The catch block in wakeAgentMentions now records wake_state on each
mention (woken/failed) instead of silently discarding the error. Failed
mentions render with a strikethrough and muted color so the user sees
which agents could not be reached, with a tooltip on hover. Migration
044 adds the wake_state column to chat_mentions.
Agents can now call external tools (GitHub, Slack, Gmail, etc.) via
Composio's managed API. A thin HTTP client calls the Composio REST API
directly — no heavy SDK dependency. External tools are loaded per-org
from composio_settings and merged with internal tools at run time. The
agent run loop routes composio_* tool calls to the Composio execution
endpoint while internal tools dispatch as before.
Includes: migration 043 for composio_settings table, IPC handler for
saving Composio API key and toolkit configuration, and the composio
client/bridge modules.
Agents now have a read_messages tool to check their inbox and channel
history during runs, enabling them to catch up on conversation before
acting. Task prompts are rewritten to encourage checking messages first
and discussing work with teammates rather than posting status updates.
MentionTextarea inserted raw agent names (e.g. @Sarah Chen) but the
backend regex captured only the first word (@Sarah), causing mention
resolution to fail silently. Moved canonicalMentionHandle to shared/
so the renderer can produce space-free handles that the backend resolves
correctly.
ChatService was 542 lines, mixing message delivery with the dispatch of
nine slash commands. Move the command cluster — /board, /approve, /deny,
/agent, /skill, /file, /runbook, /handoff, /escalate — into a
ChatSlashRouter that owns the approval and board-intervention services it
needs. ChatService keeps message delivery and delegates two calls.
Behaviour identical: chat-service 376 lines, router 196, full suite and
e2e green; the slash-routing governance test exercises every branch.
OrganizationService was 552 lines, over the 500-line ceiling. The
80-line createOrganization routine — seed the org, the Board Chair,
identity/governance/operations/SDLC/policy defaults, starter channels,
the provider catalog — is a self-contained concern. Move it to
organization-seed.ts; the service now delegates and drops the two
service fields (operations, sdlc) only that routine used. Pure
relocation, behaviour identical: 469 lines, full suite and e2e green.
The Codex flow guarded with `result.state && result.state !== state`,
so a callback that simply omitted state skipped CSRF validation while a
sign-in window was open. Match the Anthropic and Google flows: compare
strictly, so a missing or mismatched state is rejected. PKCE already
blocks code injection; this closes the login-CSRF gap.
writeDeliverable guards against path traversal, but recordWorkspaceFile
resolved and readFileSync'd its path with no containment check — safe
only because today's lone caller hands it the already-sanitized result.
Any future caller passing a raw agent path (../../.ssh/id_rsa) could read
an arbitrary file into the org's file store.
Extract the resolve+reject-".."/absolute guard into one workspace-path
helper and route both the writer and the reader through it, so the
defense lives on the function rather than the caller.
send_message answered the agent with "Posted to #x (message_id=...,
thread_id=...)" — a transaction receipt that gives no sense anyone is
there. An agent posting into that void has no reason to converse.
Now it answers with presence: who was addressed and can reply, who else
is in the channel, and the thread_id to stay in the conversation. The
message_id noise is gone. The tool's own description now frames it as
talking to teammates rather than posting a coordination status.
A woken agent was handed a "Chat reply" ticket whose description was
four source ids and "you were @mentioned... reply", then a closing that
told it to "report progress and completion" and "update the task status"
— so it answered like a status bot.
Now the wake reason names who spoke, in which channel, and quotes what
they said, and asks the agent to reply to that person in their own voice.
When a run carries a thread, buildTaskPrompt ends as a conversation
("reply to the people here... not a status report") and skips the work
closing entirely. The ordinary work closing also drops "reporting your
progress and completion" for "tell the team what you did and what's next,
the way you'd bring a colleague up to speed." The thread_id reply token
is preserved. This is the second half of making the company feel human.
The system prompt was a single flat role line, so every agent sounded
like the same status bot. Replace it with an identity (name, title,
department, company, "a colleague, not a generic assistant") and a voice
that differs by standing — executive, manager, or individual contributor
— plus an explicit "talk like a person on a team, not a status bot"
directive. The classification is derived from existing fields (top of
org = CEO, has reports = manager), so there is no schema change.
This is the first half of why the old app felt human and the new one
did not: identical model calls now carry a distinct person behind them.
