Runs record a context manifest once a model turn succeeds; the chat
marks each human message with the agents that read its current version.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Write-granted runs whose directories overlap (same folder, symlink or
junction, or nested) wait for each other, at claim time and for manual
runs.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claims skip agents that already have a running run, manual runs refuse a
busy agent, a run whose task was reassigned fails before executing, and
tasks with a run in progress cannot be deleted.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
create_task now pins upstream files by content hash and carries the
source task, source thread, verified facts, open questions and
acceptance criteria into the assignee's prompt.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The three third-party keys were hardcoded or read from the environment.
Operators now enter them in Settings > Integrations; they are stored
encrypted in the app-wide integration_keys table (migration 058).
An agent holding connector grants fails loudly when no Composio key is
saved. The goal-review test no longer depends on absolute dates.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Wake runs stamp sourceSenderType into their secret scope, and the claim
query orders interactive-first: a wake born from an employee or board
message means a person is waiting at the keyboard, so it outranks any
autonomous backlog; within each lane FIFO holds. Non-chat runs carry no
stamp and stay in the autonomous lane. With wake expiry this closes the
live incident: the founder's CEO DM now claims on the next tick instead
of 13th in line. Red-green tested.
Live incident: the founder DMed the CEO and got silence — the reply run
sat 13th in a FIFO queue behind chat wakes from four days ago, which the
reopened app was faithfully executing (necro-replies that woke yet more
agents). A wake run is a conversation reply; past one hour the moment is
gone. expireStaleChatWakes (chat-agent-wake-service) fails queued
chat_mention runs older than the TTL and closes their vehicle tasks;
the worker tick runs it before claiming, with one log line. Real work
(heartbeat/kickoff triggers) never expires. Red-green tested.
buildConversations learns the org's CEO (c_suite, the runtime predicate)
and ranks the viewer's own direct conversation with them above recency —
counterpart resolution already works from the viewer's seat, so agent-to-
agent CEO chats never match. The row carries a CEO badge and, before the
first exchange, invites asking about progress, goals, or direction.
Being first, it is also the default conversation the page opens on.
Two c_suite tools in agent-goal-tools: company_status reports every goal
with real-work task counts (wake tasks and review checkpoints stay out),
work in flight, recent completions, and team headcount, so progress
answers come from the board instead of the model's memory; update_goal
rewrites an active goal's title/description when the founder redirects
the company, audited, CEO-or-owner gated. Both route through the sync
dispatcher; complete_goal now shares the same permission helper.
Every org now keeps one direct chat between the primary employee and the
active c_suite agent: created at onboarding bootstrap (hard failure if
missing), repaired for existing orgs at startup. The existing
wakeDirectCounterparts path already wakes the CEO on every founder
message, so this space is the pinned "talk to your AI CEO" entry the
Messages page will surface next.
Root cause of the stuck kimi-for-coding kickoff (task running forever, no
team spawning): runKimiCodeTurn posted stream:false, and a non-streaming
completion returns headers only after the whole generation — once the CEO's
context grew past ~150k chars, a reasoning turn needed >60s and provider-http's
header deadline aborted it ("Provider request timed out."), failing the run;
every heartbeat retry re-read the workspace, regrew the context, and died at
the same wall until the 3-run budget was gone.
Fix at the family level: accumulateChatStream (chat-format.ts) consumes the
standard OpenAI chunk deltas — content appends, indexed tool_call fragments,
reasoning_content keeping the 60s idle deadline fed — and a shared
runChatCompletionsStream drives makeChatCompletionsTurn (kimi, kimi-cn, xai,
groq, mistral, cerebras, openrouter) and runKimiCodeTurn.
Wire shape verified against MoonshotAI/kimi-cli (kosong streams by default)
and a live streamed probe of api.kimi.com/coding/v1 with the stored plan
credential: role opener, reasoning deltas, tool_call fragments (id+name then
argument text), finish_reason, [DONE] — all reassembled correctly.
verify:ci green including e2e.
- New oauth/kimi-code-flow.ts mirrors the official kimi-cli: form-encoded device_authorization + token polling at auth.kimi.com with the required X-Msh-* device headers, pure poll interpreter, 15-minute tokens refreshed via refresh_token grant
- Agent loop runs chat completions at https://api.kimi.com/coding/v1 with the Bearer access token; default model kimi-for-coding (kimi-for-coding-highspeed for Allegretto+ plans)
- Live model listing via /coding/v1/models; credential refresh wired into the turn credential resolver
- Migration 057 seeds the provider (disabled) into existing organizations
- Wire protocol probed live: device code issued, authorization_pending poll observed, 401 envelope on the coding API confirmed
Adopted from studying Claude Code's /goal (completion judged by a fresh
model, never the one doing the work) and pi-goal-x's auditor (isolated
session, read-only tools, explicit approval marker, rejects scaffold-only
and proxy-metric completions).
complete_goal no longer closes a goal on the caller's word:
- It now takes a required `evidence` claim, and permission/argument gates
run before any audit spend.
