.tool_choice,
        );
        let observer =
            crate::session::helpers::full_replace_compaction::ShellFullReplaceObserver::new(
@@ -1943,7 +1958,7 @@ impl SessionActor {
                    == SUPPRESS_NONE
                {
                    self.send_xai_notification(XaiSessionUpdate::AutoCompactFailed {
-                        error: String::new(),
+                        error: Self::failure_with_retry_guidance(&e),
                    })
                    .await;
                }
@@ -2198,6 +2213,7 @@ mod inline_auto_compact_flow_tests {
                previous_model: std::cell::Cell::new(None),
                compaction_mode: kigi_chat_state::CompactionMode::Transcript,
                verbatim_input: true,
+                tool_choice: Default::default(),
                prefire: crate::session::compaction_config::PrefireState::default(),
                prefix_released: std::sync::atomic::AtomicBool::new(false),
            },
@@ -3283,3 +3299,28 @@ mod inline_auto_compact_flow_tests {
            .await;
    }
 }
+
+#[cfg(test)]
+mod failure_message_tests {
+    use super::SessionActor;
+    use agent_client_protocol as acp;
+
+    #[test]
+    fn transient_failure_message_carries_guidance_and_the_decoded_detail() {
+        let err = acp::Error::internal_error().data(
+            "Compact failed: compact failed: HTTP 500 from upstream\n\nRequest URL: http://u:TEST_PASSWORD@host/v1",
+        );
+        let text = SessionActor::failure_with_retry_guidance(&err);
+        let (headline, detail) = text.split_once('\n').expect("guidance + detail");
+        assert_eq!(
+            headline,
+            "it'll retry on the next turn, or start a new session using /new."
+        );
+        assert_eq!(detail, "HTTP 500 from upstream");
+        assert!(!text.contains("TEST_PASSWORD"));
+        assert_eq!(
+            SessionActor::failure_with_retry_guidance(&acp::Error::internal_error().data("   ")),
+            "it'll retry on the next turn, or start a new session using /new."
+        );
+    }
+}
@@ -146,6 +146,9 @@ pub struct CompactionConfig {
    pub compaction_mode: kigi_chat_state::CompactionMode,
    /// When `true`, feed the summarizer the verbatim conversation instead of the lossy rewrite (the retry loop may still fall back).
    pub verbatim_input: bool,
+    /// `tool_choice` for summarizer requests with tools attached (`auto` |
+    /// `none`); some backends reject `"none"`.
+    pub tool_choice: crate::util::config::CompactionToolChoice,
    /// Prefire two-pass state (background NOTE₁ cache + in-flight guard).
    /// `Default` (empty cache, not in-flight).
    pub prefire: PrefireState,
@@ -519,6 +519,9 @@ pub(crate) async fn capture_plan_changes(
    }
 }

+/// Pin the `a/` `b/` header prefixes [`extract_changed_files`] parses.
+const DIFF_PREFIX_ARGS: [&str; 2] = ["--src-prefix=a/", "--dst-prefix=b/"];
+
 /// Run `git diff <baseline>` and return the FULL (untruncated) stdout;
 /// the caller extracts the changed-file list before truncating.
 async fn run_git_diff_against_baseline(
@@ -526,7 +529,7 @@ async fn run_git_diff_against_baseline(
    workspace_root: &Path,
 ) -> Result<String, ChangesCaptureError> {
    let mut cmd = git_command(workspace_root);
-    cmd.arg("diff").arg(baseline);
+    cmd.arg("diff").args(DIFF_PREFIX_ARGS).arg(baseline);
    let output = match tokio::time::timeout(DIFF_COMMAND_TIMEOUT, cmd.output()).await {
        Ok(Ok(output)) => output,
        Ok(Err(err)) => return Err(ChangesCaptureError::DiffCommandFailed(err.to_string())),
@@ -579,10 +582,14 @@ async fn lazy_git_baseline_diff(
    // support SHA-256 repos.
    let mut cmd = git_command(workspace_root);
    if git_has_parent(workspace_root, &oldest).await {
-        cmd.arg("diff").arg(format!("{oldest}^..{head}"));
+        cmd.arg("diff")
+            .args(DIFF_PREFIX_ARGS)
+            .arg(format!("{oldest}^..{head}"));
    } else {
        let empty_tree = derive_empty_tree_sha(workspace_root).await;
-        cmd.arg("diff").arg(format!("{empty_tree}..{head}"));
+        cmd.arg("diff")
+            .args(DIFF_PREFIX_ARGS)
+            .arg(format!("{empty_tree}..{head}"));
    }

