handle_verdict(approved): removed set_issue_in_progress(work_item_id) before
_try_advance_stage. _try_advance_stage -> advance_stage -> plane_notify_stage
already PATCHes the issue to the NEXT stage status, so the reset only made the
board flicker In Progress before the next stage (part of bug 3).
Status-only verdict model: comments NEVER drive the pipeline. Removed the
whole comment-based control mechanism from handle_comment (:approved: /
:rejected: / answer-to-questions) which caused bug 3 (echo self-hit): the
analyst posts its own "waiting for approval" comment, handle_comment catches
its own comment and reverts In Review -> In Progress. handle_comment is now a
pure logger with no side effects.
handle_status_start: a return to In Progress on an EXISTING task (Slava
answered the analyst questions in Needs Input) now RELAUNCHES the stage agent
instead of being a no-op. Distinguished from a duplicate In Progress webhook
via has_active_job_for_task() (new db helper): no active job => agent idle =>
relaunch; active job => busy => skip (no double launch).
ET-006 was handed to two different tasks because M-6 derives work_item_id from
the Plane sequence_id, which can collide -> the two tasks shared a branch/worktree
slug prefix and stepped on each other.
2a: ensure_unique_work_item_id() is a uniqueness-guard LAYERED ON TOP of the M-6
derive (derive is untouched): if the derived ET-NNN already exists in tasks for
the repo, it walks forward to the next free number. Applied in start_pipeline
after the derive.
2b (defense-in-depth): worktree is keyed by branch; if the resulting branch is
already owned by another task in the repo, disambiguate it with the unique
work_item_id + plane id so two tasks can never share a worktree.
Plane issue.updated (status -> In Progress) ships only changed fields, so the
webhook payload has no description and QG-0 wrongly blocked issues. start_pipeline
now pulls the full description from the Plane issue detail API (reusing the same
GET endpoint + shared token as fetch_issue_sequence_id) when the payload field is
empty/short, before QG-0 runs. Empty API -> honest QG-0 fail (truly empty ticket).
A pytest run on prod was sending REAL Telegram messages to Slava: some tests
(e.g. test_webhook_dedup advancing a stage) reach notify_stage_change ->
send_telegram, which read the live .env token/chat_id and actually POSTed.
Add an autouse fixture stubbing send_telegram to a no-op for every test. Patch
the SOURCE src.notifications.send_telegram (covers all notify_* helpers and the
many modules that do a local from .notifications import send_telegram inside
functions) AND src.stage_engine.send_telegram (module-level binding, would not be
intercepted by the source patch alone). webhooks/plane, launcher, queue_worker are
patched defensively with raising=False.
Verified: full suite run with FAKE telegram creds + an un-swallowable httpx.post
trip-wire (BaseException, so send_telegram except Exception can not hide it) shows
ZERO calls to api.telegram.org. Without the fixture the trip-wire fires, proving
the guard is real.
Feature 4. claude is now launched with --output-format json; the run-log trailing
result JSON is parsed (defensively, never fatal) for usage + total_cost_usd. New
idempotent ALTERs add input_tokens/output_tokens/cache_read_tokens/cost_usd to
agent_runs; the launcher monitor records usage per run, posts a per-agent finish
comment under that agent bot (e.g. Developer gotov · 45.2k in / 12.1k out · $0.21),
and the deployer posts an end-of-task summary (SUM over agent_runs GROUP BY agent)
on done. New src/usage.py holds parse/format/record/summary helpers; test_usage.py
covers parsing a real CLI JSON blob, NULL-on-garbage, recording, formatting, and the
per-task aggregate.
Feature 2. The issue updated dispatch (shipped with the status-trigger handler)
also routes Approved -> _try_advance_stage (== :approved: comment) and Rejected ->
_rollback_stage (== :rejected: comment). The :rejected: comment branch was
refactored into the shared _rollback_stage so both mechanisms behave identically;
a status reject passes Reason: (rejected via status, see latest comment) since no
inline reason arrives with a status change. Comments stay fully working. This
commit adds test_verdict_status.py proving both status and comment paths funnel
into the same advance/rollback logic.
Feature 1. work_item.created no longer starts the pipeline (soft QG-0 log only);
the issue stays in the backlog until moved to In Progress. The pipeline-start body
is extracted into start_pipeline(); a new issue updated handler routes a state
change to In Progress -> handle_status_start, which is idempotent: an existing task
for the plane_id is NOT re-created or restarted (protects handle_comment, which also
flips issues to In Progress). Real Plane payload: event=issue, action=updated,
data.state.id. Existing m6/plane_webhook/dedup tests updated to drive the new
trigger; new test_status_trigger.py covers created-no-op / start / idempotent.
