Changelog — English current-release summary¶
Chinese full history: CHANGELOG.md
V7.17.3¶
Post-release update: the stable version was published at 2026-10-29 13:37:22 UTC. All six anonymous downloads, four release-validation combinations and signed provenance passed verification. Downloads and bilingual reports reflect actual publication; account installation, Desktop restart and activation remain separate facts.
- Diagnose component updates, confirmed management-contract changes and unavailable probes separately; enhanced runtime evidence remains unverified instead of demanding reinstall for every update.
- Add native-turn workspace provenance and allowance cursor v2 for the first unbudgeted project selection, preserving old sources and root ownership. Historical recovery excludes automatic continuations and foreign-project input.
- Share source read/write locks to fix Windows atomic-replacement races; advance the internal stable-component window to 0.164.0 with frozen official artifact/source evidence.
- Share verified account-root alias resolution across parent provenance, child identity and late callbacks, preserving inner-link, file-identity, project and budget-ownership checks.
- Complete the listed native reading, approval recovery, cancellation and account-installation acceptance, with an independent source supplement passed. Record current-version full CI, public publication and artifact verification separately in the release reports.
V7.17.2¶
- Fix repeated-confirmation dead ends with host-observed provenance recovery, bounded scanning and migration diagnostics.
- Preserve project, file identity, authorization, budget and host trust boundaries with targeted regressions.
7.17.1 - 2026-10-17¶
- Advance the Codex Desktop internal-component window to official stable 0.163.0, retain the ten preceding stable releases, and remove 0.157.0 without creating a standalone CLI support track.
- Freeze the official npm artifact, annotated-tag commit, and Hook/apply_patch source digests. Version 0.163.0 has no contract drift from 0.162.x and continues to reuse
result-v162. - Preserve the six Skill-quality workflows, task allowances, and gate-recovery behavior published in V7.17.0. Package, account Plugin, actual Desktop, CI, tag, assets, provenance, and public downloads remain separate acceptance layers.
7.17.0 - 2026-10-16¶
Post-release status update: the public Release was published at 2026-10-16 20:23:25 UTC. All four release-validation combinations, provenance, six anonymous asset downloads and official payload/installation checks passed. Bilingual downloads and reports reflect the actual release; historical failures, review protocol limits and uncovered host activation remain explicit.
- Add six workflows for artifact comparison, independent readers, frontend design, grouped routing evaluation, browser acceptance and MCP contracts, reusing existing tools and templates.
- Replace chat-wide budget ownership with registered task scopes; repairs, continuations, renamed tasks, subagents and late events retain their original ledger.
- Add role-bound Worker/Explorer allowances distinct from Reviewer grants, preserving historical charges and bounded storage for authorized attempts and settlement. Registered plans can share one explicit quote.
- Recover an explicit human correction to a prior allowance refusal within the original request, retaining declined decisions, limits and historical spend. A bounded native rescan can recover corrections missed by an older reader.
- Fix external task-artifact path rejection, prose-entropy checks on ordinary filenames and broad blocking after missing post-write receipts. Retain actual permissions, credentials, links, identity and budget protections.
- Shorten the read-only journal lock while retaining full replay and atomic writer validation, fixing contention that blocked parallel tool reads.
- Give the complete package test suite a separate bounded 90-minute limit while other validation commands retain 60 minutes, accommodating slower Windows runners without removing tests or timeout/failure checks.
- Complete the planned document, frontend and MCP synthetic executions and eighty frozen routing observations. Preserve failures and UNKNOWNs; trigger descriptions remain unchanged. See candidate reports for final independent review, matrix and release status.
7.16.2 - 2026-10-15¶
Post-release status update (2026-10-15 UTC): publication, all six assets and anonymous downloads were read back. All four release-validation jobs passed, and local native approval, review dispatch and settlement were verified. Refresh bilingual downloads and reports while retaining the UNKNOWN host outcome and missing source-write audit receipts.
