codex-orchestrator 2.0.3 → 2.0.5

package/CHANGELOG.md

@@ -6,12 +6,60 @@ The format is based on Keep a Changelog, and this project follows SemVer.
 
 ## [Unreleased]
 
+## [2.0.5] - 2026-07-28
+
+### Added
+- Added durable continuation of successful direct runs from newly observed,
+  trusted unresolved pull-request feedback. The same run now freezes an
+  immutable batch, reuses affected Closure, reruns checks and Acceptance Proof,
+  and appends one fast-forward commit plus one marker-bound summary to the same
+  draft PR.
+
+### Changed
+- Removed the configured Codex CLI version pin. Setup and runtime now use the
+  installed `codex` command, while the containment canary binds its actual
+  version, canonical executable path and digest, plus the orchestrator package
+  version, and rejects a stale certificate after any of those identities change.
+- The daemon now discovers both `agent:auto` and `agent:review` issues, while
+  suppressing unchanged `review-ready` output only within the current process.
+- Run state now emits schema version 2 and losslessly reads version 1; migrated
+  `review-ready` runs baseline existing feedback without executing it.
+- Replaced overlapping legacy live-smoke profiles and scenario aliases with a
+  single supplemental `v2-regression` matrix whose scenarios each exercise a
+  distinct current V2 policy, recovery, diagnostics, proof, or quality gate.
+- Pinned every model-backed live-smoke operation to real `gpt-5.6-luna`, with
+  per-scenario model audit evidence and deterministic fault injection retained
+  only around the real model result.
+- Reduced the universal proof generation schema from 52 KB to 22 KB by keeping
+  successful-proof semantics in generation while moving duplicated platform
+  combinations to the existing strict runtime validator, reducing
+  structured-output load without weakening final proof acceptance.
+- Treats a zero-exit Codex invocation with a missing output report as a
+  resumable transport failure, allowing the existing bounded retry to recover
+  without weakening malformed-report validation.
+- Repairs one schema-valid implementation report whose cumulative
+  `changedFiles` omits current product changes, without consuming another
+  implementation cycle or permitting worktree mutation; repeated mismatch
+  remains fail-closed.
+
+### Removed
+- Removed unsupported configuration conversion, old workflow-manifest readers,
+  obsolete run-state fallbacks, superseded prompts, and historical workflow
+  planning artifacts. The repository and package now contain only the current
+  V2 runtime and contracts.
+
+### Security
+- Review bodies can authorize repair only when bound to the exact PR head and a
+  fresh immutable-user `write` or `admin` permission receipt. Source, permission,
+  ref, worktree, or publication drift fails closed; no thread-resolution or
+  force-push authority was added.
+
 ## [2.0.3] - 2026-07-18
 
 ### Changed
 - Added durable `agent:waiting-human` questions, current-WRITE trusted answers,
   conflict clarification, permission revocation, and same-run rerouting before
-  implementation. Exact Config V1 now migrates atomically to Config V2.
+  implementation.
 - Replaced the two-file runtime skill lookup with one generated, manifest-bound
   package workflow containing the declared skills, profiles, operation wrappers,
   schemas, and shared review contracts.
@@ -28,441 +76,17 @@ The format is based on Keep a Changelog, and this project follows SemVer.
 ### Changed
 - Re-published the current V2 package line for downstream workspace upgrades.
 
-## [2.0.1] - 2026-07-17
-
-### Fixed
-- `setup --fresh` now checks for a pre-existing V2 owner before acquiring its
-  own setup lock, so a safe Legacy cutover no longer blocks on itself.
-
 ## [2.0.0] - 2026-07-17
 
 ### Changed
-- Replaced the compatibility and experimental runtimes with one public V2 CLI,
-  one strict configuration contract, and one `runIssue` lifecycle shared by
-  direct runs and the serial daemon.
+- Introduced one public V2 CLI, one strict configuration contract, and one
+  `runIssue` lifecycle shared by direct runs and the serial daemon.
 - Reduced the default live release smoke to package install, normal default
   Codex, browser proof, and safety-negative scenarios; broader policy scenarios
   remain opt-in.
 
-### Removed
-- Removed the Legacy CLI, bridge manifest, bundled workflow prompts, alternate
-  planning/scoped execution paths, and their compiled/test assets from the npm
-  package.
-
 ### Security
 - Contained tool environments exclude GitHub, SSH, npm, and cloud publication
   credentials while preserving the explicitly accepted shared Codex-auth and
   same-user local-read behavior. Proof rejects credentials in every text
   artifact and applies public-only host-identity restrictions to evidence.
