The inbox projection used to call `markAgentInboxDelivered` after
appending each message to the agent's session shadow. That helper
sets `agent_messages.read_at = now()` for direct messages and
inserts into `agent_message_reads` for channel messages — the SAME
flags `listAgentInboxUnread` and the public `?unreadOnly=true` API
use to define "unread". So the moment we put a message into the
recipient's transcript, the recipient's own API call returned
nothing pending. The IC prompt told the agent to "respond to ALL
unread messages"; the API said zero unread; the agent posted a
status comment instead of replying. Every silent-employee report
traces back here.
Replay safety lives entirely on `agent_message_projections`
(installed in v46) — that ledger is keyed by `(message_id,
session_id)` and `claimProjection` already short-circuits a re-run
without touching `read_at` at all. The two concepts are now
properly separated:
- "delivered" (we put it in front of the agent) — tracked by
the projection ledger, owned by the runtime bridge;
- "read" (the agent acknowledged it) — tracked by `read_at` and
`agent_message_reads`, only flipped by the explicit POST
/messages/{id}/read endpoint the agent calls itself.
`markAgentInboxDelivered` is removed because there is no remaining
caller and its semantics were the bug — keeping it around would
just invite the same mistake again.
`retireDeletedArtifacts` used to iterate `deleted.slice(0,
MAX_BASH_PROJECTED_FILES)` and emit no signal for paths past the cap
— a large `rm -r` would leave the 51st row onward visible in the
DocumentCenter with no indication anything was missed. The cap was
inherited from the registration path, where it bounds file-stat
round-trips; retirement is a pure metadata update with no filesystem
cost, so the cap was value-free safety theatre.
New DB surface `archiveWorkProductsByExternalPaths` runs a single
`UPDATE … SET status = 'archived' … WHERE external_id IN (…)
AND status <> 'archived' RETURNING id, external_id` per 500-path
chunk — one round-trip per chunk instead of one per path. The
projector feeds the full `deleted` list in, maps the returned ids
back to `retired` outcomes, and no deletion is left behind regardless
of the run's cleanup fan-out size.
Fix#16 kept deletions from registering, but rows created by earlier
runs stayed `status = "active"` forever even after the agent `rm`'d
the file. The DocumentCenter kept showing zombies and
`openWorkProduct` resolved to an empty path.
`ChangeSet` now carries both `changed` and `deleted` paths, populated
from the two detectors:
- git porcelain: the `D` status entries that were previously
dropped are now captured into `deleted` instead;
- baseline diff: paths present in the baseline but no longer on
disk at diff time are `deleted`.
A new `retireDeletedArtifacts` pass runs alongside
`projectChangedPaths` and flips the matching
`work_products.status` to `archived`, bumping `last_touched_by_run_id`
to the retiring run so the provenance timeline records the deletion.
`upsertFilesystemArtifact` sets `status = "active"` on update so a
subsequent run that recreates the path brings the row back into the
visible set. A new `ProjectionOutcome` variant — `retired` — lets the
hook and reconciliation logs report both sides of the fan-out.
`captureBaselineIfRequired` stops walking at `MAX_BASELINE_ENTRIES`
and flags `truncated = true`, but nothing downstream honoured the
flag. `diffAgainstBaseline` then treated every file past the sample
cap as "new" because `baseline.mtimeByRelPath.get(rel)` was
`undefined`, so a non-git workspace with more than 20k files could
report thousands of untouched files as deliverables on every bash
call.
When the baseline is truncated, the diff now reports only the
sampled paths whose mtime advanced. Unsampled paths stay outside the
diff — we cannot distinguish "always there" from "just written"
without the full snapshot, and silently mis-classifying is worse
than missing the tail. `collectChangedPaths` propagates the flag as
`coverageTruncated` and both the bash and the reconciliation entry
points append a `baseline_truncated` skip outcome so operators see
the degraded signal (typical remediation: `git init` the workspace
so `git status` becomes the authoritative diff source).
`isGitRepo` only accepted a `.git` directory, so linked worktrees and
submodules — both of which store `.git` as a file containing
`gitdir: …` — silently fell into the non-git branch. The baseline
diff then replaced `git status` as the source of truth for workspaces
that git itself recognises, and edge cases where the baseline is
missing (a freshly-created worktree mid-run, a submodule without a
captured snapshot) bled into `no_baseline` skips instead of running
the authoritative diff.