    let output = match tokio::time::timeout(DIFF_COMMAND_TIMEOUT, cmd.output()).await {
@@ -2192,4 +2199,38 @@ mod tests {
            "patch must carry the untracked-files note",
        );
    }
+    #[tokio::test]
+    async fn changed_files_ignore_the_host_diff_prefix_config() {
+        // Both keys rewrite the `diff --git a/<old> b/<new>` header the parser keys on.
+        for key in ["diff.mnemonicprefix", "diff.noprefix"] {
+            let tmp = tempfile::tempdir().unwrap();
+            init_repo(tmp.path());
+            git(tmp.path(), &["config", key, "true"]);
+            tokio::fs::write(tmp.path().join("tracked.txt"), b"seed\n")
+                .await
+                .unwrap();
+            let goal_created_at: i64 = 1_700_000_100;
+            git_at(tmp.path(), &["add", "."], goal_created_at + 100);
+            git_at(
+                tmp.path(),
+                &["commit", "-q", "-m", "post-goal"],
+                goal_created_at + 100,
+            );
+            let lazy = capture_changes_diff(None, tmp.path(), goal_created_at)
+                .await
+                .expect("lazy baseline must succeed");
+            assert_eq!(lazy.changed_files, ["tracked.txt"], "{key}: lazy baseline");
+            tokio::fs::write(tmp.path().join("tracked.txt"), b"changed\n")
+                .await
+                .unwrap();
+            let explicit = capture_changes_diff(Some("HEAD"), tmp.path(), goal_created_at)
+                .await
+                .expect("layer-1 git diff must succeed");
+            assert_eq!(
+                explicit.changed_files,
+                ["tracked.txt"],
+                "{key}: explicit baseline"
+            );
+        }
+    }
 }
@@ -70,6 +70,8 @@ pub(crate) struct ShellCompactionSampler {
    /// Wall-clock budget (secs) forwarded to `generate_session_compact` as the
    /// reasoning-runaway backstop; `0` disables it.
    wall_clock_budget_secs: u64,
+    /// Summarizer `tool_choice` forwarded to `generate_session_compact`.
+    tool_choice: crate::util::config::CompactionToolChoice,
    /// Full output of the most recent successful sample (for L5 telemetry).
    last_success: Mutex<Option<CompactOutput>>,
 }
@@ -86,6 +88,7 @@ impl ShellCompactionSampler {
        sampling_config: SamplingConfig,
        idle_timeout: Duration,
        wall_clock_budget_secs: u64,
+        tool_choice: crate::util::config::CompactionToolChoice,
    ) -> Self {
        Self {
            use_short_prompt,
@@ -97,6 +100,7 @@ impl ShellCompactionSampler {
            sampling_config,
            idle_timeout,
            wall_clock_budget_secs,
+            tool_choice,
            last_success: Mutex::new(None),
        }
    }
@@ -135,6 +139,7 @@ impl CompactionSampler for ShellCompactionSampler {
            &self.sampling_config,
            self.idle_timeout,
            self.wall_clock_budget_secs,
+            self.tool_choice,
        )
        .await
        {
@@ -339,9 +339,10 @@ where
 /// prefix and force a full prefill on the summarizer call — attaching them
 /// keeps the request prefix byte-identical to the turn requests so the
 /// engine reuses the session's KV cache (the whole point of the verbatim
-/// input path). Tool *use* is forbidden via `tool_choice: none` where the
-/// backend can express it (ChatCompletions, Responses); the Messages wire
-/// enum has no `none`, so that path relies on the prompt instruction alone.
+/// input path). Tool *use* is forbidden by the prompt instruction;
+/// `tool_choice` says what goes on the wire next to the attached tools
+/// (ChatCompletions, Responses — some servers reject `"none"`, hence the
+/// `auto` default). The Messages wire enum has no `none` either way.
 ///
 /// Errors carry a [`CompactFailure`] classification so the caller can
 /// short-circuit retries on deterministic failures (4xx schema violations,
@@ -356,8 +357,17 @@ pub(crate) async fn generate_session_compact(
    sampling_config: &SamplingConfig,
    idle_timeout: std::time::Duration,
    wall_clock_budget_secs: u64,
+    tool_choice: crate::util::config::CompactionToolChoice,
 ) -> Result<CompactOutput, CompactFailure> {
    let num_messages = chat_history.len();
+    let wire_tool_choice = match tool_choice {
+        crate::util::config::CompactionToolChoice::Auto => ToolChoice::auto(),
+        crate::util::config::CompactionToolChoice::None => ToolChoice::none(),
+    };
+    let conversation_tool_choice = match tool_choice {
+        crate::util::config::CompactionToolChoice::Auto => ConversationToolChoice::Auto,
+        crate::util::config::CompactionToolChoice::None => ConversationToolChoice::None,
+    };
    let output = match sampling_config.api_backend {
        ApiBackend::ChatCompletions => {
            let chat_messages: Vec<ChatRequestMessage> =
@@ -373,7 +383,7 @@ pub(crate) async fn generate_session_compact(
                            .map(|t| ToolDefinition::function(t.name, t.description, t.parameters))
                            .collect(),
                    )
-                    .with_tool_choice(ToolChoice::none());
+                    .with_tool_choice(wire_tool_choice);
            }
            let sid = session_id.to_string();
            message.x_kigi_conv_id = Some(sid.clone());
@@ -466,7 +476,10 @@ pub(crate) async fn generate_session_compact(
        ApiBackend::Responses => {
            let request = ConversationRequest {
                items: chat_history,
-                tool_choice: (!tools.is_empty()).then_some(ConversationToolChoice::None),
+                // Hosted tools count: a hosted-only request with the choice
+                // omitted would leave the hosted tool callable mid-compact.
+                tool_choice: (!tools.is_empty() || !hosted_tools.is_empty())
+                    .then_some(conversation_tool_choice),
                tools,
                hosted_tools,
                model: Some(sampling_config.model.to_owned()),
@@ -1573,6 +1586,7 @@ mod reasoning_compaction_regression_tests {
            &config,
            std::time::Duration::from_secs(30),
            0,
+            crate::util::config::CompactionToolChoice::Auto,
        )
        .await
        .unwrap_or_else(|_| panic!("compaction must succeed"));
@@ -1661,6 +1675,7 @@ mod reasoning_compaction_regression_tests {
            &config,
            std::time::Duration::from_secs(30),
            0,
+            crate::util::config::CompactionToolChoice::Auto,
        )
        .await;
        let output = result
@@ -1669,11 +1684,11 @@ mod reasoning_compaction_regression_tests {
        let _ = shutdown_tx.send(());
    }
    /// The compaction request must carry the turn loop's tool definitions
-    /// (prompt-prefix/KV-cache alignment) with `tool_choice: "none"`, and
-    /// must omit both keys when no tools are passed (Chat Completions rejects a bare
-    /// `tool_choice`).
+    /// (prompt-prefix/KV-cache alignment) with the configured `tool_choice`
+    /// (`auto` default; `none` when configured), and must omit both keys when
+    /// no tools are passed (Chat Completions rejects a bare `tool_choice`).
    #[tokio::test]
-    async fn chat_completions_compaction_attaches_tools_with_tool_choice_none() {
+    async fn chat_completions_compaction_attaches_tools_with_configured_tool_choice() {
        use std::sync::{Arc, Mutex};
        let captured: Arc<Mutex<Vec<serde_json::Value>>> = Arc::new(Mutex::new(Vec::new()));
        let cap = captured.clone();
@@ -1726,12 +1741,13 @@ mod reasoning_compaction_regression_tests {
            &config,
            std::time::Duration::from_secs(30),
            0,
+            crate::util::config::CompactionToolChoice::Auto,
        )
        .await
        .unwrap_or_else(|_| panic!("compaction with tools must succeed"));
        let client = Client::new(config.clone()).unwrap();
        generate_session_compact(
-            chat_history,
+            chat_history.clone(),
            vec![],
            vec![],
            client,
@@ -1739,16 +1755,36 @@ mod reasoning_compaction_regression_tests {
            &config,
            std::time::Duration::from_secs(30),
            0,
+            crate::util::config::CompactionToolChoice::Auto,
        )
        .await
        .unwrap_or_else(|_| panic!("compaction without tools must succeed"));
+        let tools = vec![ToolSpec {
+            name: "read_file".to_string(),
+            description: Some("Reads a file".to_string()),
+            parameters: json!({ "type" : "object", "properties" : {} }),
+        }];
+        let client = Client::new(config.clone()).unwrap();
+        generate_session_compact(
+            chat_history,
+            tools,
+            vec![],
+            client,
+            acp::SessionId::new("test-session"),
+            &config,
+            std::time::Duration::from_secs(30),
+            0,
+            crate::util::config::CompactionToolChoice::None,
+        )
+        .await
+        .unwrap_or_else(|_| panic!("compaction with tool_choice none must succeed"));
        let bodies = captured.lock().unwrap();
-        assert_eq!(bodies.len(), 2, "mock must have served both requests");
+        assert_eq!(bodies.len(), 3, "mock must have served all three requests");
        let with_tools = &bodies[0];
        assert_eq!(
            with_tools["tool_choice"],
-            json!("none"),
-            "tool use must be disabled at decode time"
+            json!("auto"),
+            "default tool_choice is auto — some backends reject none; the prompt forbids tool use"
        );
        let sent_tools = with_tools["tools"]
            .as_array()
@@ -1764,6 +1800,12 @@ mod reasoning_compaction_regression_tests {
            without_tools.get("tool_choice").is_none(),
            "tool_choice without tools is rejected by OpenAI-compat backends"
        );
+        let none_configured = &bodies[2];
+        assert_eq!(
+            none_configured["tool_choice"],
+            json!("none"),
+            "configured none must reach the wire"
+        );
        let _ = shutdown_tx.send(());
    }
    #[tokio::test]
@@ -1802,6 +1844,7 @@ mod reasoning_compaction_regression_tests {
            &config,
            std::time::Duration::from_millis(150),
            0,
+            crate::util::config::CompactionToolChoice::Auto,
        )
        .await;
        match result {
@@ -1869,6 +1912,7 @@ mod reasoning_compaction_regression_tests {
            &config,
            std::time::Duration::from_millis(150),
            0,
+            crate::util::config::CompactionToolChoice::Auto,
        )
        .await;
        match result {
@@ -1940,6 +1984,7 @@ mod reasoning_compaction_regression_tests {
            &config,
            std::time::Duration::from_millis(150),
            0,
+            crate::util::config::CompactionToolChoice::Auto,
        )
        .await;
        match result {
@@ -2008,6 +2053,7 @@ mod reasoning_compaction_regression_tests {
            &config,
            std::time::Duration::from_millis(150),
            0,
+            crate::util::config::CompactionToolChoice::Auto,
        )
        .await;
        match result {
@@ -467,7 +467,24 @@ fn compute_normalized_blocking(
            });
        }
    };
-    if buf.len() >= original_bytes {
+    // A resize that falls under the vision floors cannot be sent; keep the original with the fallback note.
+    if new_w < MIN_VISION_SIDE_PX
+        || new_h < MIN_VISION_SIDE_PX
+        || u64::from(new_w) * u64::from(new_h) < MIN_VISION_TOTAL_PX
+    {
+        tracing::warn!(
+            index,
+            width = new_w,
+            height = new_h,
+            "image resize fell under the minimum dimensions; keeping original attachment"
+        );
+        return Ok(NormalizedEntry::ReEncodingOversized {
+            bytes: Bytes::from(raw_bytes),
+            mime: Cow::Borrowed(orig_mime),
+        });
+    }
+    // An image over the dimension caps is resized even when that saves no bytes.
+    if !exceeded_dimensions && buf.len() >= original_bytes {
        return Ok(NormalizedEntry::Unchanged {
            bytes: Bytes::from(raw_bytes),
            mime: Cow::Borrowed(orig_mime),
@@ -551,6 +568,50 @@ mod tests {
            image::ImageFormat::Png
        );
    }
+    /// A flat image at the best PNG compression is tinier than any re-encode the fast encoder produces.
+    fn make_best_compressed_png_content(width: u32, height: u32) -> ImageContent {
+        use image::codecs::png::{CompressionType, FilterType, PngEncoder};
+        use image::{ImageBuffer, ImageEncoder, Rgba};
+        let img: ImageBuffer<Rgba<u8>, Vec<u8>> =
+            ImageBuffer::from_pixel(width, height, Rgba([128, 64, 32, 255]));
+        let mut buf = Vec::new();
+        PngEncoder::new_with_quality(&mut buf, CompressionType::Best, FilterType::Adaptive)
+            .write_image(img.as_raw(), width, height, image::ExtendedColorType::Rgba8)
+            .unwrap();
+        ImageContent::new(
+            base64::engine::general_purpose::STANDARD.encode(&buf),
+            "image/png",
+        )
+    }
+    /// Over the dimension caps, a re-encode that saves no bytes is still a resize.
+    #[tokio::test]
+    async fn over_dimension_image_is_resized_even_when_re_encode_saves_no_bytes() {
+        let img = make_best_compressed_png_content(3000, 2000);
+        let cache = fresh_cache();
+        match normalize_one_in(img, 1, false, &cache).await {
+            Outcome::Compressed { info, .. } => {
+                assert!(
+                    info.compressed_bytes >= info.original_bytes,
+                    "test premise: the re-encode must not shrink ({} vs {})",
+                    info.compressed_bytes,
+                    info.original_bytes
+                );
+                assert!(info.compressed_width <= MAX_ENCODE_SIDE_PX);
+                assert!(info.compressed_height <= MAX_ENCODE_SIDE_PX);
+            }
+            other => panic!("expected the oversized image to be downscaled, got {other:?}"),
+        }
+    }
+    /// A 3000x8 strip resizes to 2000x5, under the 8 px floor: it is kept with the fallback note.
+    #[tokio::test]
+    async fn resize_under_the_minimum_side_keeps_the_original() {
+        let img = make_best_compressed_png_content(3000, 8);
+        let cache = fresh_cache();
+        match normalize_one_in(img, 1, false, &cache).await {
+            Outcome::ReEncodingOversized(content) => assert_eq!(content.mime_type, "image/png"),
+            other => panic!("expected the strip to be kept with a fallback note, got {other:?}"),
+        }
+    }
    fn make_gif_content(width: u32, height: u32) -> ImageContent {
        use image::{ImageBuffer, Rgba};
        let img: ImageBuffer<Rgba<u8>, Vec<u8>> =
@@ -359,11 +359,13 @@ pub enum PersistenceMsg {
    /// stays sequential with other summary.json mutations.
    GeneratedTitle(String),
    Flush,
-    /// Flush all pending writes, then signal the caller once the flush is complete.
-    /// Unlike `Flush` (fire-and-forget), this is a **sync barrier**: the caller's
-    /// oneshot only resolves after `flush_pending()` finishes writing to disk.
+    /// Flush all pending writes AND fsync the session files, then signal the
+    /// caller. Unlike `Flush` (fire-and-forget, page cache only), this is a
+    /// **durability barrier**: `Ok` means every prior write is on stable
+    /// media; `Err` reports a storage write that failed since the last
+    /// barrier or a failed sync.
    FlushAndAck {
-        respond_to: tokio::sync::oneshot::Sender<()>,
+        respond_to: tokio::sync::oneshot::Sender<io::Result<()>>,
    },
    /// Flush all pending writes, then copy the current session directory contents and return
    /// the in-memory snapshot to the caller (who can tar.gz + upload to GCS, etc.).
@@ -1392,6 +1394,9 @@ struct SessionPersistence {
    storage: Arc<dyn StorageAdapter>,
    /// Pending ACP notification for merging consecutive text chunks
    pending_notification: Option<acp::SessionNotification>,
+    /// First storage failure since the last barrier: `FlushAndAck` must not
+    /// report `Ok` for state that never reached the file.
+    pending_write_error: Option<io::Error>,
    rx: mpsc::UnboundedReceiver<PersistenceMsg>,
    /// Session title generation lifecycle.
    summary: crate::session::summary::SummaryGenerator,
@@ -1491,29 +1496,49 @@ impl SessionPersistence {
        }
    }