-
-## [0.1.51] - 2026-07-15
-
-### Fixed
-- Bridge publication normalizes the package CLI executable mode before hashing,
-  so installed-package verification matches npm's required `bin` mode.
-
-## [0.1.50] - 2026-07-15
-
-### Added
-- Added a target-scoped shared/exclusive activity fence for daemon lifetime,
-  targeted claims, setup, and skill-runtime-v2 preparation.
-- Added `setup --prepare-skill-runtime-v2`, strict Darwin/Linux legacy-process
-  drain checks, canonical prepared-generation evidence, and a verified
-  `bridge-runtime.json` package manifest.
-
-### Changed
-- Runner state can preserve a forward-compatible empty version-2 envelope while
-  bridge config remains version 1; bridge execution still stores only legacy
-  run records and does not change prompt or Codex transport behavior.
-
-### Security
-- Skill-runtime-v2 preparation now fails closed on foreign or ambiguous process
-  ownership, active local/GitHub claims, GitHub read failure, and bridge package
-  provenance mismatch.
-
-## [0.1.49] - 2026-07-15
-
-### Added
-- Added a Runner-owned Publication saga that pins validated candidates and
-  reconciles branch, draft PR, managed-label, and terminal-comment mutations
-  without duplicating non-idempotent GitHub writes.
-- Added atomic Publication preparation for both scoped Missions and durable Plan
-  Parents, including exact applied-candidate and final-validation fencing.
-
-### Changed
-- GitHub issue and pull-request adapters now expose canonical fully paginated
-  enumeration contracts with immutable remote identities.
-
-### Fixed
-- Publication recovery remains durably indexed through restart, counts only
-  completed postcondition observations, fails closed after exhausted state CAS,
-  and rejects attempts to overwrite an existing recovery schedule.
-- Plan Parent cancellation now atomically cancels linked Publication work so a
-  cancelled owner cannot continue remote mutations.
-- Live-smoke completion fixtures now satisfy the current proof-plan contract and
-  isolate each scenario from prior configuration overrides.
-
-## [0.1.48] - 2026-07-14
-
-### Added
-- Added the Resolution Mission safety core with capability-fenced execution,
-  deterministic evaluation, audited quarantine output, and bounded scope
-  expansion.
-- Added a restart-safe Resolution Mission loop that drives recoverable blockers
-  through diagnosis, authorization, execution, reconciliation, and reevaluation
-  without turning internal retry conditions into user blockers.
-- Added indexed resumable Mission scheduling, fenced daemon claims, durable
-  process ownership, cancellation reconciliation, and terminal retention.
-- Added typed durable Plan Parent aggregates with deterministic dependency waves,
-  immutable child descriptors, serial integration checkpoints, recovery Missions,
-  and atomic final Publication linkage.
-
-### Fixed
-- Exact runner idle timeouts before the first safe file change now keep a
-  dedicated typed blocker and receive one bounded recovery retry instead of
-  being reported as normal completion without changes.
-
-## [0.1.47] - 2026-07-13
-
-### Changed
-- Scoped review handoffs now distinguish genuinely maintainer-only checks from
-  checks that an implementation agent can run before handoff.
-- Failed configured checks are preserved as explicit residual-risk warnings
-  after the agent's own validation succeeds instead of blocking publication.
-- The repo-local self-improvement workflow now resolves `codex` through `PATH`
-  rather than depending on a macOS application-bundle path.
-
-### Fixed
-- Review evidence validation now rejects maintainer-only claims that omit a
-  concrete reason or merely defer agent-verifiable work.
-
-## [0.1.46] - 2026-07-13
-
-### Changed
-- Automatic Acceptance Proof routing now treats issue text and changed paths as
-  applicability hints while the validated proof plan remains authoritative.
-- Self-improvement discovery now emits an explicit proof strategy selected from
-  observable behavior instead of relying on file-location inference.
-
-### Fixed
-- Non-visual frontend refactors no longer require browser screenshots merely
-  because they touch frontend paths.
-- Runner command dispatch and strict legacy visual gates no longer override an
-  accepted non-visual proof plan.
-
-## [0.1.45] - 2026-07-07
-
-### Added
-- Added runner support for agent-authored Acceptance Proof plans so scoped
-  implementation handoffs can carry explicit proof-plan artifacts.
-- Added proof-plan report validation coverage across scoped runs, plan-auto
-  children, status output, and handoff decisions.