Call `git rev-parse --is-inside-work-tree` and take git's own yes/no
answer. Exit 0 means git would happily run `status` there; any other
outcome (no git binary, corrupted ref, genuinely non-git) lands us on
the baseline path exactly as before.
`prepareRun` used to call `captureWorkspaceBaseline` on every
workspace with a path, even when `.git` was present. The baseline was
silently discarded by `collectChangedPaths` (the git branch wins) but
the walk still paid for itself — an O(tree) scan on every run start,
capped at 20k files, with the only effect being that the first 20k
entries got captured before the capper truncated the tail.
Collapse the decision into `captureBaselineIfRequired`: the function
short-circuits to null for git workspaces so `git status` stays the
sole source of truth, and only scans the tree when the baseline is
the only source available. The call site in `runtime-bridge` now
passes through without the per-call conditional.
Two bleed-throughs let a removed file land in `work_products` and
plant a ghost row:
- `listChangedRelPaths` pushed every non-`!!` git porcelain entry,
including worktree-deleted (`Y = D`) and staged-deleted-with-
synced-worktree (`XD` + ` `) rows, so the projector tried to
upsert a path that was no longer on disk.
- `upsertFilesystemArtifact` never stat'd the absolute path; a
deleted file (from either source detector) still reached
`createWorkProduct` and left the DocumentCenter "open" link
pointing to nothing.
Filter the two deletion forms at the git list step and add an
`existsSync + isFile` gate in `upsertFilesystemArtifact`. The check
also catches edge cases the baseline diff surfaces — a file touched
and then `rm`'d, a regular file replaced by a symlink, a path that
raced with a `cleanup` shell script — so neither change detector can
register a row that would not open.
`projectBashArtifacts` and `reconcileWorkspaceArtifacts` used to bail
with `not_a_git_repo` whenever the workspace had no `.git`, so bash
output in such workspaces never reached `work_products` and the
DocumentCenter looked empty while files piled up on disk.
`captureWorkspaceBaseline` now walks the workspace at `prepareRun`
time, skipping the same categories the registrability filter rejects
(`.git`, `node_modules`, `.next`, caches) and capping at
`MAX_BASELINE_ENTRIES` so large trees stay bounded. The baseline rides
on `ActiveRunContext.workspaceBaseline` and is passed into every
artifact projection.
`collectChangedPaths` is the single change-detector for both the
per-bash projection and the end-of-run reconciliation. It picks `git
status` when a `.git` is present and falls back to a
`diffAgainstBaseline` walk otherwise. Paths already on the ignore list
still never land in `work_products`; the identity stays
`(company, workspace, provider, externalId)`; the upsert still refreshes
`last_touched_by_run_id` so the deliverable signal follows the latest
touching run.
`registerInboundInjection` used to index correlations by
`Date.parse(record.createdAt)` — a millisecond-precision wall-clock
value. Two inbox rows that landed in the same millisecond collapsed
onto one Map slot, and the thread-event writer lost one peer's
identity on `message_end`.
Identity now lives on `record.id`. `projectInboundMessage` allocates a
monotonic `piTimestamp` per run (base = `Date.parse(createdAt)`,
bumped to `max(base, lastAllocated + 1)` so it is always unique
within the run) and stamps that value onto the projected pi message.
The correlation registry keeps `byMessageId` as the primary index and
`byPiTimestamp` as a reverse lookup so the `message_end` callback —
which only surfaces the pi-message shape — still resolves back to the
author without depending on wall-clock precision.
The runtime message injector now defers projection to
`RuntimeBridge.projectInboundMessage` so the steered message, the
correlation entry, and the allocated timestamp stay in lockstep.
The old `prepareRun` projection drifted: append-to-shadow, thread-event,
and `markAgentInboxDelivered` ran as three independent writes, so a
crash between the append and the delivered mark replayed the same
inbox row on the next run and double-populated the transcript.
v46 installs `agent_message_projections` keyed by (message_id,
session_id) as the idempotent gate. The projection closure now wraps
the ledger claim, the shadow append, the thread-event write, and the
delivered-mark inside one `BEGIN IMMEDIATE` transaction (exposed via
`AppDatabase.withTransaction`). `claimProjection` returns false on a
replay, so the whole closure short-circuits — no double transcript,
no double thread event.