- Add scoped review-allowance requests and native approval recovery through append-only events in the existing root, retaining base capacity, historical charges and active calls.
- Reserve each approved review from its own work/role/profile-bound allocation; other Workers, Explorers and reviews cannot consume it.
- Separate stable work identity from short-lived dispatch permits. In-scope packet refresh preserves approval; scope expansion, closed work and unapproved profiles remain restricted.
- Accept direct and structured user replies, distinguish human input from automatic context, and prevent duplicate grants after replay, concurrency or a post-append interruption. Both locale renderings share one request.
- Register required review and repair work ahead of spending. Distinguish units, attempts, active calls, native-source and installed-reader failures; recover existing grants and unknown outcomes first.
- Add installed-reader consistency checks, bilingual operating guidance and isolated subprocess regressions. Package checks, independent review, installation and actual Desktop effectiveness remain separate evidence.
7.16.1 - 2026-10-12¶
Post-release status update (2026-10-13 Asia/Shanghai): publication, six attested-workflow assets and anonymous downloads were read back; managed installation and a native large-result call were verified. Refresh bilingual downloads/reports while retaining original-project business-replay and other limits.
- Fix the child-context Hook rejecting parent tool inputs/results above 2 MiB before identifying the caller, or injecting an incorrect incomplete-result instruction.
- Validate complete large JSON envelopes with bounded streaming and top-level identity only; reject duplicates, invalid escapes/UTF-8, trailing data and conflicting aliases without logging or spooling bodies.
- Preserve the existing 2 MiB child limit and G6/V5 ownership and permission checks; bound transport, elapsed parsing time, depth, nodes and key memory.
- Add parent/child, byte-boundary, image-result, late-identity and malformed-payload regressions. Publication, installation and field effectiveness remain separate checks.
7.16.0 - 2026-10-12¶
Post-release documentation correction (2026-10-12): synchronize current downloads, release/validation status, and landing-page registration counts. Download sections now follow recorded public Release metadata. Published tags and signed artifacts retain their original content.
- Fix native-write rejection of in-repository absolute paths and inert metadata, preserving scope and identity protection with actionable preparation parameters.
- Inventory actual registered Hooks and align execution/review controllers with GPT-6 defaults. Missing review or index material permits scoped continuation without manufacturing verification.
- Isolate child-ledger failures from ordinary parent reads and reconcile long or late native receipts without erasing accounting, refunding unknown calls, or redispatching.
- Support authorized capacity templates, same-root continuations, and verified Agent ID mappings. Reserve each new turn separately; old receipts cannot settle a new turn.
- Preserve frozen legacy replay, logical-readonly limitations, and the Desktop-only product boundary. Artifact, installation, runtime, and publication evidence remain distinct.
7.15.3 - 2026-10-10¶
- Advance the Codex Desktop internal component window to official stable 0.162.1, retain the ten preceding stable releases, and remove 0.156.1 and 0.156.0.
- Freeze official npm, annotated-tag commits, and source-contract evidence; add
result-v162while retaining historical result profiles still consumed by 0.161.0 and 0.157.x. - Preserve V7.15.2's published GPT-6 script-first defaults and bounded flexibility; package, account Plugin, fresh Desktop task, CI, tag, assets, provenance, publication, and anonymous downloads remain separate acceptance layers.
7.15.2 - 2026-10-09¶
Target version: 7.15.2. These are source-candidate feature changes; installation, native acceptance, commit and public-release states are recorded separately in the version validation report.
- Default new tasks to script-controlled GPT-6 Luna/Sol/Astra low/medium/high profiles for Reviewers, Workers and Explorers. Ordinary or unknown work uses Sol/medium without statistical qualification, gain cards or historical samples as dispatch prerequisites.
- Move scope measurement, default selection, planning weights, shared budgets, atomic reservations, required holds, retries and result checks into scripts. Models retain business judgment and bounded proposals; scripts recalculate exact parameters before execution. Scope or one failure does not establish difficulty.
