v0.11.0 — observation-only compatibility evidence

Highlights

• Records agy 1.1.22 release/source/distribution and bounded help evidence as an observation without activating it; the accepted agy baseline, model matrix, and inventory binding remain at 1.1.16.
• Adds version-bound 1.1.22 account-inventory capture preparation and runner surfaces with synthetic positive/negative coverage. The single authorized account call failed, was not retried, and advanced no metadata or routing authority.
• Advances the Codex observational baseline to 0.150.1 without making Codex a worker backend or granting agy model/dispatch authority.
• Pins every actions/checkout use to the reviewed v6.0.2 commit with credentials disabled and exact-ref/full-history behavior preserved.
• Rejects partial/promisor clones before provider dispatch, avoiding a consumed worker cycle on unsupported Git object state.
• Streamlines development checks around one canonical stable-candidate gate, fixes nested updater coverage accounting, and removes scheduler-sensitive 10 ms assumptions from two test-only lease-boundary fixtures.
• Aligns plugin/source metadata to v0.11.0 and corrects the immutable v0.10.0 package-history discrepancy without rewriting the historical tag.

Verification

• PR #74 protected merge-ref test: success with head-range hygiene bound to 871d7299987ff77ba382b39211371028a51bb4a9 (run 33110439083).
• Manual exact-head test: success on 871d7299987ff77ba382b39211371028a51bb4a9 (run 33111372507).
• Squash-merged exact main: 7fe59599ef26773f9ab6537e1ecf31ec8ddc00b9; tree 2636a911dc3839c75bb29642e4e1b5862399b1d0.
• Manual exact-main test: success for a433664c9bc0ef888fb5585e80f027febd8d9913..7fe59599ef26773f9ab6537e1ecf31ec8ddc00b9 (run 33115398617).
• Canonical local offline gate: all 35 suites passed on the final candidate, including updater 325/325, packaging 390/390, doctor 257/257, dispatcher 334/334, remediation 89/89, 1.1.22 capture 45/30/58, and retained 1.1.16 activation 22/22.
• git diff --check, focused packaging, ground-truth interface observation, and expected 1.1.22-versus-1.1.16 drift-review passed their documented contracts.
• Separate fresh-context reviewers accepted the compatibility candidate and final release-scope wording; the AGENTS hierarchy audit remained A (97/100).

Known limitations

• agy 1.1.22 is not an active compatibility baseline. The authorized account inventory attempt exited nonzero with no accepted inventory; raw stdout/stderr remain owner-private and the failure cause is not inferred from public evidence.
• The release does not claim a provider backend, general model ranking, persona promotion, or reduced Codex allowance usage. Driver-side Codex usage observation remains tracked in issue #69; privacy-safe account-capture classification and CI timing/sharding remain tracked in #72 and #73.
• P2-B/P2-C telemetry and pruning, signing, Windows parity, daemon/MCP, dashboard/cloud, multiple backends, and automatic model/routing/publication behavior remain deferred or rejected as documented.

Upgrade

After checking out this exact tag, run ./install.sh to refresh the global Codex skill bundle. The installer does not modify agy or Codex configuration.

v0.10.0 — reliable lifecycle and Codex-owned assurance

Highlights

• Separates provider terminal status, schema-valid worker reports, physical worktree evidence, and Codex-owned driver disposition.
• Preserves valid reports from provider errors as unverified candidates without promoting them to success; cancellation remains fail-closed.
• Adds additive lifecycle/state compatibility, bounded resume/continue/finalize recovery, strict verification binding, and private actionable text output.
• Hardens repository identity, symlink, Git-probe, state-SHA, concurrency, deadline, and legacy-result boundaries.
• Binds provider/canonical schemas and caller-owned model/runtime selection while keeping version drift observational rather than an automatic product blocker.
• Removes descendant-held-pipe probe delays and reduces safe worktree snapshot process amplification.

Verification

• PR #62 protected check: success on exact candidate 5fa7cc4d63fca3c8f7781922a6f03ae8f4b10ff2 (run 32957450748, 50m28s).
• Squash-merged main 2ac893ef837dbcbb0ff0b3b9d62eecd164445bf6 has the identical Git tree 201e807a47325414d6b8a7026d8267fc4a6ff50b.
• Canonical offline runner: 32 suites, 3,374 checks; syntax, compile, diff, package/runtime/persona/benchmark parity, ground truth, help, and bytecode hygiene passed.
• Fresh independent Sol xhigh reviews accepted the final diff with no blocking finding; two unchanged UX/user-proxy rounds scored every critical axis at least 4/5.