The natural key used to be (companyId, provider, externalId), so two
workspaces with the same workspace-relative path collided onto one
row. Provenance was also frozen at the first write: `created_by_run_id`
never refreshed, so a run that edited a pre-existing file never
qualified as a deliverable and the DocumentCenter "open" link could
still jump into the first workspace after a repo move.
Migration v45 adds `workspace_id` + `last_touched_by_run_id`, backfills
`workspace_id` via the task/run join, seeds `last_touched_by_run_id =
created_by_run_id`, and installs a partial unique index on
`(company_id, workspace_id, provider, external_id)` scoped to rows
that have a workspace. The artifact projector now records
`workspace_id` on insert and refreshes `last_touched_by_run_id` on
every upsert. The deliverable signal query reads either provenance
column so a run that only edited a pre-existing file still counts.
`resolveWorkProductTarget` prefers the stored `workspace_id` and falls
back to the task/run join for legacy NULL rows.
Per-tool `afterToolCall` projection stays the primary write path but is
now backed by an end-of-run reconciliation sweep on `RuntimeBridge`.
`reconcileWorkspaceArtifacts` re-scans the workspace via `git status`
before the run's resources are released, idempotently upserting any
file the per-tool hook skipped — the exact gap that was leaving the
DocumentCenter empty while the disk filled up.
Projection failures stop being swallowed silently: the hook now logs
a tool-level summary at INFO and throws are routed to ERROR so they are
visible in the desktop log tail. The reconciliation sweep surfaces its
own failures the same way, so an operator can tell the difference
between "no files produced" and "files produced but registration
broken".
recordThreadEvent used to drop the event whenever `envelope.taskId` was
null, so communication wakes produced nothing in the thread view even
while the agent actively replied to peers. Taskless runs now write into
a per-agent general thread (`ensureAgentCommunicationThread`), so the UI
surfaces peer conversations alongside task activity.
Peer-authored user turns inside the transcript used to be labelled with
the recipient's id — pi erases the sender once the message is projected.
A new `inboundCorrelations` map on `RuntimeBridge` pins each pi-message
timestamp to its source `agent_messages` row at injection time, so the
thread-event writer can recover the real author id and `senderKind` on
`message_end`. prepareRun projections (no `message_end` fires for
priorMessages) write their thread event directly as the projection lands
so they are not silently lost.
Replace the fire-and-forget steer with a correlated ACK round-trip so main
only marks an inbox row delivered once the worker confirms the message
landed on the live Agent. Steer against a run that is still in the
`preparing` window or has already started `closing` now falls back to the
wake + inbox projection path instead of being silently dropped.
Worker publishes `run-agent-state` transitions at Agent construction and
teardown; main tracks these on `ActiveRunContext.state` and refuses to
steer runs that are not `agent_active`. A rejected or timed-out ACK
triggers `wakeAgentIfPossible` so the next `prepareRun` replays the row.
Previous Task 2 pass gated every destructive tool on any blocking
approval or any open wait-state for the agent. That was too coarse:
a pending deliverable_review or approve_ceo_strategy (both decision-
type approvals, waiting for a human to respond to completed or
proposed work) froze the agent's entire toolchain — including the
`bash` call the agent uses to POST /api/messages. Result: CEO
talked, employees woke on heartbeat, tried to reply, the message
POST was blocked as a "destructive tool", the run ended silent.
Every reply collapsed to an empty heartbeat.
Two corrections land together:
1. Approval gate narrows to the five approval types that actually
gate EXECUTION — `dangerous_command`, `secret_access`,
`production_deploy`, `social_post`, `budget_override`. Review /
decision types are ignored at the tool layer; they remain visible
in the board UI and still block via the heartbeat scheduler when
the execution mode demands it.
2. Wait-state check leaves the hook entirely. The heartbeat scheduler
(`checkpoint-engine.shouldPauseHeartbeat`) is the single owner of
"agent should pause because a thread is waiting" — mirroring the
rule at tool level produced the same company-wide freeze for
any reply on any thread. Heartbeat is the right scope; tools
are not.
The destructive-tool set and the operational-approval set are both
named, readonly, and centralised so the next addition reads like a
whitelist instead of a patch.
The session_entries supersede pipeline was a design that never
shipped: `SqliteSessionStore.replaceMessages` / `appendCompaction`,
`sessionEntries.replace` / `listAll` / `deleteForSession`, and the
`"compaction" | "summary"` kind variants had zero live callers.