The manual runTask path enforces assertExecutionApproval, but the
autonomous heartbeat path (enqueueHeartbeatRuns -> enqueueRun -> the
queue worker) did not, so a high-risk or filesystem task whose approval
was still pending — the normal state in ai_cofounder mode, or after an
operator declines — was executed unsupervised.
Apply the same gate in enqueueRun. Pending/declined approvals leave the
task unqueued (the heartbeat's existing try/catch skips it) and it is
picked up on the next tick once approved. ai_ceo mode auto-approves at
task creation, so autonomous companies are unaffected.
An @mention by an agent wakes the mentioned teammate by creating a task
run, and the woken agent's reply can mention back, an unbounded A->B->A
loop whose only backstop was the per-provider monthly request cap (which,
once tripped, fails every agent on that provider). Before enriching agent
conversation this had to be made safe.
Cap the agent-authored messages a thread may accumulate since its last
human/board message; past the cap, agent-triggered wakes pause until a
person re-engages. A human message always wakes mentions and resets the
budget. Counted from chat_messages, so no schema change.
Root cause of the GPT-account "empty turn: no text and no tool calls"
failures, found by replaying the real stored response: the CEO actually
called hire_agent five times, but the ChatGPT/Codex backend delivers each
output item in a response.output_item.done event and then sends a
response.completed whose response.output is []. accumulateResponsesStream
read only the completion event, so every tool call was discarded and the
turn parsed as empty — the run failed and the company never staffed.
Collect items from output_item.done and fall back to them when the
completion carries no output of its own; keep using the completion output
when it is populated (non-Codex Responses backends). Verified against the
real captured stream (5 hires recovered) and unit-tested with the exact
event shape.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
The GPT-account (Codex) path supports tool calls and worked before; my
forced tool_choice broke it and was the wrong fix. The real regression was
strict:null in the rebuild's tool definitions, which made the endpoint drop
the tool list so the model had nothing to call — a "Done" run that hired no
one. Forcing then turned that silent failure into a red one.
Restore the original "auto" loop (no forced tool_choice, no finish tool) and
keep only the one-line strict:null omission. Codex now receives valid tools
on the working "auto" choice, as in the version that worked.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
The gpt-account (Codex) turn built its function tools with strict:null. The
Responses API expects strict to be a boolean or omitted; null is malformed,
so the endpoint dropped the whole tool list — leaving the model with nothing
to call. That is exactly the "model returned no tool calls" failure on the
GPT-account path. Omit strict when null; keep the boolean otherwise. Locked
with a test, alongside the Bedrock/Gemini forced-tool-choice encoding.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Forcing only the opening turn let a weak model take one token action and
stop, so a kickoff hired one agent (or none) and went quiet. Now every turn
is forced (tool_choice required) and the agent ends only by calling the new
`finish` tool — so it must keep hiring, delegating, and messaging until the
work is actually done. Finishing with zero actions, or returning no tool
call at all, fails the run loudly instead of faking success.
Prompts tell each agent it must call a tool every turn and finish only when
done. Tests drive weak-model stubs that act then finish, proving a full team
forms rather than a single gesture.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
The whole premise is cheap models + skills + a team — so the control plane,
not the model, must guarantee action. Every adapter hardcoded
tool_choice:"auto", letting a weak model answer the kickoff with prose and
"succeed" having hired no one, delegated nothing, and said nothing.
Now the agent loop compels a tool call on the opening turn (tool_choice
plumbed through every backend: OpenAI/Responses/Chat, Anthropic, Codex,
Copilot, Azure, Custom, Gemini ANY mode, Bedrock any). If the forced turn
still yields no tool call, the run fails loudly instead of recording a
deceptive success — surfacing the problem (let it crash) rather than a
silent dead company. After it has acted, the model chooses freely.
Tested: a stub that rambles on "auto" but acts when forced now posts a
kickoff message; a prose-only first turn fails the run.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Engine-side writes - hires, delegated tasks, chat messages, run state -
only reached the renderer after a run finished: the queue worker sent
one snapshot-changed at the end, and nothing fired at claim time, so
the window sat frozen for the minutes a CEO run takes unless the user
reloaded by hand. The execution service now takes a notifyDataChanged
callback, the agent loop fires it after every tool-bearing iteration,
and the worker announces the claim itself, so the UI follows the
company working in real time.
Interrupted runs now return to the queue instead of failing - an app
shutdown is not a model error and must not spend the task's bounded
retry budget. Startup work moves behind the single-instance lock so a
losing second instance can never run recovery against the live
instance's database. Background provider requests that died before
receiving an external response id now fail at boot (the background
worker can only resume accepted ones). Agent updates resolve the
department through the same canonical lookup as hires, replacing the
free-text write and its invented "General" default.