-
-### Changed
-- Acceptance Proof guidance now documents the proof-plan/report contract more
-  directly in runner docs and scoped workflow prompts.
-
-## [0.1.44] - 2026-07-07
-
-### Added
-- Added typed runner blocker evidence and bounded report/evidence repair so
-  safe completion-report or review-handoff failures can be repaired without
-  weakening publication gates.
-- Added focused live-smoke and unit coverage for incomplete progress retry,
-  typed blocker repair, and daily self-improvement issue limits.
-
-### Changed
-- Scoped review reports now keep handoff output shorter while preserving the
-  evidence the runner needs for publication decisions.
-- Daily self-improvement runs now limit issue churn so repeated runs produce
-  fewer redundant follow-up issues.
-
-### Fixed
-- Scoped implementation runs can retry incomplete agent progress after an idle
-  timeout when the runner can prove safe progress in the worktree.
-
-## [0.1.43] - 2026-07-06
-
-### Added
-- Added `codex-orchestrator acceptance-proof validate --report <path>` so proof
-  agents can check machine-readable Acceptance Proof reports before handoff.
-
-### Changed
-- Adaptive Acceptance Proof prompts now include a minimal valid JSON report
-  template generated from the runner-owned report contract instead of a long
-  prose schema line.
-- Malformed adaptive proof reports now return all structural schema errors at
-  once and get one proof-report repair attempt before hard-blocking.
-
-## [0.1.42] - 2026-07-06
-
-### Changed
-- Runner terminal handoff now uses one shared outcome module for review-ready,
-  blocked, and promotion-requested finalization across scoped, recovered, parent,
-  and child paths.
-- Draft PR handoff, terminal labels, comments, Durable Run Summary evidence, and
-  child state cleanup now share focused ordering tests instead of relying on
-  repeated per-path mutation sequences.
-
-## [0.1.41] - 2026-07-03
-
-### Added
-- Added a focused `plan-auto-tree-recovery` live-smoke scenario that prepares
-  stale runner-owned parent state, recovered closed child evidence, and
-  retryable blocked child rework in the scratch repository.
-
-### Fixed
-- `codex-orchestrator run --issue` can now enter plan-auto parent recovery for
-  open `agent:plan-auto` issues that still carry `agent:running`, while keeping
-  child-labeled and unsafe cases behind the existing recovery safety checks.
-
-## [0.1.40] - 2026-07-03
-
-### Changed
-- Live smoke now defaults to a smaller `core-release` profile, exposes explicit
-  scenario profiles, and folds related commit, safety, and Acceptance Proof
-  checks into matrix scenarios.
-- Runner rework now uses a first-class `ReworkDecision` contract with explicit
-  retry, exhausted, and hard-block outcomes across scoped, recovery, and
-  issue-tree child paths.
-- Quality gates now prefer structured TDD evidence in `validation[]`, while
-  keeping legacy validation summaries as a fallback.
-- Figma MCP policy now distinguishes optional design context from required
-  design access: optional failures retry without MCP, required failures block.
-- Added focused live-smoke coverage for tree-child quality-gate rework:
-  `npm run smoke:live -- --scenario tree-child-quality-rework --cleanup`.
-
-## [0.1.39] - 2026-07-02
-
-### Fixed
-- Fixed full live smoke plan-child fixtures so issue-tree children write only
-  to their declared ownership paths.
-- Fixed browser-proof live smoke execution so packaged visual proof reads target
-  config while validating proof artifacts from the issue worktree.
-
-## [0.1.38] - 2026-07-02
-
-### Added
-- Added risk-routing review gates so runner handoffs can require plan review,
-  cleanup review, or code review evidence before publication.
-- Added risk and review-gate metadata to scoped, parent, and child orchestration
-  handoffs so downstream agents inherit the correct validation expectations.
-
-### Changed
-- Updated bundled workflow prompts to report risk classification and review-gate
-  completion consistently across PRD, issue breakdown, scoped implementation,
-  parent orchestration, and Acceptance Proof flows.
-
-## [0.1.37] - 2026-07-01
-
-### Changed
-- Publishability checks now drop runner-owned visual proof skip notes when the
-  changed files do not require the visual proof gate.
-- Local self-improvement daily phase summaries now share one helper for failure
-  classification, exit-code impact, and summary rendering.
-
-## [0.1.36] - 2026-05-21
-
-### Added
-- Added package-owned browser visual proof for web UI Acceptance Proof, with
-  scenario-driven navigation, screenshots, DOM snapshots, console/network logs,
-  and machine-readable UI Evidence reports.