-    async fn write_update(&mut self, update: &SessionUpdate) {
-        if let Err(e) = self.storage.append_update(&self.info, update).await {
-            tracing::warn!(?e, "failed to write update");
+    fn note_storage_failure(&mut self, error: &io::Error) {
+        if self.pending_write_error.is_none() {
+            self.pending_write_error = Some(io::Error::new(error.kind(), error.to_string()));
        }
    }

-    /// Flush any pending merged ACP notification to disk.
-    async fn flush_pending(&mut self) {
-        // Write any pending merged ACP notification
-        if let Some(notification) = self.pending_notification.take() {
-            self.write_update(&SessionUpdate::Acp(Box::new(notification)))
-                .await;
+    /// Append an update; a failure is logged and latched for the next barrier.
+    async fn write_update(&mut self, update: &SessionUpdate) -> io::Result<()> {
+        let result = self.storage.append_update(&self.info, update).await;
+        if let Err(error) = &result {
+            tracing::warn!(%error, "failed to write update");
+            self.note_storage_failure(error);
+        }
+        result
+    }
+
+    /// Write the pending merged ACP notification.
+    async fn flush_pending(&mut self) -> io::Result<()> {
+        match self.pending_notification.take() {
+            Some(notification) => {
+                self.write_update(&SessionUpdate::Acp(Box::new(notification)))
+                    .await
+            }
+            None => Ok(()),
        }
    }

-    /// Flush pending writes and sync all session files to disk.
-    /// Called before CopyFile to ensure all data is persisted.
-    async fn flush_and_sync(&mut self) {
-        self.flush_pending().await;
-        // Sync all session files to disk to ensure they're actually written
-        if let Err(e) = self.storage.sync_session_files(&self.info).await {
-            tracing::warn!(?e, "Failed to sync session files to disk");
+    /// Fsync every session file to stable media.
+    async fn sync_files(&self) -> io::Result<()> {
+        let synced = self.storage.sync_session_files(&self.info).await;
+        if let Err(error) = &synced {
+            tracing::warn!(%error, "failed to sync session files to disk");
        }
+        synced
+    }
+
+    /// The `FlushAndAck` barrier. First error wins: a write that failed since
+    /// the last barrier outranks a failed sync.
+    async fn flush_and_sync(&mut self) -> io::Result<()> {
+        let flushed = self.flush_pending().await;
+        let prior_write = self.pending_write_error.take().map_or(Ok(()), Err);
+        let synced = self.sync_files().await;
+        flushed.and(prior_write).and(synced)
    }