Compaction wires through pi's `transformContext` and leaves the
persisted transcript alone, so the shadow-row machinery was pure
cruft with two tests propping it up.
Rip it out: the store exposes only `loadMessages` and
`appendMessage`, `sessionEntries` exposes only `append` and
`listActive`, and `SessionEntryKind` narrows to `"message"` so new
writers can't re-introduce a dead kind by accident.
`is_superseded` stays on the table — the DB column is harmless
with no writer, and dropping it is a separate schema migration
concern.
`listTopLevelAgentIds` loses its only caller in the previous
commit (company channel now fans out through
`listCompanyBroadcastAgentIds`), so it goes too. Three tests that
validated the removed behaviour are deleted; the one remaining
session-store test still covers the live `append + load` path.
pi's `transformContext` hook fired on every LLM call but returned
the transcript unchanged, so a long-running agent paid per-token
for its entire history on every turn. The README promised "sessions
persist + auto-compact" but the main loop had no compaction wired
in at all — only rotation at the long-cycle boundary.
`context-compactor` closes that gap. A pure, provider-free function
estimates tokens (~4 chars/token), short-circuits below a 120K
threshold, and on overflow keeps the last 20 messages verbatim
while collapsing older ones into a single "context summary" user
turn. The summary preserves the original task and the first twelve
assistant highlights so the agent has continuity without forwarding
every historical turn.
Compaction is non-destructive: the Agent's in-memory transcript
and the SQLite session store still hold the full history, feeding
UI timelines and audit. Only the LLM payload shrinks. Rotation
remains the long-cycle boundary (separate commit) — compaction is
what keeps each individual LLM call within budget.
The artifact projector only registered write/edit. Everything bash
produced — build artifacts, files written via shell redirection,
git commits, generated code — stayed invisible to the desktop UI
and the audit layer. Filesystem truth and UI truth were
systematically out of sync.
This lifts bash onto the same filesystem artifact protocol as
write/edit. When a bash tool call succeeds in a git workspace, the
projector runs `git status --porcelain=v1 -z --untracked-files=all`
and upserts a work_products row for every modified/untracked entry,
reusing the existing `(companyId, provider='filesystem', externalId)`
identity so repeated edits never duplicate. Non-git workspaces fall
through with an explicit "not_a_git_repo" skip reason — there is no
authoritative baseline there to diff against, and silently
half-registering would be worse than skipping.
A 50-file cap per invocation bounds the scan so a runaway rebuild
cannot flood the table; anything past the cap is logged as a skip.
Rename entries (emitted as `R new\0old`) advance the iterator past
the old-path entry so deletions of the rename source are not
double-registered.
The projector now always returns an array so `write`/`edit` (one
path) and `bash` (many paths) share one shape, and the hook handler
can log uniform outcomes.
.gitignore gains `AgentCompany-Rust/` — the separate Rust rewrite
tree lives in the same repo root but is not part of the desktop app
build and should not drift into feature commits.
Incoming messages (human → agent, agent → agent, channel broadcast)
persisted to the inbox and were only projected into the recipient's
transcript on the next `prepareRun`. When the recipient was already
running, every message waited for the next wake — a Slack-like
system operating at heartbeat cadence, not in real time.
`RuntimeMessageInjector` closes the gap: it looks up the recipient's
active run through `RuntimeBridge.findActiveRunForAgent`, projects
the persisted `AgentMessageRecord` via the existing inbox-projection
helper, and posts a `steer-run` ParentMessage into the worker. The
worker's runner enqueues the projection on the Agent's steering
queue; pi drains it at the next turn boundary and emits the usual
`message_end` → `persistMessage` sequence, so the session store
shadow is preserved without double-writing.
`dispatchAgentMessage` gains an optional `injectRuntimeMessage` hook
used by every API-side dispatch. Injection succeeds → we mark the
inbox row delivered and skip `wakeAgentIfPossible`. Injection fails
(no active run) → the legacy inbox + wake path still fires, so
sleeping agents continue to receive their messages on the next turn.
The injector is owned by the service graph alongside `workerService`
and `piRuntimeRef`, which gives it both the worker postMessage
channel and the RuntimeBridge active-run map without additional
coupling. Internal dispatch call sites (sweepStaleReviews,
automation actions) leave the hook unset; those system messages
target idle agents by design and do not need the runtime path.