-- Added `visual-proof auto` dispatch so setup can route web changes to browser
-  proof and mobile changes to device-backed proof from one package-owned
-  command.
-- Added live smoke coverage for the browser proof path, including packaged CLI
-  execution and browser runtime evidence.
-
-### Changed
-- Browser proof now prefers an explicit or locally installed Chrome, Chromium,
-  or Edge executable before falling back to Playwright-managed Chromium
-  download.
-- Live smoke daemon retries now tolerate transient GitHub failures during
-  browser-proof runs.
-
-## [0.1.35] - 2026-05-21
-
-### Added
-- Added a runner-validated UI Evidence Contract for Acceptance Proof reports,
-  covering workflow, viewport, freshness, layout, copy, and source-input
-  evidence for screenshot and UI-dump artifacts.
-- Added live smoke coverage for UI Evidence pass and blocking cases, including
-  missing UI Evidence and too-narrow desktop viewport proof.
-
-### Changed
-- Runner-owned visual proof no longer treats screenshot-only command success as
-  a pass path; proof commands must produce a valid machine-readable Acceptance
-  Proof report.
-- Updated the legacy `visual-proof` live smoke scenario to emit the same
-  machine-readable UI Evidence report required by the runner.
-
-## [0.1.34] - 2026-05-20
-
-### Added
-- Added an Adaptive Proof Agent Codex phase for scoped and issue-tree child
-  runs, with runner-provided proof report paths, artifact directories, changed
-  file context, and proof-owned repair policy.
-- Added durable Acceptance Proof attempt evidence in lifecycle events, run
-  summaries, blocked comments, review reports, and issue-tree PR handoff.
-- Added package-bundled Acceptance Proof workflow prompts and setup routing for
-  the new proof phase.
-
-### Changed
-- Parent `agent:plan-auto` child waves now block parent publication when a child
-  Acceptance Proof attempt fails, requests rework, or is blocked.
-- Proof attempts now use isolated Codex homes and preserve proof artifacts while
-  keeping publication authority runner-owned.
-
-## [0.1.33] - 2026-05-20
-
-### Added
-- Added canonical `reviewGates.acceptanceProof` policy with proof-owned path
-  classification and machine validation for high-confidence proof reports.
-- Added live smoke scenarios for canonical Acceptance Proof pass, proof rework,
-  low-confidence blocking, and proof-phase product-diff blocking.
-
-### Changed
-- Kept `reviewGates.visualProof` as a compatibility adapter while routing
-  runner prompts and proof policy through Acceptance Proof.
-- Proof-phase product-code changes now block publishability instead of being
-  silently committed as verification output.
-
-## [0.1.32] - 2026-05-19
-
-### Added
-- Added package-owned `visual-proof mobile`, `visual-proof android`, and
-  `visual-proof ios` commands for reusable UI launch proof across installed
-  repositories.
-- Mobile visual proof now supports Flutter Android, native Android, Flutter iOS,
-  and native iOS projects, with screenshots saved into runner proof artifacts.
-
-### Changed
-- Setup now defaults visual proof to
-  `codex-orchestrator visual-proof mobile --issue ${issueNumber}` instead of a
-  target-repo local proof script.
-- Android proof resolves SDK tools from environment variables, `PATH`, and
-  default macOS, Linux, and Windows SDK locations.
-- On macOS, mobile proof falls back to the iOS simulator when Android tooling or
-  devices are unavailable and the repo has an iOS target.
-
-## [0.1.30] - 2026-05-18
-
-### Added
-- Runner setup now records an explicit remote base branch so scoped and
-  issue-tree worktrees start from the same branch that PRs target.
-- Context snapshots and live smoke now prove the resolved remote base ref and
-  SHA used for autonomous runs.
-
-### Changed
-- Doctor now checks the configured remote base branch instead of trusting the
-  currently checked-out local branch, and warns about legacy base config.
-- Existing Codex branches from a different base are blocked on resume instead
-  of being reset, deleted, or rebased automatically.
-- Prompt conflict reports now give agents explicit keep, merge, and replace
-  actions so users get a concrete choice instead of a bare warning.
-- Mobile proof guidance now separates native Android, Flutter Android, and
-  native iOS validation paths so platform-specific cache/tooling recovery is not
-  applied to the wrong project type.
-
-## [0.1.29] - 2026-05-18
-
-### Added
-- Setup now records a prompt manifest and supports `--sync-prompts` modes for
-  safe prompt refreshes, local-edit preservation, replacement, and appended
-  package updates.