- Give active gates consistent default/degraded continuation paths. Missing evidence remains UNKNOWN without disabling gates or inventing PASS. Confirmed authorization, safety, identity, accounting and budget violations restrict their affected action; supervised-write boundaries remain.
- Separate permit preparation, atomic reservation, native creation receipts, genuine terminal outcomes and no-start proof. Preserve in-flight, failed, cancelled and unknown accounting; reentry and recovery cannot omit charges, double-charge or reset limits.
- Preserve old roots, unresolved records and late-event ownership during same-chat handoff. Replay V3/V4/V5 and previous research/reader contracts unchanged. Retain earlier context, output-attribution, payload-integrity and recovery fixes for compatibility, without promoting development results into default-model qualification.
-
Align bilingual core rules, Skill entries, user guides and model policy. Distinguish planning weights from actual fees, permission to continue from verified success, and source candidates from installed or published versions. Support Codex Desktop only; internal scripts do not create a standalone CLI product track.
-
Define computable metrics and whole-collection deadlines; enforce required-work profiles, hold transfer, no-start retries and split adjustment accounting. Stop derives minimal observations from existing journals without zero-filling unknown quality or actual charges.
- Fix legacy Context Hook rejection of new G6 children, terminal accounting when final messages precede task_complete, and current/historical release-report consumers. Isolate legacy synthetic test accounts without relaxing runtime gates.
7.14.6 - 2026-10-06¶
- Advance the Codex Desktop internal component window to official stable 0.160.1, retain the ten preceding stable releases, and remove 0.155.0.
- Freeze official npm, annotated-tag commit, and source-contract evidence. Version 0.160.1 continues to reuse
result-v158, while 0.155.1 retainsresult-v155. - Preserve Desktop-only product scope, the 3600-second bounded validation, and the master/tag ancestry gate; package, account Plugin, fresh Desktop process, CI, tag, assets, provenance, publication, and anonymous downloads remain separate acceptance layers.
7.14.5 - 2026-10-04¶
- Advance the Codex Desktop internal component window to official stable 0.160.0, retain the ten preceding stable releases, and remove 0.154.0.
- Freeze official npm, annotated-tag commit, and source-contract evidence. Version 0.160.0 continues to reuse
result-v158, while the exiting release's uniquely usedresult-v154is removed atomically. - Preserve Desktop-only product scope, the 3600-second bounded validation, and the master/tag ancestry gate; package, account Plugin, fresh Desktop process, CI, tag, assets, provenance, publication, and anonymous downloads remain separate acceptance layers.
7.14.4 - 2026-10-02¶
- Preserve the Codex Desktop 0.159.2 component window and all V7.14.3 compatibility evidence.
- Raise the bounded validation subprocess timeout from 1800 to 3600 seconds after three immutable-tag Windows Python 3.11 runs timed out without assertion failures.
- Preserve fail-closed timeout evidence, Desktop-only scope, the master/tag ancestry gate, and all independent publication gates.
7.14.3 - 2026-10-01¶
- Advance the Codex Desktop internal component window to official stable 0.159.2, include stable 0.159.1, retain nine earlier stable releases, and remove 0.153.4/0.153.3.
- Freeze official npm, tag-commit, and source-contract evidence for both patch releases; their Hook/apply_patch contracts match 0.159.0 and reuse
result-v158. - Preserve Desktop-only product scope and keep package, account Plugin, fresh Desktop process, CI, tag, assets, provenance, publication, and anonymous downloads as separate evidence gates.
7.14.2 - 2026-09-30¶
- Advance the Codex Desktop internal component window to official stable 0.159.0, retain the ten preceding stable releases, and remove 0.153.2.
- Freeze official npm, tag-commit, and source-contract evidence; 0.159.0 matches 0.158.0 and therefore reuses
result-v158. - Fix a Windows concurrent selection-lock race that could falsely classify an in-root recovery guard as unmanaged, while retaining link/reparse rejection.
