Skip to content

RELEASE_CHECKLIST.md

Latest commit

 

History

History
52 lines (41 loc) · 4.68 KB

RELEASE_CHECKLIST.md

File metadata and controls

52 lines (41 loc) · 4.68 KB
Raw

Release checklist

Record candidate-specific proof in RELEASE_EVIDENCE.md; never infer a checked gate from a source-only test.

Automated gates

  • Plugin manifest and skill validation pass.
  • Router and activator tests pass on Python 3.10, 3.11, 3.12, 3.13, and the current runtime.
  • The current six-role routing implementation passes the hosted macOS, Ubuntu, and native Windows Python 3.10–3.14 matrix in run 29233219101; the Autorouter branding/visual candidate repeats the full gate locally.
  • Clean install, identical reinstall, conflict refusal, reviewed replacement, upgrade, repair, deactivation, and modified-file preservation pass.
  • Exception and hard-crash failure injection at every declared failpoint restore the exact pre-operation agent tree.
  • Concurrent lifecycle commands serialize without lost ownership state.
  • Standalone install, exact-owned update, uninstall, durable-controller recovery, and purge pass without overwriting foreign or drifted skill trees.
  • Every declared standalone install/update/uninstall crash boundary rolls back or commits deterministically, and post-crash edits are preserved.
  • Plugin cache contains no source-machine absolute paths, secrets, symlinks, bytecode, or transient state.
  • Installed-cache hooks use only plugin-relative commands and plugin data.
  • Missing retired cache paths migrate to the newest installed hook, emit at most one no-cache diagnostic, and duplicate terminal Stop delivery is idempotent.
  • Nested Codex commands and JSON-wrapped results preserve successful verification evidence; XML-wrapped Stop feedback cannot create a recursive routing turn, and post-route mutations cannot retain an answer check gate.
  • Stop and SubagentStop never emit a block decision across satisfied, unmet, duplicate, active-stop, malformed-child, and routing-council states; the release validator rejects any reintroduction.
  • Offline session repair covers top-level and compacted poison, mixed legitimate content, byte-exact backup, atomic replacement, malformed JSON fail-closed behavior, links, hard links, idempotency, and privacy-safe reporting.
  • scripts/check_hooks.py observes exactly five trusted, enabled, error-free plugin hooks and six active current agent definitions after explicit review on candidate 0.1.0+codex.20260714074631.
  • All README SVGs parse, render at 1,600 pixels wide, retain accessible titles/descriptions, and pass visual clipping/legibility review; all local README links resolve.
  • Routing golden cases, product taxonomy, adversarial prompts, continuations, Fast reporting, and 5,000-case fuzzing pass.
  • ADVISORY, local-agent-ready, authoritative full-ready, and MANAGED_BLOCKED diagnostics are covered.

Real-host acceptance

Complete every row with the released archive or marketplace snapshot, not the source directory. Follow HOST_ACCEPTANCE.md and retain its evidence fields for each row.

OS Surface Install/cache relocation Hook trust Route smoke Activate/deactivate Result
macOS Desktop required required required required pending
macOS CLI pass pass precursor pass pass exact candidate delegation replay pending after hook-adapter change; precursor 0.1.0+codex.20260713113358 passed
Linux CLI required required required required pending
Windows Desktop required required required required pending
Windows CLI required required required required pending

IDE extensions are excluded from the full-plugin matrix because Codex plugins are not currently supported there. The standalone advisory skill still requires an IDE discovery, routing, custom-agent, update, and uninstall smoke on each supported host.

Policy and degraded modes

  • Untrusted hooks are skipped and status does not claim full activation.
  • Changed hook definitions require renewed trust.
  • Hooks-disabled and multi-agent-disabled configurations remain advisory.
  • allow_managed_hooks_only reports managed blocked and never attempts bypass.
  • Missing custom agents remain advisory without breaking normal Codex work.
  • Fast and effort labels never claim unavailable runtime attestation.
  • The public repository or marketplace exposes a tested private vulnerability-reporting channel (GitHub makes repository private reporting available only after public visibility).

Public-release decision

Do not mark stable or publish broadly until every automated gate passes and every real-host row has evidence. A private marketplace beta may proceed with explicitly documented pending rows.