| name | onboard |
| description | Seat-onboarding checklist for SCMessenger - load when taking the CTO or orchestrator seat in this repo, before any dispatch, merge, or verdict. |
onboard — taking a seat in this repo
Run this checklist IN ORDER when taking any seat (CTO or orchestrator) in
SCMessenger, on any frontend. Each step exists because skipping it already
cost a session.
1. Load state — read, then re-derive
Read the top of HANDOFF/CTO_STATE.md first: the standing rule, the handoff
banner, merge authority. Then re-derive live state, because the file ages and
the repo does not:
git fetch origin
gh pr list --limit 10
gh pr checks <open PR> # for each open PR
gh run list --branch main -L 5
git log --oneline -8
Derive from origin/main, NEVER from the shared checkout — it can be dozens
of commits behind, and acting on it merges stale verdicts.
2. Session launch audit
Confirm the orientation hook output appeared at session start — its absence
means the hook did not run, so the gates below may not be attached either. If
a model gate hook blocked the session, re-launch with the exact model it
names; do not work around the gate.
3. Gate inventory — verify before relying
A gate you have not confirmed exists is a gate you do not have. Check each one
before relying on it:
Scripts (scripts/): pr_scope.sh, check_wiring.py, orchestrator_guard.py,
apply_branch_protection.sh, delegate.py, lane_probe.py, agy_run.sh,
rules_check.py, orchestrate_strict.py, orchestration_contract.py,
session_orchestration_audit.py, reclaim_safe.py.
Hooks (.claude/hooks/): model_gate.sh, session_orientation.sh,
preflight_guard.py, check_no_emoji.py.
Git hooks path: git config core.hooksPath must point at .githooks —
without it the pre-commit and pre-push guards are not attached at all.
4. Live traps — already paid for
Re-reading them here is cheaper than re-learning them:
- GitHub runner hangs:
gh run cancel <id> then gh run rerun <id> --failed.
- Lint red on every PR, including markdown-only ones:
cargo deny / RustSec
environmental, not your regression.
adb logcat main buffer hides crashes: use -b crash first.- Never set
CARGO_TARGET_DIR for Android gradle builds — the APK ships
without the .so while gradle still says BUILD SUCCESSFUL.
- agy has a ~90s per-tool timeout separate from
--print-timeout; never ask
it to run a cold full build.
- zai lanes need thinking disabled or they return empty content — a silent
vacuous success, not a refusal.
- Machine-verify fixtures and read the output before acting.
- Consumer searches must include
tests/ directories.
- Wiring verdicts come from
python scripts/check_wiring.py (exit 1 =
findings), never by eye.
5. Queue
HANDOFF/todo/_QUEUE.md is the backlog order;
HANDOFF/V1_0_0_EXECUTION_PLAN.md is the sequencing authority. Do not
re-derive priority from anywhere else.
6. Session close
Update HANDOFF/CTO_STATE.md — standing rule, section 0-rule: immediately on
any important change, not batched to the end. A session that dies mid-run
leaves the next one reading fiction.
Run python scripts/session_orchestration_audit.py to audit delegation.
Caveat: its STATUS column is not trustworthy; the token/step accounting is.