Approval creation already flowed through the API server, but pi's
`bash` / `write` / `edit` tools never traversed that layer, so the
gate was purely advisory: the agent kept mutating the workspace while
the board tried to hold it. The fix: policy runs at the one place
every destructive tool invocation must pass, pi's `beforeToolCall`.
`checkApprovalBlock` fails closed on lookup errors — an unreadable
approvals table is a safety incident, not a licence to execute. It
re-reads `getBlockingApproval` on every call so an approval created
mid-run takes effect on the next tool without any cache.
`checkWaitStateBlock` rides the unified
`hasUnresolvedWaitStateForAgent` query so task, approval, and
freeform thread wait states all gate the same tools. Read-only tools
are untouched; only bash/write/edit are gated because gating
observation would slow investigation without adding safety.
The destructive-tool set is promoted to a named predicate so the
Task 5 artifact projector can share the same definition when it
extends to bash outputs.
The worker previously constructed `new Agent({systemPrompt, model, tools,
messages, streamFn})` and stopped there — transformContext,
beforeToolCall, afterToolCall, sessionId, thinkingLevel, and steer/
followUp were all dark. Approval enforcement, artifact projection,
auto-compaction, and mid-run steering all had to reach the agent
through other, non-canonical channels (event observers, prompt-block
injection at prepareRun, API-layer gates). That left the whole pi
contract at one-tenth utilisation and forced business logic into the
wrong seams.
This change rebuilds the contract assembly layer:
- `pi-agent-factory` builds the Agent with the full option surface:
sessionId (taskSessionId), thinkingLevel (mapped from
agent.metadataJson), and three RPC-proxied hooks.
- Each hook shuttles a thin payload over the existing worker RPC
channel into `RuntimeBridge`, which routes to `hook-handlers`:
* beforeToolCall → approval-gate policy (blocks destructive
tools while a pending approval is outstanding).
* afterToolCall → filesystem artifact projection, previously
nested in the tool_execution_end event observer.
* transformContext → pass-through today; compaction wires in
next commit.
- ParentMessage gains `steer-run` / `followup-run`; `WorkerRunner`
keeps a runId→Agent registry so main can inject mid-run messages
into the active pi loop.
`worker-types` extends the rpc-request method enum; `runtime-bridge`
loses the per-event artifact projector and gains a `findActiveRunForAgent`
lookup used by the upcoming inbox → steer path. No public main API
changes yet — the wiring is the foundation for the next five commits.
Sessions are already agent-scoped post-v34, but the code kept routing
rotation / accumulation / handoff through the task indirection. Heartbeat
runs (no task) silently skipped all three, letting taskless transcripts
grow past the model's context window — the 7M-token explosion.
- DB adds getSessionById(sessionId) — direct lookup bypassing
tasks.active_session_id
- session-management-service: generateHandoff(session) is now agent-scoped,
walking recent runs across tasks + heartbeats; getHandoffForAgent(agentId)
is the canonical lookup, getHandoffContext(taskId) is a thin shim for
callers that still hold a task id
- turn-context.ts::resolveSession drops the `if (!task)` early return:
rotation evaluation runs on the agent session unconditionally
- run-finished-helpers.ts::accumulateAndRotateSession drops the `!task`
gate: tokens accumulate and rotation fires for every run, keyed on the
agent session, not the task
- Rotation evaluator contract: evaluatePreRunSessionRotation(agentId,
connectorId) → { rotated, handoff, sessionId }; old (taskId, connectorId)
signature removed
IPC contract tightening: ipc-support::emitEvent becomes a best-effort
sink. "Render frame was disposed" during renderer reload is a lifecycle
signal, not a logic error; silently drop it instead of letting it abort
downstream event emissions.
Artifact projection: pi's write/edit tool results now upsert a
`work_products` row via a new agent-runtime/artifact-projector.ts. No
more agents writing 54 files into the workspace that the UI never sees.
Migration v40 relaxes work_products.task_id to nullable so heartbeat
artifacts can register too.
File size: split agent-event-projection.ts out of runtime-bridge.ts to
keep the bridge under the 400-line cap after adding the projector hook.
Tests: 604 unit + integration green; generateHandoff callsites updated
for the single-argument signature; artifact-projector has pure-function
coverage via the projector module's own shape.