    async fn run(mut self) {
@@ -1530,34 +1555,38 @@ impl SessionPersistence {
            }
            match msg {
                PersistenceMsg::Flush => {
-                    self.flush_pending().await;
+                    // Logged and latched by write_update; the next barrier reports it.
+                    let _ = self.flush_pending().await;
                }
                PersistenceMsg::FlushAndAck { respond_to } => {
-                    self.flush_pending().await;
-                    let _ = respond_to.send(());
+                    let result = self.flush_and_sync().await;
+                    let _ = respond_to.send(result);
                }
                PersistenceMsg::Update(update) => {
                    match update {
                        SessionUpdate::Acp(notification) => {
                            // ACP notifications use merging to coalesce consecutive text chunks
                            if let Some(to_write) = self.maybe_merge_notification(&notification) {
-                                self.write_update(&SessionUpdate::Acp(Box::new(to_write)))
+                                // Logged and latched by write_update.
+                                let _ = self
+                                    .write_update(&SessionUpdate::Acp(Box::new(to_write)))
                                    .await;
                            }
                        }
                        SessionUpdate::Xai(_) => {
-                            // xAI notifications are written directly without merging
-                            self.write_update(&update).await;
+                            // Written directly without merging; logged and latched by write_update.
+                            let _ = self.write_update(&update).await;
                        }
                    }
                }
                PersistenceMsg::Chat(chat_msg) => {
-                    if let Err(e) = self
+                    if let Err(error) = self
                        .storage
                        .append_chat_message(&self.info, &chat_msg)
                        .await
                    {
-                        tracing::warn!(?e, "failed to write chat message");
+                        tracing::warn!(%error, "failed to write chat message");
+                        self.note_storage_failure(&error);
                    }
                }
                PersistenceMsg::ReplaceChatHistory(messages) => {
@@ -1571,6 +1600,7 @@ impl SessionPersistence {
                        .await
                    {
                        tracing::warn!(?e, "failed to replace chat history");
+                        self.note_storage_failure(&e);
                    }
                }
                PersistenceMsg::CurrentModel {
@@ -1589,21 +1619,25 @@ impl SessionPersistence {
                        .await
                    {
                        tracing::warn!(?e, "failed to update current model");
+                        self.note_storage_failure(&e);
                    }
                }
                PersistenceMsg::PlanState(state) => {
                    if let Err(e) = self.storage.write_plan_state(&self.info, &state).await {
                        tracing::warn!(?e, "failed to write plan state");
+                        self.note_storage_failure(&e);
                    }
                }
                PersistenceMsg::PlanModeState(state) => {
                    if let Err(e) = self.storage.write_plan_mode_state(&self.info, &state).await {
                        tracing::warn!(?e, "failed to write plan mode state");
+                        self.note_storage_failure(&e);
                    }
                }
                PersistenceMsg::GoalModeState(state) => {
                    if let Err(e) = self.storage.write_goal_mode_state(&self.info, &state).await {
                        tracing::warn!(?e, "failed to write goal mode state");
+                        self.note_storage_failure(&e);
                    }
                }
                PersistenceMsg::GraphModeState(state) => {
@@ -1613,6 +1647,7 @@ impl SessionPersistence {
                        .await
                    {
                        tracing::warn!(?e, "failed to write graph mode state");
+                        self.note_storage_failure(&e);
                    }
                }
                PersistenceMsg::ContentChunk(content_chunks) => {
@@ -1686,12 +1721,14 @@ impl SessionPersistence {
                        }
                        Err(e) => {
                            tracing::warn!(?e, "failed to persist generated session title");
+                            self.note_storage_failure(&e);
                        }
                    }
                }
                PersistenceMsg::RewindPoint(point) => {
-                    if let Err(e) = self.storage.append_rewind_point(&self.info, &point).await {
-                        tracing::warn!(?e, "failed to write rewind point");
+                    if let Err(error) = self.storage.append_rewind_point(&self.info, &point).await {
+                        tracing::warn!(%error, "failed to write rewind point");
+                        self.note_storage_failure(&error);
                    }
                }
                PersistenceMsg::TruncateRewindPoints { from_index } => {
@@ -1701,6 +1738,7 @@ impl SessionPersistence {
                        .await
                    {
                        tracing::warn!(?e, from_index, "failed to truncate rewind points");
+                        self.note_storage_failure(&e);
                    }
                }
                PersistenceMsg::MergeRewindPointsFrom { target_index } => {
@@ -1710,6 +1748,7 @@ impl SessionPersistence {
                        .await
                    {
                        tracing::warn!(?e, target_index, "failed to merge rewind points");
+                        self.note_storage_failure(&e);
                    }
                }
                PersistenceMsg::CollectionId(collection_id) => {
@@ -1719,6 +1758,7 @@ impl SessionPersistence {
                        .await
                    {
                        tracing::warn!(?e, "failed to write collection id");
+                        self.note_storage_failure(&e);
                    }
                }
                PersistenceMsg::NextTraceTurn {
@@ -1731,11 +1771,13 @@ impl SessionPersistence {
                        .await
                    {
                        tracing::warn!(?e, "failed to write next trace turn");
+                        self.note_storage_failure(&e);
                    }
                }
                PersistenceMsg::Signals(signals) => {
                    if let Err(e) = self.storage.write_signals(&self.info, &signals).await {
                        tracing::warn!(?e, "failed to write session signals");
+                        self.note_storage_failure(&e);
                    }
                }
                PersistenceMsg::AnnouncementState(state) => {
@@ -1745,16 +1787,19 @@ impl SessionPersistence {
                        .await
                    {
                        tracing::warn!(?e, "failed to write announcement state");
+                        self.note_storage_failure(&e);
                    }
                }
                PersistenceMsg::Feedback(entry) => {
                    if let Err(e) = self.storage.append_feedback(&self.info, &entry).await {
                        tracing::warn!(?e, "failed to write feedback entry");
+                        self.note_storage_failure(&e);
                    }
                }
                PersistenceMsg::Btw(entry) => {
                    if let Err(e) = self.storage.append_btw(&self.info, &entry).await {
                        tracing::warn!(?e, "failed to write btw entry");
+                        self.note_storage_failure(&e);
                    }
                }
                PersistenceMsg::GitHead { commit, branch } => {
@@ -1764,6 +1809,7 @@ impl SessionPersistence {
                        .await
                    {
                        tracing::warn!(?e, "failed to persist git HEAD");
+                        self.note_storage_failure(&e);
                    }
                }
                PersistenceMsg::CompactionCheckpoint(checkpoint) => {
@@ -1773,6 +1819,7 @@ impl SessionPersistence {
                        .await
                    {
                        tracing::warn!(?e, "failed to write compaction checkpoint file");
+                        self.note_storage_failure(&e);
                    }
                }
                PersistenceMsg::CompactionRequest(request) => {
@@ -1782,11 +1829,13 @@ impl SessionPersistence {
                        .await
                    {
                        tracing::warn!(?e, "failed to write compaction request artifact");
+                        self.note_storage_failure(&e);
                    }
                }
                PersistenceMsg::RecapRequest(request) => {
                    if let Err(e) = self.storage.write_recap_request(&self.info, &request).await {
                        tracing::warn!(?e, "failed to write recap request artifact");
+                        self.note_storage_failure(&e);
                    }
                }
                PersistenceMsg::CompactionSegment(segment) => {
@@ -1796,20 +1845,25 @@ impl SessionPersistence {
                        .await
                    {
                        tracing::warn!(?e, "failed to write compaction segment");
+                        self.note_storage_failure(&e);
                    }
                }
                PersistenceMsg::CopyFile { one_shot } => {
-                    // Flush pending writes and sync all session files to disk before copying.
-                    self.flush_and_sync().await;
-
-                    let result = self.copy_session_dir_to_memory().await;
+                    // Not the durability barrier: the write-failure latch stays for FlushAndAck.
+                    let flushed = self.flush_pending().await;
+                    let result = match flushed.and(self.sync_files().await) {
+                        Ok(()) => self.copy_session_dir_to_memory().await,
+                        Err(error) => {
+                            Err(anyhow::Error::new(error).context("session files are not on disk"))
+                        }
+                    };
                    let _ = one_shot.send(result);
                }
            }
        }