A reasoning-heavy model can burn the whole max_tokens budget on
thinking blocks; the parsed turn then has neither text nor tool calls
and the loop treated it as a successful completion, marking a run done
when nothing happened. Such a first turn now fails the run so the
failure is visible and the heartbeat retries; later empty-text turns
after real tool work still complete normally.
The CEO's first run hired five reports into departments "general" and
"engineering"; the hiring-policy lookup was exact-match against the
seeded names (Executive, Engineering, ...), so every hire failed and
the kickoff produced nothing. The tool schema now enumerates the
canonical departments, the policy lookup matches case-insensitively
and stores the canonical name on the agent, and an unknown department
error lists the valid choices so the model can self-correct.
The abort timer was cleared as soon as fetch resolved with headers, so
a server that stalled the response body hung the request forever - and
one hung run froze the autonomous queue behind the worker re-entrancy
guard. JSON requests now carry a header deadline plus a longer body
deadline on one abort controller, and event streams fail when no chunk
arrives within an idle window, cancelling the reader. Timeouts surface
as ProviderHttpError so the run records a clear failure and the
heartbeat retries.
Adversarial review of the kickoff feature surfaced two gaps. First, the
kickoff task was inserted with raw SQL, so a workspace-bound kickoff
carried requires_filesystem without the execution approval every other
task receives - manual runs failed and review-mode orgs had no approval
trail. Kickoffs (in-app and onboarding) now go through createTask and
its policy gates. Second, a goal created before any agent existed
stayed inert forever; creating an agent now operationalizes every
active goal that has no task yet. Goal owners must be active agents:
an archived owner falls back to the top-level agent and the owner
dropdown only offers active agents.
Every source file is now under 500 lines. Task prompt construction
moves from task-execution-service into task-prompts; goal assignee
resolution moves from organization-service into goal-kickoff; the
shared command types become a domain-structured shared/commands
directory with unchanged import paths; the provider adapter contract
types move to provider-adapter-contract, re-exported through
provider-adapters so callers keep a single public surface.
A goal created from the Work page only wrote a goals row; nothing ever
read it, so the company never started working on it. Creating a goal now
produces an "Operationalize goal" kickoff task for the chosen owner (or
the top-level agent when unowned) and re-enables that agent's heartbeat,
so the autonomous queue worker starts the run - the same activation
contract the previous runtime had via goal_activated wakes and CEO
bootstrap. Onboarding shares the same kickoff module instead of its own
copy.
Migration 041 adds parent_task_id to tasks for subtask hierarchy.
TaskDetailDialog now has three tabs: Overview (description, assignee,
project, goal, workspace, runs with error messages), Subtasks (child
tasks with status), and Activity (audit events and approval history).
Inline edit overlay covers title, description, assignee, risk, priority,
and parent task. Task delete is available from the dialog footer.
Agents can now create versioned company documents (PRD, spec, memo,
report, runbook, proposal, policy, guide) via the create_document tool.
Documents are persisted in the documents table and visible in the
Documents page.
UpdateAgentInput now includes department and budgetCents. The edit dialog
on the Team page shows department and monthly budget alongside the
existing role, title, seniority, provider, and model fields.
Full-stack: IPC channels, command types, OrganizationService methods,
IPC handlers, preload bridge, desktop-api types, and UI dialogs on the
Work page. Users can now create goals and projects directly instead of
relying on the onboarding bootstrap.
When a task_run is triggered by a chat @mention, the agent now receives
the full thread transcript (up to 20 messages) in its prompt so it
understands what it is replying to and can use the correct thread_id.
Agents were deaf to @mentions — only /agent slash commands triggered a
task_run. Now ChatService.sendMessage auto-wakes every mentioned agent
(skipping self-mentions and /agent which already handles it) by creating
a queued task_run through ChatAgentWakeService.wakeFromMention.
Every catalog provider can now drive agent execution:
- Azure OpenAI -> Chat Completions at the deployment URL with api-key
- AWS Bedrock -> Converse API signed natively with SigV4 (node:crypto)
- Vertex AI -> Gemini generateContent with a service-account JWT (RS256)
exchanged for a cached access token
The credential resolver maps each cloud profile to the turn's needs; the router
covers all 16 providers. Pure encoders, the SigV4 signer and the GCP auth are
unit-tested; no AWS/GCP SDK dependency.
ZacharyZhang-NY