The Co-Founder beforeToolCall hook (introduced in 77f81a7) gated every
non-readonly pi tool call as a `dangerous_command` approval. Because
agents call `POST /api/companies/:id/messages`, `/agents` (hire), and
every other company mutation via bash→node/curl, the first bash call in
every Co-Founder run got parked on a 30-minute approval timeout. Users
observed it as "agents never talk to each other" and "every bash command
waits for approval"; both symptoms are the same bug.
The business-level approval flow (`hire_agent`, `document_review`,
`plan_review`, etc. in execution-policy.ts) already gates the events
users actually care about, at the HTTP route layer where the event
originates (routes-agents.ts:368 for hiring, document routes for
deliverables). Stacking a pi-level hook on top was duplicate gating with
the wrong taxonomy.
Removed:
- src/main/agent-runtime/approval-hook.ts (entire file, 137 LOC)
- `beforeToolCall` conditional in worker-runner.ts and the
corresponding RPC method in worker-types.ts and runtime-bridge.ts
- `approvalHook` field + `resolveApproval` callback + shutdown
`cancelAll` in runtime-bridge.ts / service-graph-pi-runtime.ts /
wire-api-server.ts
Unchanged:
- `dangerous_command` approval type itself (still used by automation
workflows and routes-misc.ts)
- CEO mode (already never attached the hook)
- All business-level approval types and their wiring
The earlier sweep cleaned the hot path (worker/runtime/automation/api-
server/ipc core — 59 sites). This pass finishes the deferred long tail:
main/* services, database/* CRUD modules, api-server/* routes, ipc-*
handlers, thread-integration, automation helpers, browser, skills.
Default strategy: catch (error) { logger.warn("[module] context:", ...) }
— preserve current behavior, make the failure visible instead of
silently hidden. Widened 3 logger interface signatures to include warn
where callers only passed info/error previously (thread-helpers,
host-environment-service, ipc-domain-common).
Legitimate boundaries left intact: JSON.parse on untrusted input, DNS
resolver no-records paths, explicit safe* helpers (safeJson, parseJson,
parseTriggerDetail), URL-parse predicates, and `<unserialisable>`
fallbacks inside the prompt-builder summarisers. Each kept catch now
either has a reason in context or is a documented safe* helper.
Tests adjusted only where test-constructed no-op loggers needed a warn
stub (message-dispatch, company-mode-switch, ipc-agent-graph-handlers).
With the earlier pass, AgentCompany no longer has silently-swallowing
catch blocks in main-process or renderer business code.
Cost was written through three independent paths: worker-runner shipped
`finalUsage.costUsd` in run-finished -> finishRun wrote `runs.cost_usd`,
`tasks.cost_usd += costUsd`, `agents.spent_monthly_usd += costUsd`, and
inserted a per-run aggregate row in cost_entries. CostWriter
additionally inserted per-message rows in cost_entries from main. Any
CostWriter failure created silent drift between `SUM(cost_entries)`
(read by budget enforcement) and `runs.cost_usd` (read by the UI).
Make CostWriter the single writer, finishRun the single reader.
- Add `db.recordCostEntry(...)` as the first-class insert path for
per-message cost rows. CostWriter calls it; tests call it directly.
- Remove the aggregate INSERT from finishRun. finishRun now reads
`SELECT COALESCE(SUM(amount_usd), 0) FROM cost_entries WHERE run_id=?
AND unattributed=0` and writes that as the authoritative run total to
runs.cost_usd, tasks.cost_usd, and agents.spent_monthly_usd.
- Drop `costUsd` from `finishRun` input, the worker `run-finished`
ChildMessage shape, the automation `RunFinishedEvent` type, and every
caller + test fixture. Cost is owned by the per-message path.
- worker-runner stops extracting `usage.cost.total` — token counts stay
for session accumulation, cost comes from pi-ai via CostWriter.
- run-finished-helpers now gates budget incident checks on
`event.status === "succeeded"` instead of a cost field that no longer
exists on the event.
- automation-run-completion fires `budget_threshold` on every succeeded
run (post-run spent_monthly_usd already reflects the SUM path).
- Update budget tests to write a `recordCostEntry` before `finishRun`
to simulate the per-message cost path; drop `costUsd` from cross-
subsystem, edge-case-budget, database, and related fixtures.
One writer, one reader, one source of truth. No drift.