-        // Drain the merge buffer on channel close.
-        self.flush_pending().await;
+        // Drain the merge buffer on channel close; a failure is logged by write_update.
+        let _ = self.flush_pending().await;
    }

    async fn copy_session_dir_to_memory(&self) -> anyhow::Result<SessionStateCopy> {
@@ -2002,6 +2056,7 @@ pub(crate) async fn new(
            info: info_clone,
            storage: storage.clone(),
            pending_notification: None,
+            pending_write_error: None,
            rx,
            summary: crate::session::summary::SummaryGenerator::new(
                crate::session::summary::SummaryConfig {
@@ -2070,6 +2125,7 @@ pub async fn new_with_explicit_dir(
            info: info_clone,
            storage: storage.clone(),
            pending_notification: None,
+            pending_write_error: None,
            rx,
            summary: crate::session::summary::SummaryGenerator::new(
                crate::session::summary::SummaryConfig {
@@ -2181,6 +2237,7 @@ pub(crate) async fn load_light(
            info: loaded_info,
            storage: storage.clone(),
            pending_notification: None,
+            pending_write_error: None,
            rx,
            summary: summary_gen,
            registry_title_sync,
@@ -3625,3 +3682,120 @@ mod repo_wide_resolution_tests {
        );
    }
 }
+
+#[cfg(test)]
+mod barrier_tests {
+    use super::*;
+
+    fn sampling_client() -> crate::sampling::Client {
+        crate::sampling::Client::new(kigi_sampler::SamplerConfig {
+            api_key: Some("test-key".to_string()),
+            base_url: "http://localhost".to_string(),
+            model: "test".to_string(),
+            max_completion_tokens: None,
+            temperature: None,
+            top_p: None,
+            api_backend: Default::default(),
+            chat_compat: Default::default(),
+            auth_scheme: Default::default(),
+            anthropic_oauth: false,
+            github_copilot: false,
+            openai_codex: false,
+            extra_headers: Default::default(),
+            context_window: 100_000,
+            force_http1: false,
+            max_retries: None,
+            stream_tool_calls: false,
+            idle_timeout_secs: None,
+            reasoning_effort: None,
+            origin_client: None,
+            attribution_callback: None,
+            bearer_resolver: None,
+            supports_backend_search: false,
+            compactions_remaining: None,
+            compaction_at_tokens: None,
+            doom_loop_recovery: None,
+            header_injector: None,
+        })
+        .expect("sampling client should build")
+    }
+
+    async fn barrier(handle: &PersistenceHandle) -> io::Result<()> {
+        let (tx, rx) = tokio::sync::oneshot::channel();
+        handle
+            .tx
+            .send(PersistenceMsg::FlushAndAck { respond_to: tx })
+            .unwrap();
+        rx.await.unwrap()
+    }
+
+    async fn spawn(dir: &std::path::Path) -> (Info, PersistenceHandle) {
+        let info = Info {
+            id: acp::SessionId::new("barrier"),
+            cwd: dir.to_string_lossy().into_owned(),
+        };
+        let handle = new_with_explicit_dir(
+            &info,
+            dir.to_path_buf(),
+            acp::ModelId::new("test-model"),
+            sampling_client(),
+            crate::test_support::TEST_MODEL.to_owned(),
+        )
+        .await
+        .unwrap();
+        (info, handle)
+    }
+
+    /// A chat append that never reached the file fails the next barrier once;
+    /// the barrier after that is clean again.
+    #[tokio::test]
+    async fn flush_and_ack_reports_a_lost_chat_append_once() {
+        let dir = tempfile::tempdir().unwrap();
+        let (info, handle) = spawn(dir.path()).await;
+        barrier(&handle).await.unwrap();
+
+        // A directory where the chat log should be makes the append fail.
+        let chat_file = JsonlStorageAdapter::with_explicit_session_dir(dir.path().to_path_buf())
+            .chat_file(&info);
+        let _ = std::fs::remove_file(&chat_file);
+        std::fs::create_dir(&chat_file).unwrap();
+        handle
+            .tx
+            .send(PersistenceMsg::Chat(ConversationItem::user("lost")))
+            .unwrap();
+        // CopyFile is an ordering fence that keeps the write-failure latch.
+        let (tx, rx) = tokio::sync::oneshot::channel();
+        handle
+            .tx
+            .send(PersistenceMsg::CopyFile { one_shot: tx })
+            .unwrap();
+        let _ = rx.await.unwrap();
+        // With the path restored only the lost append can fail the barrier.
+        std::fs::remove_dir(&chat_file).unwrap();
+        let error = barrier(&handle).await.unwrap_err();
+        assert_ne!(error.kind(), io::ErrorKind::NotFound, "got {error}");
+
+        barrier(&handle).await.unwrap();
+    }
+
+    /// A history rewrite that never landed fails the next barrier.
+    #[tokio::test]
+    async fn flush_and_ack_reports_a_failed_history_replacement() {
+        let dir = tempfile::tempdir().unwrap();
+        let (info, handle) = spawn(dir.path()).await;
+        let chat_file = JsonlStorageAdapter::with_explicit_session_dir(dir.path().to_path_buf())
+            .chat_file(&info);
+        // The rewrite stages a sibling temp file; a directory there fails it.
+        std::fs::create_dir(chat_file.with_extension("jsonl.tmp")).unwrap();
+        handle
+            .tx
+            .send(PersistenceMsg::ReplaceChatHistory(vec![
+                ConversationItem::user("rewritten"),
+            ]))
+            .unwrap();
+        let error = barrier(&handle).await.unwrap_err();
+        assert_ne!(error.kind(), io::ErrorKind::NotFound, "got {error}");
+
+        barrier(&handle).await.unwrap();
+    }
+}
@@ -86,7 +86,7 @@ impl JsonlStorageAdapter {
    fn updates_file(&self, info: &Info) -> PathBuf {
        self.session_dir(info).join("updates.jsonl")
    }
-    fn chat_file(&self, info: &Info) -> PathBuf {
+    pub(crate) fn chat_file(&self, info: &Info) -> PathBuf {
        self.session_dir(info).join("chat_history.jsonl")
    }
    fn summary_file(&self, info: &Info) -> PathBuf {
@@ -700,7 +700,7 @@ impl JsonlStorageAdapter {
        options: super::CopySessionOptions,
    ) -> io::Result<super::CopySessionResult> {
        let target_dir = self.session_dir(target_info);
-        std::fs::create_dir_all(&target_dir)?;
+        super::create_dir_all_durable(&target_dir)?;
        let source_summary = self.read_summary_sync(source_info)?;
        let chat_format_version = source_summary.chat_format_version;
        let mut chat_to_copy: Vec<ConversationItem> =
@@ -920,7 +920,7 @@ async fn next_compaction_segment_index(compaction_dir: &std::path::Path) -> u64
 impl StorageAdapter for JsonlStorageAdapter {
    async fn init_session(&self, info: &Info, model_id: acp::ModelId) -> io::Result<Summary> {
        let dir = self.session_dir(info);
-        std::fs::create_dir_all(&dir)?;
+        super::create_dir_all_durable(&dir)?;
        let summary_path = self.summary_file(info);
        if Path::new(&summary_path).exists() {
            tracing::info!("Loading existing session from JSONL");
@@ -1283,29 +1283,10 @@ impl StorageAdapter for JsonlStorageAdapter {
            .await
    }
    async fn sync_session_files(&self, info: &Info) -> io::Result<()> {
-        let info_clone = info.clone();
-        let adapter_clone = self.clone();
-        tokio::task::spawn_blocking(move || -> io::Result<()> {
-            use std::fs::OpenOptions;
-            let adapter = adapter_clone;
-            let files_to_sync = [
-                adapter.updates_file(&info_clone),
-                adapter.chat_file(&info_clone),
-                adapter.summary_file(&info_clone),
-                adapter.plan_file(&info_clone),
-                adapter.rewind_points_file(&info_clone),
-            ];
-            for file_path in &files_to_sync {
-                if file_path.exists()
-                    && let Ok(file) = OpenOptions::new().write(true).open(file_path)
-                {
-                    let _ = file.sync_all();
-                }
-            }
-            Ok(())
-        })
-        .await
-        .map_err(io::Error::other)?
+        let dir = self.session_dir(info);
+        tokio::task::spawn_blocking(move || super::sync_tree_durable(&dir))
+            .await
+            .map_err(io::Error::other)?
    }
    async fn replace_chat_history(
        &self,
@@ -2794,3 +2794,24 @@ async fn graph_mode_state_round_trips_and_tombstones() {
    let after = adapter.load_session_without_updates(&info).await.unwrap();
    assert!(after.graph_mode_state.is_none(), "cleared graph must not resurrect");
 }
+
+/// The barrier covers nested state files, and an unsyncable one fails it.
+#[cfg(unix)]
+#[tokio::test]
+async fn sync_session_files_reports_a_nested_file_that_cannot_be_synced() {
+    use std::os::unix::fs::PermissionsExt;
+    // Root opens anything; the permission probe is meaningless there.
+    if unsafe { libc::geteuid() } == 0 {
+        return;
+    }
+    let dir = TempDir::new().unwrap();
+    let adapter = JsonlStorageAdapter::with_explicit_session_dir(dir.path().to_path_buf());
+    let info = create_test_info();
+    adapter.init_session(&info, default_model_id()).await.unwrap();
+    let nested = adapter.goal_mode_state_file(&info);
+    std::fs::create_dir_all(nested.parent().unwrap()).unwrap();
+    std::fs::write(&nested, b"{}").unwrap();
+    std::fs::set_permissions(&nested, std::fs::Permissions::from_mode(0o000)).unwrap();
+    let error = adapter.sync_session_files(&info).await.unwrap_err();
+    assert_eq!(error.kind(), io::ErrorKind::PermissionDenied, "got {error}");
+}
@@ -1,5 +1,5 @@
 use async_trait::async_trait;
-use std::io::{self, BufRead, BufReader, Seek, SeekFrom};
+use std::io::{self, BufRead, BufReader, Seek, SeekFrom, Write};
 use std::path::Path;