Known limitation

• A fixed Gemini Pro task canary returned provider SUCCESS with an empty response and no structured_output in direct agy capture. agy-worker correctly rejects this as missing_structured_output; it is reported as an external provider/agy-path blocker, not a product pass and not bypassed by fallback parsing.

Upgrade

After checking out this exact tag, run ./install.sh to refresh the global Codex skill bundle. The installer does not modify agy or Codex configuration.

v0.9.0 — agy 1.1.16 compatibility

Highlights

• Activates the exact agy 1.1.16 version, source, distribution, and 14-model inventory binding.
• Preserves the reviewed Gemini 3.7 Flash low, medium, and high mappings with no tier remap.
• Accepts Codex 0.148.0 as an observational compatibility baseline without granting agy authority.
• Adds an independent, bounded, one-call/no-retry 1.1.16 capture evidence chain.
• Extends updater, notifier, doctor, resolver, packaging, and evidence-receipt compatibility.
• Keeps historical 1.1.12 evidence and the narrow 1.1.13 quota classifier distinct.

Verification

• Required PR check: success on exact candidate tree.
• Canonical offline runner: 31 suites, 3,194 tests.
• Independent fresh-context Sol audit: ACCEPT; P0-P3 = 0.
• Package parity, syntax, diff, RTK, bytecode hygiene, and AGENTS audit passed.

Upgrade

After checking out this exact tag, run ./install.sh to refresh the Codex skill. Existing local notifier installations intentionally report maintenance-required and need an explicit ./update-notifier.sh refresh after source update.

v0.8.0 — version-aware quota diagnostics

Highlights

• Recognizes the reviewed agy 1.1.13 structured individual-quota terminal as exit 24 (provider_quota_exhausted) and exposes only a bounded decreasing retry duration.
• Preserves same-conversation recovery without automatic retry, sleep, restart, or model changes.
• Adds explicit notifier maintenance/refresh behavior and annotated GitHub tag support.
• Records agy 1.1.16 and Codex 0.148.0 interface observations without falsely advancing the reviewed model matrix.

Verification

• Required PR test check passed on the exact head in one natural run.
• Local offline contract: 27 suites / 3,031 cases, plus syntax, compile, diff, bytecode hygiene, and RTK verification.

Fixes #59

v0.7.0

What changed

• Adds usability-first explore, task, and project workflows so Codex can delegate broad reviews, bounded features, and whole-project implementation.
• Adds same-conversation repair cycles with a five-attempt total cap and Codex-owned verified, partially_verified, or blocked assurance.
• Makes personas, exact file lists, and predeclared test commands optional instead of dispatch gates.
• Preserves useful candidate changes when verification is incomplete, while retaining hard repository, credential, permission, and publication boundaries.
• Supports Gemini 3.7 Flash low, medium, and high selections from reconciled agy inventory evidence.

GitHub Actions efficiency

• Runs the full macOS suite on pull requests and exact-SHA manual release checks, avoiding a duplicate full run after squash merge.
• Cancels stale runs for the same pull request.
• Moves the metadata-only feedback watch to Ubuntu.
• Adds scripts/ci-offline.sh as the canonical 27-suite local fallback when hosted Actions are unavailable or quota-limited. Local evidence remains local evidence; the required GitHub check must still run when service or quota returns.

Validation

• 27 offline suites, 3,007 cases.
• Exact-head pull-request CI passed before merge.
• Exact-merge-SHA manual release verification is required before publication.

v0.6.0 — Gemini 3.7 Flash support

Gemini 3.7 Flash support

• Adds reviewed direct selections for Gemini 3.7 Flash at low, medium, and high.
• Keeps minimal unsupported, preserves all tier mappings, and continues to forward exactly one resolved --model without a thinking flag.
• Reconciles the active agy 1.1.12 inventory from 11 to 14 exact slugs using a separately authorized, marker-valid account capture and official Google model-level documentation.

Reliability hardening

• Applies a child-only private umask to the explicit-account capture process while retaining strict mode-0600 cache validation.
• Serializes mutable dispatch-state snapshots with the existing state lock, eliminating a narrow replace-on-read inode race without weakening stale-SHA controls.

Verification

• 27 offline suites, 2,988 cases, 0 failures on the PR head and merged main SHA.
• Dispatcher 257/257, updater 319/319, packaging 365/365.
• Bash syntax, isolated Python compilation, committed diff hygiene, RTK 154/154, and a complete no-findings security diff scan.

Native Windows remains unblocked but untested/best-effort; POSIX/WSL-compatible environments remain the supported path.

v0.5.0 — Secure feedback reporting loop

v0.5.0 — Secure feedback reporting loop

This release adds a local-first feedback path for evidenced agy-worker defects and concrete improvement requests.