An earlier cleanup deleted HumanApprovalHook wholesale on the rationale
that the body always returned {}. That was the stub implementation, not
the feature: 版本更新.md specifies AI Co-Founder mode as "human-reviewed
execution that pauses until human review arrives". With no beforeToolCall
attached, Co-Founder and CEO modes behaved identically — a silent
regression of the product's core differentiator.
Bring the hook back with real gating logic and split the Agent
construction path by execution mode so CEO runs pay zero IPC overhead.
- HumanApprovalHook.before(): create a `dangerous_command` approval row
for non-whitelisted tool calls, emit approval-updated, return a promise
that resolves when a human decides.
- AUTO_APPROVED_TOOLS whitelist: read/ls/glob/grep/find/tree. Anything
else (bash, write, edit, …) waits for human approval.
- approvalHook.resolveApproval(): invoked from wire-api-server's
onApprovalResolved *before* automation and thread follow-ups, so the
pi Agent unblocks the instant the human decides.
- approvalHook.cancelAll(): release every pending wait on shutdown so the
worker can drain cleanly; wired into service-graph-pi-runtime.release.
- worker-runner: attach beforeToolCall only when job.executionMode ===
"ai_cofounder". CEO mode constructs Agent with no hook, no RPC hops.
- WorkerRpcClient.invoke: accept per-call timeoutMs override so approval
RPCs can wait 30min instead of the 30s default.
- worker-types schema: restore beforeToolCall in the method enum.
- RuntimeBridge.handleRpc: restore beforeToolCall branch.
No changes to approval DB schema, approval lifecycle, or existing
automation/thread handling. CEO-mode runs retain zero-IPC-per-tool
behavior from the prior cleanup.
Inter-agent messages previously entered the recipient's context as a
separate prompt block (recentMessages), rebuilt every turn. This double-
charged tokens and kept messages outside pi's native AgentMessage stream.
Now at prepareRun: pull the recipient's unread inbox, project each
AgentMessageRecord into a pi user-role AgentMessage, persist to
SqliteSessionStore, and mark delivered. The recipient reads messages
as natural conversation turns in the transcript.
- db: listAgentInboxUnread + markAgentInboxDelivered (transactional)
- agent-runtime/inbox-projection.ts: AgentMessageRecord -> pi AgentMessage
- runtime-bridge.prepareRun: inject + deliver before loadMessages
- prompt-builder: delete recentMessages context block + buildMessagesSection
- message-normalize: drop unused normalizeMessagesForPrompt
- delete prompt-builder/attachment-resolver.ts (orphan after section removal)
Send path (api-server/routes-messages, ipc, UI) unchanged — inbox table
remains the UI source. Transactional delivery marking prevents duplicate
injection across runs. Failed append -> no mark -> retry on next run.
karpathy-guidelines pass — never guess at types, read the source.
Root cause of the silent "等待日志输出" Console:
1. runtime-bridge's projectAgentEventToLogChunk switched on event types
that never exist in pi. Real types per pi-mono-reference/packages/
agent/src/types.ts `AgentEvent`:
agent_start / agent_end / turn_start / turn_end
message_start / message_update / message_end
tool_execution_start / tool_execution_update / tool_execution_end
(NOT the `assistant_message_delta` / `tool_call_start` / `error` I
invented.) Every case therefore fell through to `default → null`,
so zero log chunks were emitted even when the agent was streaming
tokens.
Rewrote the switch to match the real union. Streaming assistant
text and thinking now route through `message_update` ->
`assistantMessageEvent.text_delta` / `thinking_delta` (see
pi-mono-reference/packages/ai/src/types.ts `AssistantMessageEvent`
for the inner protocol). Tool activity routes through
`tool_execution_start` / `tool_execution_end`.
2. worker-runner had no Console-visible breadcrumbs between
`run-status: running` and eventual turn output, so when the agent
stalled (credential fetch, model resolution, first API call) the
user saw a bare "RUNNING" with no clue where time was being spent.
Added 5 targeted `run-log` probes:
[runner] starting run ...
[runner] prepareRun ok, resolving model
[runner] model=... provider=... — constructing Agent
[pi-event] <type> (inside agent.subscribe, one per event)
[runner] calling agent.prompt (N chars)
[runner] agent.prompt resolved
These are structured runtime signals, not test/mockup data. They
surface the exact phase a stuck run is in.
typecheck 0 / lint 0-0 / test:unit 508/508 / build:verify 0-0.