 use crate::extensions::notification::SessionNotification;
@@ -15,6 +15,137 @@ use agent_client_protocol as acp;
 use kigi_sampling_types::ReasoningEffort;
 use kigi_workspace::session::file_state::RewindPoint;

+/// Fsync `file` to stable media; macOS `fsync` may stop at the drive cache.
+#[cfg(target_os = "macos")]
+pub(crate) fn sync_file_durable(file: &std::fs::File) -> io::Result<()> {
+    use std::os::fd::AsRawFd;
+    file.sync_all()?;
+    fullfsync_raw(file.as_raw_fd())
+}
+
+#[cfg(target_os = "macos")]
+fn fullfsync_raw(fd: std::os::fd::RawFd) -> io::Result<()> {
+    // SAFETY: fcntl on a live descriptor with no pointer arguments.
+    if unsafe { libc::fcntl(fd, libc::F_FULLFSYNC) } == -1 {
+        return Err(io::Error::last_os_error());
+    }
+    Ok(())
+}
+
+#[cfg(not(target_os = "macos"))]
+pub(crate) fn sync_file_durable(file: &std::fs::File) -> io::Result<()> {
+    file.sync_all()
+}
+
+/// Fsync `dir` so entries just created or renamed into it survive power loss.
+#[cfg(target_os = "macos")]
+pub(crate) fn sync_dir_durable(dir: &Path) -> io::Result<()> {
+    use std::os::fd::AsRawFd;
+    let file = std::fs::File::open(dir)?;
+    file.sync_all()?;
+    fullfsync_raw(file.as_raw_fd())
+}
+
+#[cfg(all(unix, not(target_os = "macos")))]
+pub(crate) fn sync_dir_durable(dir: &Path) -> io::Result<()> {
+    std::fs::File::open(dir)?.sync_all()
+}
+
+/// Windows has no directory-handle fsync; NTFS journals directory metadata.
+#[cfg(windows)]
+pub(crate) fn sync_dir_durable(_dir: &Path) -> io::Result<()> {
+    Ok(())
+}
+
+/// Write `bytes` to `tmp`, fsync them, rename over `dest`, fsync the parent:
+/// a crash never leaves `dest` torn, and the rename never publishes
+/// unsynced bytes over a synced file.
+pub(crate) fn write_bytes_atomic(tmp: &Path, dest: &Path, bytes: &[u8]) -> io::Result<()> {
+    let written = std::fs::File::create(tmp).and_then(|mut file| {
+        file.write_all(bytes)?;
+        sync_file_durable(&file)
+    });
+    if let Err(error) = written {
+        let _ = std::fs::remove_file(tmp);
+        return Err(error);
+    }
+    crate::util::fs::replace_file(tmp, dest)?;
+    sync_parent_dir_durable(dest)
+}
+
+fn sync_parent_dir_durable(path: &Path) -> io::Result<()> {
+    let parent = path
+        .parent()
+        .ok_or_else(|| io::Error::new(io::ErrorKind::InvalidInput, "path has no parent"))?;
+    sync_dir_durable(parent)
+}
+
+fn open_for_sync(path: &Path) -> io::Result<std::fs::File> {
+    let mut options = std::fs::OpenOptions::new();
+    options.read(true);
+    // FlushFileBuffers needs a writable handle; fsync does not.
+    options.write(cfg!(windows));
+    options.open(path)
+}
+
+/// Fsync every file and directory under `dir`, `dir` included, so all
+/// session state reaches stable media. A path that vanished is not an error.
+pub(crate) fn sync_tree_durable(dir: &Path) -> io::Result<()> {
+    let entries = match std::fs::read_dir(dir) {
+        Ok(entries) => entries,
+        Err(error) if error.kind() == io::ErrorKind::NotFound => return Ok(()),
+        Err(error) => return Err(error),
+    };
+    for entry in entries {
+        let entry = entry?;
+        let file_type = match entry.file_type() {
+            Ok(file_type) => file_type,
+            Err(error) if error.kind() == io::ErrorKind::NotFound => continue,
+            Err(error) => return Err(error),
+        };
+        if file_type.is_dir() {
+            sync_tree_durable(&entry.path())?;
+        } else if file_type.is_file() {
+            match open_for_sync(&entry.path()) {
+                Ok(file) => sync_file_durable(&file)?,
+                Err(error) if error.kind() == io::ErrorKind::NotFound => {}
+                Err(error) => return Err(error),
+            }
+        }
+    }
+    match sync_dir_durable(dir) {
+        Err(error) if error.kind() == io::ErrorKind::NotFound => Ok(()),
+        result => result,
+    }
+}
+
+/// `create_dir_all`, then fsync every directory that gained an entry so the
+/// chain itself survives power loss. A directory that exists but is still
+/// empty re-syncs a few ancestors: an earlier create may have died before
+/// its sync.
+pub(crate) fn create_dir_all_durable(dir: &Path) -> io::Result<()> {
+    let mut gaining_an_entry = Vec::new();
+    let mut cursor = dir;
+    while !cursor.exists() {
+        let Some(parent) = cursor.parent() else { break };
+        gaining_an_entry.push(parent.to_path_buf());
+        cursor = parent;
+    }
+    std::fs::create_dir_all(dir)?;
+    if gaining_an_entry.is_empty() && std::fs::read_dir(dir)?.next().is_none() {
+        gaining_an_entry.extend(dir.ancestors().skip(1).take(4).map(Path::to_path_buf));
+    }
+    for parent in gaining_an_entry.iter().filter(|parent| !is_fs_root(parent)) {
+        sync_dir_durable(parent)?;
+    }
+    Ok(())
+}
+
+fn is_fs_root(path: &Path) -> bool {
+    path.parent()
+        .is_none_or(|parent| parent.as_os_str().is_empty())
+}
+
 pub mod jsonl;
 pub mod search;
 pub mod search_fts;
@@ -228,12 +228,13 @@ fn read_summary(path: &Path) -> io::Result<Summary> {
 fn write_summary_atomic(summary_path: &Path, summary: &Summary) -> io::Result<()> {
    let bytes = serde_json::to_vec_pretty(summary)
        .map_err(|e| io::Error::new(io::ErrorKind::InvalidData, e))?;
-    let tmp = summary_path.with_extension("json.tmp");
-    std::fs::write(&tmp, &bytes)?;
-    // Windows-safe replace: this is the write that persists a session's
-    // CURRENT MODEL — a bare rename made a switched model silently revert
-    // on resume whenever AV/indexer held summary.json open on Windows.
-    crate::util::fs::replace_file(&tmp, summary_path)
+    // Durable and Windows-safe: this write persists a session's CURRENT MODEL,
+    // and other writers (`/rename`) race the persistence barrier on this path.
+    super::write_bytes_atomic(
+        &summary_path.with_extension("json.tmp"),
+        summary_path,
+        &bytes,
+    )
 }

