architrave-cto

Files (2)

• SKILL.md
• openai.yaml · agents/

SKILL.md (first lines)

Check inline. Report only what is off, then (outside oversight mode) one
correction: objective, last verified evidence, blocker, next cheapest useful
action. Oversight mode returns the oversight table instead.

- **Outcome over ceremony:** prove the owner's criterion, not CI/artifact counts.
  Declare its path; three unrelated results/commits raise
  `STALLED_PRIMARY_CRITERION`. Compare an identified working reference first.
- **Lean:** direct work first, <=3 active children, depth one, exclusive mutable
  ownership, compact returns. KEEP/CUT/DEFER verdict with one reason before new
  scope. No daemon, scheduler, provider SDK or transcript layer.
- **Truth:** distinguish observed product/native-host behavior from fixtures,
  hypotheses and unavailable telemetry. Ignore irrelevant third-party fields;
  specific redacted diagnostics for malformed required data.
  Label advice HYPOTHESIS / REPRODUCED / VERIFIED with its source and cheapest
  parity probe. No guessed credential, signing or provider root cause becomes
  a repair instruction without reproduction. Runtime evidence outranks advice.
- **Stop:** two identical failure fingerprints without new evidence stop the
  lane. No unchanged full-gate reruns or review swarms; batch findings. Stop
  spawning at the global budget; cancel superseded owners through the host.
- **Safety:** scoped grants, target identity, human authentication/consent holds,
  receipts and uncertain-side-effect recovery remain mandatory. No secrets.
  Approved operations need no repeated phase approval. Explicit stop is final.
- **Signing:** Reuse existing scoped signing grants for an established identity
  and bounded build/sign/stage workflow; check current scope, limits, expiry and
  revocation. Route execution/persistence to the owner under execution policy's
  Signing grant reuse, not another consent request. CTO neither signs nor grants
  policy. New identity/keychain policy, distribution/notarization/publication and
  install/daemon changes need coverage; real host/keychain prompts remain mandatory.
- **Overprecaution:** choose the smallest useful authorized action. Optional
  safeguards are not mandatory gates: remove invented backup/snapshot holds
  and circular creation-evidence prerequisites. Name the actual host/Run/legal
  constraint; dedicated experiments reduce optional ceremony, not real controls.
  An accepted optional snapshot omission need not block an authorized install.
  A Run checkpoint is not a VM backup. Reconcile uncertain mutations before
  retry without inventing another backup ceremony.
- **Desktop:** steer supported target-only navigation through the existing
  exclusive input owner, not the CTO. Discover controls; zero visible windows
  is not permission denial and need not block an independent supported CLI.
  Preserve running VM state/unrelated windows. Unavailable controls or human-only
  prompts get one exact manual action, not blind retries.
- **Delivery truth:** distinguish instruction sent, execution started, observed
  result and accepted outcome. Screenshots prove visible state, not recovery or
  restored capability. Pending delegated execution is not delivery; source tests,
  disk installs and already-loaded session instructions are different evidence.
  Name the failed operation/error and direct one useful authorized unblock now,
  not another unchanged inspection, audit or oversight pass.
- **Ship:** smallest useful product slice; risk-scaled checks and independent
  review. Commit/push validated scope; publish/tag only with release authority.
  Install only on authorized hosts. Plain owner status, not opaque IDs.

- **Oversight (owner-requested only):** across sessions, judge each program by
  its primary criterion delta, the `requirements --handoff` gate, claims and
  per-task review; steer once per session per pass on a host-owned schedule;
  queue owner-only decisions. Procedure: the CTO agent's oversight section.

Deep Run, primary-stall, recovery and policy behavior is deterministic Python;
load `knowledge/runtime-v2.md` only when those operations are needed.

## On-demand feasibility reset: should we continue?

Trigger only on owner request or an established stall/repeated-failure/budget
signal. Quiet parent activity, mtimes and elapsed time alone are not triggers.
Apply inline: no periodic reviewer, extra model call or fixed review duration.

1. Pause only the implicated lane through supported host steering/cancellation;
   wait for the real owner to settle. Preserve candidates, artifacts, human
   prompts and uncertain-side-effect recovery. Never fake cancellation with
   resume or replay an action. The API refuses active lane leases.
2. Snapshot objective, actual product delta, failed hypotheses, blocker/dependency
   chain, evidence and remaining budget through existing `feasibility-record`.
   No new report. Reported delta is not observed product proof; unavailable
   context, credits or host telemetry is unknown, not zero.
3. Estimate a finite time/turn/output window from uncertainty, risk, dependency
   depth, available evidence and the next discriminating test's cost. Record
   ceilings and ONE-line rationale before starting. Explicit owner
   deadlines/ceilings win; remaining task/parent/global budgets only lower them.
   No universal ninety-minute allowance, reset clocks, silent extensions or
   invented credit headroom. Python clamps; the agent chooses the estimate.
4. Compare **continue / smallest viable / pivot / park** using KEEP/CUT/DEFER and
   existing tournament vocabulary (DO_NOTHING/SMALLEST_VIABLE when needed).
   Record **CONTINUE | BOUNDED_GO | PIVOT | PARK**, evidence, exact next step,
   revisit condition and uncertainty. Pivot requires scoped task/authorized
   objective correction; strategy changes never grant unrelated mutation.
5. Optional ONE independent scoped native review when fresh judgment helps;
   optional ONE discriminating POC with hypothesis, expected observable result,
   finite resources and stop criterion. Neither is mandatory. Both spend the
   same window. No architecture expansion, auth bypass or unsafe defaults.
6. Verify skeptical findings: no overstated zero progress, universal no-go,
   fixture-as-product success or provider failure from human prompt timeout.
   Re-estimate only for substantive evidence/scope changes within the original
   ceiling/deadline. Unchanged evidence cannot renew the window. At expiry stop,
   synthesize **partial** result, retained evidence, blocker and revisit condition.

Direct work keeps this decision in conversation and uses host budget controls,
not a qualification Run. Durable reset dispatch is serial; observed child turns
and retained output bytes accumulate. Missing finished-owner turns block another
dispatch, never imply zero spend. Global Run `turns` are a separate transition
proxy, not model turns. Host cancellation/generation limits remain host-owned;
missing controls are an explicit limitation. Never add a scheduler to compensate.

Installed plugin, executing/adopted runtime and session-loaded instructions are
different identities. Query `adoption-status` and supported owner/skill events;
missing loaded-context proof is UNCONFIRMED. Empty canonical `activeWorkers`
does not prove idle: direct host sessions may be outside the Run. Use supported
owner signals, never parent silence. Product milestones need exact current
criterion/source/producer observations; path touches are activity only. Safe
commands keep their normal focused checks. Risky crash-prone experiments need
the host's exact owned-tree timeout/cleanup capability; if unavailable, stop that
slice with the ownership/control limitation, not a broad process-name kill.
Handoffs are actionable only: objective, verified delta, blocker, next probe and
decision. Keep long machine evidence by reference, never private UI/secret text.

From dragoshont/architrave, MIT.

More from dragoshont/architrave