Feature 3 + Feature 2 infra. Extend the global PLANE_STATES with the 6 new
enduro status UUIDs (architecture/development/review/testing + approved/rejected),
remap STAGE_TO_STATE so the 4 mid-pipeline stages move the issue across its own
board column instead of all sitting in In Progress, and add the
set_issue_stage_state() helper. Needs Input / In Review / Blocked keep their own
explicit setters and stay higher priority. TODO(ORCH-10): statuses are per-project;
resolve per project when more projects are onboarded.
add_comment now accepts an optional author (agent role) and POSTs under the matching Plane bot token via _headers_for(), so Plane shows the real author (Analyst/Architect/Developer/Reviewer/Tester/Deployer/Stream) instead of a single shared account. Unknown/empty roles or missing tokens fall back to the shared orchestrator token (autonomy preserved). GET/PATCH (find_issue_id, set_state) are unchanged and stay on the shared token. Call sites in stage_engine, launcher, webhooks/plane and the plane_sync notify helpers now pass author by stage role; stage transitions use stream. Adds tests/test_plane_author.py.
Add 7 optional bot-token fields (plane_bot_analyst..stream) read from the ORCH_PLANE_BOT_* env vars, default empty. Required for per-agent comment authorship; empty values fall back to the shared orchestrator token.
18 tests: happy-path advance per stage with correct agent (ORCH-4 fix),
QG-fail no-advance, reviewer REQUEST_CHANGES rollback+retry/alert, tester FAIL
rollback+retry/block, architect conflict rollback to analysis, analyst
approved-flow no-advance, and launcher+plane both delegating to the engine.
launcher._try_advance_stage and plane._try_advance_stage are now thin
wrappers over stage_engine.advance_stage. The plane webhook calls the sync
engine via asyncio.to_thread so there is exactly one implementation. The
launcher forwards finished_agent so the agent-specific rollback branches still
fire; the webhook passes None (human :approved:), matching prior behavior.
Also fixes the agent-selection bug in the launcher path: it used to enqueue
get_agent_for_stage(next_stage) (skipping a stage, e.g. analysis->architecture
launched developer instead of architect). The unified engine uses
get_agent_for_stage(current_stage), consistent with plane and gitea.
Merge the two diverged _try_advance_stage implementations (launcher sync +
plane async) into one synchronous engine. Preserves all launcher business
logic (analyst approved-flow, reviewer REQUEST_CHANGES rollback+retry, tester
FAIL rollback+retry, architect conflict rollback) and the plane
check_review_approved PR-by-branch dispatch. Unifies the QG signature
dispatch. Fixes agent selection: advancing FROM current_stage launches
get_agent_for_stage(current_stage), not next_stage.
The watchdog used to time.sleep(timeout) then immediately SIGKILL, which cut
claude off mid-write and left half-written artifacts. It now sends SIGTERM,
polls os.kill(pid, 0) for up to agent_kill_grace_seconds, and only SIGKILL if
the process is still alive; ProcessLookupError is tolerated at every step.
Timeout is now configurable via config.py: agent_timeout_seconds (default 1800),
agent_kill_grace_seconds (default 20), and agent_timeout_overrides_json for
per-agent overrides (e.g. {"reviewer": 3600}). AGENT_TIMEOUT is kept as a
backward-compatible alias. The recorded exit_code stays -9 so the ORCH-1
monitor retry/fail logic is unchanged (timeout-kills classify as permanent and
requeue within max_attempts, no retry loop).
_auto_merge_pr had zero callers (merge is handled by the deployer agent).
Removed the method; _ensure_pr (still used by the auto-advance path) is kept.
Container ORCH_CLAUDE_BIN pointed at a non-existent /usr/bin/claude while the
launcher spawns the hardcoded /opt/claude-code/bin/claude.exe. Preflight now
follows AgentLauncher.CLAUDE_BIN (the genuinely executed path), so it no longer
falsely blocks every job in production.
_finalize_job classifies the run log: transient (429/overload) -> backoff
requeue via mark_job_transient with separate transient_attempts budget honouring
Retry-After; permanent -> normal attempts<max. on_outcome callback feeds the
circuit breaker. _backoff_seconds = min(2^n*base, max) | Retry-After.
queue_worker.QueueWorker drains the queue respecting max_concurrency. main.py
lifespan: queue-recovery (requeue running jobs) after M-1 orphan-recovery, starts
worker and stops it on shutdown. New GET /queue endpoint (counts + recent jobs).
All 8 webhook launch points (plane x4, gitea x4) now enqueue a job and return
immediately instead of synchronously spawning claude in the uvicorn process.
ORCH-6: ARCHITECTURE.md gets a project-registry section; README explains
how to add a project via ORCH_PROJECTS_JSON; BUGFIXES_2026-06-03.md
records the fix and links the 2026-06-02 webhook autorun incident.