 #[cfg(test)]
@@ -8,13 +8,66 @@ use tokio::time;
 use crate::terminal::runner::{
    AsyncTerminalRunner, TerminalError, TerminalRunRequest, TerminalRunResult,
 };
+use kigi_tty_utils::KILL_REAP_TIMEOUT;

 pub struct LocalTerminalRunner;

-async fn read_stream(mut stream: impl AsyncReadExt + Unpin) -> Vec<u8> {
-    let mut buffer = Vec::new();
-    let _ = stream.read_to_end(&mut buffer).await;
-    buffer
+/// A pipe reader whose buffer is shared with the spawner, so bytes read
+/// before an abort are not lost with the task.
+struct PipeReader {
+    task: tokio::task::JoinHandle<()>,
+    buffer: std::sync::Arc<std::sync::Mutex<Vec<u8>>>,
+}
+
+impl PipeReader {
+    fn spawn(mut stream: impl AsyncReadExt + Unpin + Send + 'static) -> Self {
+        let buffer = std::sync::Arc::new(std::sync::Mutex::new(Vec::new()));
+        let task_buffer = std::sync::Arc::clone(&buffer);
+        let task = tokio::spawn(async move {
+            let mut chunk = [0u8; 8192];
+            loop {
+                match stream.read(&mut chunk).await {
+                    Ok(0) | Err(_) => break,
+                    Ok(n) => Self::lock_buffer(&task_buffer).extend_from_slice(&chunk[..n]),
+                }
+            }
+        });
+        Self { task, buffer }
+    }
+
+    /// Never panics: a poisoned lock just means a holder panicked mid-append —
+    /// the buffer is plain bytes and stays consistent, so recover it.
+    fn lock_buffer(buffer: &std::sync::Mutex<Vec<u8>>) -> std::sync::MutexGuard<'_, Vec<u8>> {
+        buffer
+            .lock()
+            .unwrap_or_else(std::sync::PoisonError::into_inner)
+    }
+
+    fn take_buffer(self) -> Vec<u8> {
+        std::mem::take(&mut Self::lock_buffer(&self.buffer))
+    }
+
+    /// Wait for EOF (pipe closed by all writers) at most [`KILL_REAP_TIMEOUT`],
+    /// then return everything read so far.
+    ///
+    /// Callers only join after the child was reaped or killed, so EOF is
+    /// normally immediate; the bound is a backstop for writers that outlive
+    /// the command and keep the pipe open — a background descendant
+    /// (`printf hi; sleep 30 &`) or a process wedged in uninterruptible
+    /// kernel I/O (D-state), which not even SIGKILL moves.
+    async fn join_bounded(mut self) -> Vec<u8> {
+        if time::timeout(KILL_REAP_TIMEOUT, &mut self.task)
+            .await
+            .is_err()
+        {
+            self.task.abort();
+            // Await the cancellation: abort() is asynchronous, and a reader
+            // past its await point could otherwise append between the take
+            // below and its own completion, losing those bytes.
+            let _ = (&mut self.task).await;
+        }
+        self.take_buffer()
+    }
 }

 /// Truncate buffer to keep only the last `limit` bytes (drops oldest bytes).
@@ -66,12 +119,35 @@ impl AsyncTerminalRunner for LocalTerminalRunner {

        // Detach from the controlling terminal so child processes
        // (e.g. GPG pinentry) cannot open /dev/tty and corrupt the TUI.
+        // (This also makes the child its own session/group leader, which the
+        // ProcessGroup attach below relies on.)
        kigi_tools::util::detach_command(&mut cmd);

+        // process-group-killed on the request timeout; an unreapable (D-state)
+        // child is abandoned to the runtime's orphan reaper
        let mut child = cmd
            .spawn()
            .map_err(|e| TerminalError::Other(format!("Failed to start shell: {e}")))?;

+        // Own the whole tree, not just the direct shell: on timeout the group
+        // is killed so grandchildren can't keep running (and can't hold the
+        // output pipes open past the kill). Group teardown is an enhancement
+        // over the direct-kill floor, so both enrollment failures degrade
+        // softly instead of dropping an already-running shell.
+        let process_group = match kigi_tty_utils::ProcessGroup::new() {
+            Ok(mut group) => {
+                if let Err(e) = group.attach(&child) {
+                    // e.g. the child already exited.
+                    tracing::debug!("ProcessGroup::attach failed: {e}");
+                }
+                Some(group)
+            }
+            Err(e) => {
+                tracing::warn!("Failed to create process group: {e}");
+                None
+            }
+        };
+
        let stdout = child
            .stdout
            .take()
@@ -81,28 +157,46 @@ impl AsyncTerminalRunner for LocalTerminalRunner {
            .take()
            .ok_or_else(|| TerminalError::Other("Failed to capture stderr".into()))?;

-        let stdout_task = tokio::spawn(read_stream(stdout));
-        let stderr_task = tokio::spawn(read_stream(stderr));
+        let stdout_reader = PipeReader::spawn(stdout);
+        let stderr_reader = PipeReader::spawn(stderr);

        let mut timed_out = false;

        let wait_result = time::timeout(request.timeout, child.wait()).await;
-        let exit_status = match wait_result {
+        let exit_code = match wait_result {
            Ok(status_res) => status_res
-                .map_err(|e| TerminalError::Other(format!("Failed to wait for process: {e}")))?,
+                .map_err(|e| TerminalError::Other(format!("Failed to wait for process: {e}")))?
+                .code(),
            Err(_) => {
                timed_out = true;
+                // Kill the direct child AND the whole group (grandchildren),
+                // then return the synthetic timeout result without waiting
+                // for the corpse: `timed_out: true` already tells the caller
+                // everything, and a D-state child would never become reapable
+                // anyway. Both kills are unconditional — `kill()` on a group
+                // whose attach failed is a silent no-op, so the direct
+                // `start_kill` is the guaranteed floor (same as
+                // `util/subprocess.rs`). The kills close the pipes, so the
+                // bounded joins below return immediately in the normal case;
+                // the abandoned child goes to tokio's orphan reaper.
                if let Err(e) = child.start_kill() {
                    tracing::warn!("Failed to kill timed-out process: {e}");
                }
-                child.wait().await.map_err(|e| {
-                    TerminalError::Other(format!("Failed to wait for killed process: {e}"))
-                })?
+                if let Some(group) = &process_group
+                    && let Err(e) = group.kill()
+                {
+                    tracing::warn!("Failed to kill timed-out process group: {e}");
+                }
+                None
            }
        };

-        let stdout_result = stdout_task.await.unwrap_or_else(|_| Vec::new());
-        let stderr_result = stderr_task.await.unwrap_or_else(|_| Vec::new());
+        // Join concurrently (worst case one reap bound, not one per pipe).
+        // Bounded on the natural-exit arm too: a background descendant that
+        // outlives the shell (`printf hi; sleep 30 &`) holds the pipes open
+        // past the child's exit, and an unbounded join stalls on it.
+        let (stdout_result, stderr_result) =
+            tokio::join!(stdout_reader.join_bounded(), stderr_reader.join_bounded());

        let mut combined = stdout_result;
        combined.extend(stderr_result);
@@ -112,7 +206,7 @@ impl AsyncTerminalRunner for LocalTerminalRunner {