-- Doctor now warns when package prompt updates are available and points users to
-  the setup command that applies safe prompt updates.
-
-## [0.1.27] - 2026-05-18
-
-### Changed
-- Setup and docs now describe the package-bundled workflow prompt model
-  clearly, so installed repositories do not depend on local Codex `SKILL.md`
-  files.
-- `agent:auto` is reserved for standalone scoped issues, while issue-tree
-  children use only `agent:child` plus runner metadata and are executed by their
-  parent `agent:plan-auto` flow.
-- The default workflow prompts are full package-bundled workflows instead of
-  fallback stubs, with setup preserving project-owned prompt edits by default.
-
-## [0.1.26] - 2026-05-16
-
-### Added
-- Runner diagnostics wave: read-only `doctor`, `status --json`, phase-specific
-  Codex profiles, lifecycle events, and bounded context snapshots before Codex
-  sessions.
-- Live smoke coverage for diagnostics/profile evidence through the packaged CLI.
-
-### Changed
-- Status and handoff evidence now point to bounded runner artifacts instead of
-  requiring operators to inspect raw Codex transcripts.
-
-## [0.1.25] - 2026-05-15
-
-### Added
-- Runner-owned Loop Policy controls for daemon priority selection, bounded
-  rework, optional Fresh-Context Review, Durable Run Summaries, and
-  non-mutating Policy Suggestions.
-
-### Changed
-- Scoped and issue-tree handoff reports now include stronger runner-owned
-  evidence before draft PR publication, while keeping GitHub publication,
-  labels, comments, merges, releases, and deploys outside Agent authority.
-
-## [0.1.24] - 2026-05-14
-
-### Changed
-- Mobile UI verification guidance now gives an explicit adb/emulator preflight
-  for Android mobile app work, prefers a connected device before emulator
-  fallback, tells agents to try loading Test Android Apps when unavailable, and
-  treats missing plugin/device targets as a warning rather than a blocker.
-
-## [0.1.23] - 2026-05-14
-
-### Added
-- Automatic GitHub release notes on publish (tag + generated release notes), so each npm publish can have matching GitHub “what changed”.
-- Published `CHANGELOG.md` in the npm package, so the npm page can show a clear “what changed” history without leaving the registry.
-
-### Changed
-- Started maintaining this changelog so release-by-release improvements are easy to scan.
-
-## [0.1.22] - 2026-05-14
-
-### Added
-- Less “false red” friction across repos:
-  - If a repo doesn’t have a script like `typecheck`, the runner can treat it as a warning instead of blocking the whole run.
-  - Optional lint baseline mode to keep progress moving when the repo has existing lint debt, while still keeping touched files clean.
-- One automatic “rework” retry for common fixable blockers, so the agent can self-correct once instead of stopping immediately.
-
-### Changed
-- UI proof is Playwright-oriented and non-blocking: missing/failed visual proof is reported as a warning so it doesn’t stop unrelated progress.
-
-## [0.1.21] - 2026-05-14
-
-### Added
-- A more “productized” runner experience:
-  - Clearer, more consistent review handoff reports (what changed, what was validated, what remains risky).
-  - A stronger live-smoke harness so you can trust that publishing gates behave the same way every time.
-  - Safer path policy so the runner can reason about “what changed” consistently across OS/path quirks.
-  - Setup now avoids committing/generated runtime folders by default, reducing repo noise.
-
-## [0.1.20] - 2026-05-13
-
-### Added
-- A more reliable execution loop for both scoped issues and issue trees:
-  - Durable, structured completion reporting so the runner can validate outcomes instead of trusting free-form text.
-  - Better timeouts/logging so failures are actionable and don’t require re-running blindly.
-  - Runner-owned screenshot proof support (so visual validation can be enforced consistently).
-
-### Changed
-- Stronger “runner owns publication” boundaries (the runner stays the source of truth for commits/push/PRs and safety checks).
-
-## [0.1.19] - 2026-05-12
-
-### Changed
-- Cleaner, more predictable runs by isolating Codex session state away from the target repo and cleaning it up afterward (reduces leaked state and cross-run flakiness).
-
-## [0.1.16] - 2026-05-12
-
-### Added
-- Initial release: a reusable GitHub Issues-driven runner that can take an authorized issue, run Codex in an isolated worktree, enforce project checks/policy, and hand back a draft PR (including parent planning + dependency-aware child execution).