Two coupled user-visible defects were fixed.
1. Every tool call (shell / edit / write / read / ...) triggered a
synchronous human-approval prompt under BOTH execution modes. The
worker blocked inside pi's beforeToolCall waiting on a DB approval
row, so runs never progressed — "等待日志输出" forever.
The execution-mode contract is now enforced:
- AI CEO → fully autonomous, no framework gate.
- AI Co-Founder → still fully autonomous at the pi layer. Human
review is a model-driven decision (LLM calls an explicit
`request_approval` tool when it judges review necessary), not
a framework-level interrupt.
HumanApprovalHook collapses to a no-op `before` / `after` /
`cancelAll`. All machinery that queued `dangerous_command`
approvals from the worker-side hook is removed. No more starving
the agent on `node -e` self-call checks.
2. The Console panel only subscribed to `run-log` stdout/stderr
chunks, so pi's agent-event stream (assistant_message_delta,
tool_call_start/end, errors, turn_end) produced zero visible
output. With bug (1) fixed, runs now execute — but the user
would still see a blank Console.
runtime-bridge.handleAgentEvent now projects each meaningful
agent-event into a run-log chunk (assistant_message_delta → raw
delta, tool_call_start → `[tool:bash] {args}`, tool_call_end →
`[tool:done|error:name] {result}`, error → stderr). ConsoleView
picks these up via the existing run-log DesktopEvent pipe with
no renderer change.
tests/unit/approval-hook.test.ts rewritten to assert the new
autonomous contract (before always returns {}, no approval rows
written, cancelAll is a no-op). Tests 508/508 pass.
typecheck 0 / lint 0-0 / build:verify 0-0.
pi-ai's getModel/getModels does not know about the Fireworks providers
(fireworks, fireworks-firepass); both live under our local model-resolver
because Fireworks speaks the Anthropic-compat endpoint with Bearer auth
at https://api.fireworks.ai/inference, not pi-ai's built-in catalog.
Three call sites were still hitting pi-ai directly and failing silently
for Fireworks:
1. connector-model-policy.ts — listProviderModels() caught the pi-ai
throw and returned []. The "Fire Pass (CEO runtime)" card in the
Team Models onboarding step therefore showed "暂无模型 / 已启用 0/0",
blocking users on Fire Pass from continuing onboarding.
2. worker/worker-runner.ts — getModel() threw mid-run when the selected
connector was Fireworks, so the worker could never actually execute
an agent turn against Fire Pass or the curated Fireworks model set.
Now uses resolveModel() + the Fireworks-specific Bearer header from
resolveExtraStreamHeaders() so streamSimple hits the correct endpoint
with the correct auth.
3. agent-runtime/cost-writer.ts — cost observation silently wrote
amount_usd=null for Fireworks runs because getModel() threw inside
the try/catch. Now resolves via resolveModel() so the Fireworks
curated contextWindow/maxTokens are visible to calculateCost().
Verification: pnpm typecheck 0 / pnpm lint 0-0 / pnpm build:verify 0-0
/ pnpm test:unit 508/508.
Bootstraps CredentialStore/Resolver, OAuthLoginService, and RuntimeBridge
inside service-graph via the existing pi-runtime factory, and exposes the
bridge to worker-service for agent-event + rpc-request routing. Adds a
`prepareRun` RPC that resolves credentials, loads prior messages, and
registers the active run context in a single call. Session rotation drops
the adapter nativeContextManagement branch; run payloads now carry the
taskSessionId and executionMode the worker needs to drive pi-agent-core.
- credential-resolver.resolve: throw explicit CREDENTIAL_INVALID when
the connector has no auth schema, and add a default branch so the
switch has a definitive return path for every auth kind
- credential-resolver api_key branch: guard definition.provider being
unset before calling pi-ai's getEnvApiKey
- cost-writer: only attempt cost lookup when definition.provider is
defined; adapter-only connectors still write an unattributed row
P3-56: three call sites repeated the same 'lookup → fall back to
new → set if missing' dance for session stores. The duplication hid the
fact that the fallback allocation in the second and third sites could
be wasted when the entry already existed. Extract a single
getOrCreateStore helper that does one map lookup, one conditional
construction, and always returns the cached instance.
ZacharyZhang-NY