Highlights

• local bug, improvement, and private-only security feedback drafts
• exact preview digest plus a separate matching public-safety digest before any public submission
• fixed public destination and in-memory exact-body submission
• metadata-only feedback-triage.sh with 100-record, 64 KiB, one-page, and 20-second bounds
• weekly/manual read-only feedback watcher with no issue mutations or agent ingestion
• prompt-injection, spam/burst, lookalike-target, overflow, signal, process-group, and bytecode regressions

Security and privacy boundary

The periodic watcher does not request titles, bodies, comments, labels, or usernames. Only canonical issue numbers/URLs, timestamps, and bounded aggregate flags are emitted. Explicit security reports remain on GitHub's private vulnerability route. Keyword matching is an additional deny barrier, not proof that a draft is safe to publish.

Verification

• 27 offline suites: 2,964 passed / 0 failed
• reporting: 47 / 47
• feedback triage: 26 / 26
• package/CI policy: 365 / 365
• RTK: 154 / 154
• exact merge-SHA push CI and GitHub Pages: successful
• independent security and instruction audits: accepted

The release does not automatically submit, label, comment on, or close GitHub issues, and it does not dispatch agents from issue content.

codex-agy-worker v0.4.0

Highlights

• Replace post-process stream handling and the single five-minute wall timeout with a process-owning, progress-aware supervisor.
• Add explicit local start, status, wait, result, extend, cancel, resume, and restart commands while preserving synchronous run.
• Separate idle, hard, and absolute runtime limits; valid progress renews only the idle lease.
• Remove automatic fresh retry and bind continuation attempts to immutable task, selection, conversation, state, and result evidence.
• Add hash-approved pre-gate dispatch-failure recording and job.sh abort cleanup.
• Stage complete plan prompts privately while keeping slash expansion only for the fixed upstream plan driver prompt.

Safety

• Local status and cancellation do not claim provider status or verified remote cancellation.
• Unknown provider/authentication diagnostics remain unclassified without an exact version-bound reviewed signature.
• No automatic model, effort, thinking-level, retry, Git action, provider action, or Windows denylist was added.
• P2-B and P2-C remain evidence-gated and deferred.

Verification

• 26 offline suites: 2,908 passed, 0 failed.
• Dispatcher: 238/238; lifecycle/abort: 116/116; packaging: 361/361.
• Independent adversarial audit reported no P0-P3 findings.
• Exact PR-head CI and exact merge-SHA push test succeeded; the accepted and merged Git trees are identical.
• Bash syntax, external-cache Python compile, diff hygiene, RTK, instruction audit, and no-bytecode checks passed.

Coverage boundary

The lifecycle implementation is offline and synthetic-subprocess verified. A separately authorized live provider exercise is still required before claiming live plan, conversation-resume, or remote-cancellation behavior.

codex-agy-worker v0.3.3

Highlights

• Bound direct and snapshot envelope input to 1 MiB before JSON parsing.
• Bound lifecycle Git stdout incrementally and close the owned process group on overflow, timeout, or read failure.
• Disable persisted checkout credentials in CI.
• Align the Codex plugin package metadata with v0.3.3.

Safety

• No qa-gate weakening, automatic routing, provider action, or metadata authority was added.
• P2-B and P2-C remain evidence-gated and deferred.

Verification

• 26 offline suites: 2,865 passed, 0 failed.
• Exact final merge SHA test and Pages deployment succeeded.
• Local Codex skill installation matches the released bundle and the daily notifier remains loaded with unchanged source.

codex-agy-worker v0.3.2

Highlights

• Reconciles the active agy compatibility baseline to 1.1.12 after one separately authorized, no-retry JSON models capture and strict offline review of the unchanged 11-slug inventory.
• Binds Codex 0.147.0 compatibility observation to the exact stable-tag commit, avoiding alert fatigue from moving main while preserving new-release detection.
• Includes the optional macOS daily update notifier and its bounded install, status, run, and uninstall lifecycle.
• Clarifies Windows policy: there is no Windows-specific denylist; POSIX-compatible environments are best effort, native Windows remains untested, and the notifier is macOS-only.

Safety boundaries

• Failed or markerless captures cannot activate metadata or the model matrix, and capture attempts are never retried automatically.
• No routing, tier, recommendation, provider, or acceptance authority was added.
• The capture does not prove provider backend, pricing, billing, or future behavior.
• P2-B, P2-C, and decisions based on 30/60/90-day measurements remain deferred until their evidence thresholds are met.

Verification

• 26 offline suites: 2,862 passed, 0 failed.
• Exact merge-SHA GitHub Actions test and Pages workflows completed successfully.
• The local daily notifier was installed and loaded; its source binding and manual read-only compatibility check reported unchanged before release publication.