        Ok(TerminalRunResult {
            combined_output,
-            exit_code: exit_status.code(),
+            exit_code,
            truncated,
            signal: None,
            timed_out,
@@ -181,4 +275,56 @@ mod tests {
        assert_eq!(result.combined_output.trim(), "hello");
        assert_eq!(result.exit_code, Some(0));
    }
+
+    /// Timing out a command whose grandchild inherited the output pipes must
+    /// (a) return promptly — before the group kill, the joins waited for the
+    /// grandchild's EOF (the full sleep here; forever for a writer wedged in
+    /// uninterruptible kernel I/O, the production incident this guards
+    /// against) — and (b) actually kill the grandchild via the group kill,
+    /// not merely stop waiting for it.
+    #[tokio::test]
+    #[cfg(unix)]
+    async fn test_timeout_kills_grandchildren_and_returns_promptly() {
+        let mut request = make_request("sleep 5 & echo bgpid=$!; sleep 5");
+        request.timeout = std::time::Duration::from_millis(300);
+
+        let started = std::time::Instant::now();
+        let result = LocalTerminalRunner.run(request).await.unwrap();
+
+        assert!(result.timed_out, "run should report the timeout");
+        // Prompt return: request timeout (0.3s) + pipe EOF from the group
+        // kill (normally instant; KILL_REAP_TIMEOUT-bounded worst case) —
+        // anything near the 5s sleeps means we waited for the grandchild.
+        assert!(
+            started.elapsed() < std::time::Duration::from_secs(4),
+            "timeout path must not wait for the killed tree (took {:?})",
+            started.elapsed()
+        );
+
+        // Outcome: the background grandchild must actually be dead, proving
+        // the group kill reached it. Poll briefly — the SIGKILL is delivered
+        // synchronously but init may reap the orphan a beat later.
+        let bg_pid = result
+            .combined_output
+            .lines()
+            .find_map(|l| l.trim().strip_prefix("bgpid="))
+            .and_then(|p| p.trim().parse::<u32>().ok())
+            .expect("background pid should have been echoed before the kill");
+        let deadline = std::time::Instant::now() + std::time::Duration::from_secs(5);
+        loop {
+            let alive = std::process::Command::new("kill")
+                .args(["-0", &bg_pid.to_string()])
+                .status()
+                .map(|s| s.success())
+                .unwrap_or(false);
+            if !alive {
+                break;
+            }
+            assert!(
+                std::time::Instant::now() < deadline,
+                "background grandchild {bg_pid} survived the group kill"
+            );
+            tokio::time::sleep(std::time::Duration::from_millis(50)).await;
+        }
+    }
 }
@@ -912,7 +912,10 @@ fn spawn_with_argv(
            if kigi_sandbox::should_restrict_child_network() {
                // SAFETY: single prctl syscall (async-signal-safe).
                unsafe {
-                    cmd.pre_exec(|| kigi_sandbox::child_net::install_child_network_filter());
+                    let filter = kigi_sandbox::child_net::prebuilt_child_network_filter();
+                    cmd.pre_exec(move || {
+                        kigi_sandbox::child_net::install_child_network_filter(filter)
+                    });
                }
            }
        });
@@ -33,6 +33,11 @@ pub(crate) fn ctx_with_toggle(toggle: HashMap<String, bool>) -> SubagentSpawnCon
    let (tx, _rx) = mpsc::unbounded_channel();
    SubagentSpawnContext {
        lsp: None,
+        spawn_limits: kigi_tools::implementations::kigi::task::admission::SubagentLimits::default(),
+        spawn_permits: std::sync::Arc::new(tokio::sync::Semaphore::new(
+            kigi_tools::implementations::kigi::task::admission::SubagentLimits::default()
+                .effective_max_concurrent(),
+        )),
        parent_max_turns: None,
        gateway: test_gateway(),
        client_hooks: Default::default(),
@@ -1,6 +1,42 @@
 /// Default auto-compact threshold (% of context window) when no source sets it.
 pub const DEFAULT_AUTO_COMPACT_THRESHOLD_PERCENT: u8 = 85;

+/// `tool_choice` sent on compaction summarizer requests when tools are
+/// attached (prefix-cache alignment keeps the tool definitions in the
+/// request). Default `Auto` because some OpenAI-compatible servers reject
+/// `tool_choice: "none"`; the summarization prompt itself forbids tool use.
+#[derive(Clone, Copy, Debug, Default, PartialEq, Eq)]
+pub enum CompactionToolChoice {
+    #[default]
+    Auto,
+    None,
+}
+
+impl std::str::FromStr for CompactionToolChoice {
+    type Err = ();
+
+    fn from_str(s: &str) -> Result<Self, Self::Err> {
+        match s.trim().to_ascii_lowercase().as_str() {
+            "auto" => Ok(Self::Auto),
+            "none" => Ok(Self::None),
+            _ => Err(()),
+        }
+    }
+}
+
+pub(crate) const ENV_COMPACTION_TOOL_CHOICE: &str = "KIGI_COMPACTION_TOOL_CHOICE";
+
+pub(crate) fn resolve_compaction_tool_choice_from(
+    env: Option<&str>,
+    config: Option<&str>,
+    remote: Option<&str>,
+) -> CompactionToolChoice {
+    env.and_then(|s| s.parse().ok())
+        .or_else(|| config.and_then(|s| s.parse().ok()))
+        .or_else(|| remote.and_then(|s| s.parse().ok()))
+        .unwrap_or_default()
+}
+
 /// Env-var override for `auto_compact_threshold_percent`. Parsed as `u8`;
 /// out-of-range or unparseable values are ignored.
 pub(crate) const ENV_AUTO_COMPACT_THRESHOLD_PERCENT: &str = "KIGI_AUTO_COMPACT_THRESHOLD_PERCENT";
@@ -127,6 +163,37 @@ pub fn resolve_compaction_wall_clock_budget_secs(gb_global: Option<u64>) -> u64
    resolved
 }

+#[cfg(test)]
+mod compaction_tool_choice_tests {
+    use super::{CompactionToolChoice, resolve_compaction_tool_choice_from as resolve};
+
+    #[test]
+    fn precedence_env_config_remote_default() {
+        assert_eq!(resolve(None, None, None), CompactionToolChoice::Auto);
+        assert_eq!(
+            resolve(None, None, Some("none")),
+            CompactionToolChoice::None
+        );
+        assert_eq!(
+            resolve(None, Some("auto"), Some("none")),
+            CompactionToolChoice::Auto
+        );
+        assert_eq!(
+            resolve(Some("none"), Some("auto"), Some("auto")),
+            CompactionToolChoice::None
+        );
+        // Unparseable tiers fall through instead of masking lower tiers.
+        assert_eq!(
+            resolve(Some("bogus"), None, Some("none")),
+            CompactionToolChoice::None
+        );
+        assert_eq!(
+            resolve(Some("NONE "), None, None),
+            CompactionToolChoice::None
+        );
+    }
+}
+
 #[cfg(test)]
 mod compaction_wall_clock_budget_tests {
    use super::resolve_compaction_wall_clock_budget_secs as resolve;
@@ -161,6 +161,9 @@ async fn request(handle: &PermissionHandle, access: AccessKind, id: &str) -> Dec
        session_id: None,
        subagent_type: None,
        subagent_description: None,
+        // The accesses here carry no file path; the process cwd is a real cwd.
+        tool_cwd: AbsPathBuf::new(std::env::current_dir().unwrap()).unwrap(),
+        display_cwd: None,
    };
    let PermissionHandle::Actor { cmd_tx, .. } = handle else {
        panic!("expected actor handle");