- The audit is an isolated tool-turn loop: fresh conversation (the auditor
never sees the executor's transcript), read-only workspace tools (no
run_command), strict auditor instructions, at most 6 iterations.
- Only an explicit <approved/> closes the goal (recorded with the audit
verdict). <disapproved/>, no marker, or an exhausted budget all leave the
goal active and hand the decisive gap back to the caller as the tool
result.
- The handler moved from the sync dispatcher into the run loop's async
routing because it needs provider access; the goal-review prompt warns
that substanceless claims will bounce.
AGENTS.md documents both this and the web-access contract.
"Be ashamed to guess APIs blindly; be proud to look things up carefully" —
until now the agents had no way to look anything up: no search, no fetch,
so market research, API usage, and competitive claims were all invented
from model memory.
agent-web-tools.ts, carried by every agent in every run (same pattern as
the workspace tools, no grant needed):
- web_search: keyless DuckDuckGo HTML backend; parses titles, unwraps the
/l/?uddg= redirect hrefs, pairs snippets; reports honestly when nothing
parses instead of fabricating results. Verified against the live endpoint.
- fetch_url: readable-text extraction (dependency-free HTML-to-text with
entity decoding), 20k-char page cap, 2MB body cap read in chunks, 30s
timeout.
- SSRF guard on both: hostnames resolve before connecting and loopback,
private, link-local, CGNAT, and reserved ranges are refused — re-checked
on every redirect hop, so a public page cannot bounce the agent into
127.0.0.1 or the cloud metadata address. IPv6 and v4-mapped forms covered.
- The fetcher is injectable, so the test suite covers the guard, redirect
policy, truncation, extraction, and the DuckDuckGo parser without
touching the network.
The board draining to 280/280 done left the company idle under a
still-active goal, with 24 markdown files and zero code as its life's work.
Two structural gaps caused it: goals had no lifecycle (status was written
once as 'active' and never read again — nothing re-engages anyone when the
last task closes), and nothing anywhere distinguished plans from product,
so every "engineering" objective closed on documents.
The loop, closed:
- tasks.origin (migration 055) records provenance: chat_wake for mention
vehicles (backfilled from their runs), goal_review for checkpoints, null
for real work. Chatter and checkpoints can therefore never count as
progress.
- spawnGoalReviews (each worker tick): an active goal with zero open tasks
and real progress since the last review gets ONE review task for the goal
owner. The review demands evidence — list the workspace, read the code,
run the tests; documents are how work starts, not proof it happened — and
must end in exactly one of: complete_goal, or the next wave of tasks with
requires_filesystem engineering. A review closed with neither spawns
nothing: termination is structural, not a counter.
- complete_goal (CEO or goal owner) finally lets someone say a goal is
achieved, audited.
- CEO and department-head voices now state the doctrine: a goal that
promises software closes on code that builds and passes tests in the
workspace, and implementation is delegated with requires_filesystem: true.
On the live profile this arms one review: the active platform goal has 0
open tasks and 21 real completed ones, so the CEO gets the checkpoint on
next launch and the company starts building instead of stopping.
Every "engineering" deliverable the company ever produced was a markdown
skeleton, because documentation was the ceiling of the toolset:
write_deliverable and create_document were the only artifact tools. An
engineer could not see the files that exist, read a line of code, or run a
single command — and create_task hardcoded requires_filesystem to false, so
the entire approval-gated read_write mount plumbing was dead code.
New workspace capability (agent-workspace-tools.ts):
- list_files / read_file: explore the bound workspace, traversal-guarded by
the same resolveWorkspacePath used for writes, size- and count-capped.
- run_command: real shell execution with the workspace as cwd, 120s timeout,
16k output cap. Gated on the task's approved filesystem grant — the same
requires_filesystem flag that flips the mount to read_write and routes the
task through operator approval (auto-approved only in AI-CEO mode, which
is that mode's explicit contract).
- create_task now takes requires_filesystem, so the CEO and officers can
delegate engineering work that is allowed to change code and run tests.
- Prompts brief every agent on the tools, and filesystem tasks state the
bar directly: real output is code that builds and tests that pass — a
document describing the work does not close an engineering task.
Tests: tool registry, dependency-folder-skipping listing, traversal refusal,
the run_command access gate plus exit-code reporting, and the create_task
approval path for filesystem work.
The two real goal tasks from July 9 each got exactly one session: the run
completed, the agent never moved the task off backlog, and the heartbeat's
"has a completed run" exclusion locked them out forever. The agents could
not have closed them anyway — update_task_status requires a task_id the
prompt never carried.
The progress contract, made explicit end to end:
- The task prompt now carries "Task id: ..." and tells the agent exactly how
to end a session: done via update_task_status when the work is complete,
in_progress with a note on what remains when it is not.
- The heartbeat re-enqueues any open task with no run in flight — a
completed run that left its task open earns a follow-up session, with the
prompt stating how many prior sessions were spent so the agent continues
instead of redoing work. The retry budget (3 non-transient runs) still
bounds every task; a task that spends it parks with one log line naming it.
- Scripted agents in the integration suites now close their tasks the way
the contract expects real ones to.