- Preserve Desktop-only product scope and keep package, account Plugin, fresh Desktop process, CI, tag, assets, provenance, publication, and anonymous downloads as separate evidence gates.
7.14.1 - 2026-09-28¶
- Advance the frozen component window used by Codex Desktop to official stable 0.158.0, retain the ten preceding stable releases, and remove 0.153.1.
- Add
result-v158, referenced only by 0.158.0, to freeze apply_patch handler drift without rewriting the actual Desktop runtime contract. - Preserve the V7.14.0 authoritative Desktop V2 review-context protocol and validate the stable internal component, account Plugin, and fresh Desktop process as separate evidence layers.
7.14.0 - 2026-09-28¶
- Add an opt-in bounded Desktop V2 review context. Opaque transport is audited separately; the runtime verifies and delivers approved material.
- Introduce Budget 5, Review State 10 and Result 7 while preserving old V4 plaintext checks and historical journal semantics.
- Add session-first protection, exact child association, a fixed reader, actual-output receipts and controller-owned result attribution.
- Permit evaluation mode only; production remains unavailable without tool-coverage and scenario qualification evidence. Preserve GPT-5.6 compatibility and frozen V3 defaults.
- Package tests, independent review, native Desktop acceptance and publication are recorded separately in this release's validation records.
7.13.6 - 2026-09-27¶
- Register the two exact Desktop V1 messaging and resume names so reentry in a V4-bound task reaches the existing PreToolUse budget gate.
- Preserve neutral behavior for unbound tasks, existing creation handling, message-digest checks, and historical ledgers without wildcard matching.
- A real Desktop V1 run verified reentry denial and one GPT-6 Luna / Low reservation, creation receipt, completion, and charge. Evidence covers the tested local channel only; model quality qualification and GPT-6 defaults remain unchanged.
7.13.5 - 2026-09-27¶
- Advance the frozen component window used by Codex Desktop to official stable 0.157.1, retaining the ten preceding stable releases and removing 0.153.0 from the active window.
- Freeze official npm, tag-commit, Windows binary, and source-contract evidence. Upstream provides no verifiable feature highlights, so no specific fix is inferred.
- Preserve Desktop-only scope, the V7.13.4 permission and delegation fixes, frozen reviewer policy, and fail-closed native-admission boundaries.
7.13.4 - 2026-09-25¶
- Fix Windows Desktop installation permissions and delegation entry points; existing stable-component windows remain internal contract-regression references, not standalone CLI support.
- Freeze official npm, tag-commit, and source-contract evidence while preserving Desktop-only, frozen V3, recovery, and path protections.
- Fix Windows public-payload ACL inheritance, Desktop delegation-name normalization, and Pre/Post matcher registration. Positive native admission remains blocked by opaque encrypted message transport; body validation is not weakened.
7.13.3 - 2026-09-25¶
- Fix missed budget bindings for Windows Desktop path aliases while preserving historical fingerprints and ledger events.
- Serialize registration per task and compare equivalent lifecycle proofs under the ledger lock; unavailable identity, conflicts, or disappearing bindings reject the operation.
- Align external-state and managed-root protection while preserving support for valid child targets that do not exist yet.
- Keep frozen V3 defaults, GPT-5.6 compatibility, and GPT-6 qualification gates unchanged; full package, installation, and native admission are accepted separately.
7.13.2 - 2026-09-25¶
- Advance the frozen compatibility window to Codex CLI 0.156.1, retain the ten preceding stable releases from 0.156.0 through 0.152.1, and remove 0.152.0 from the active window.
- Freeze official npm, tag-commit, and source-contract evidence; the 0.156.1 Hook/apply_patch hashes continue to use
result-v156. - Preserve the Codex Desktop-only product boundary, frozen V3 default, and V7.13.1 recovery protections. GPT-6 Sol/Luna visibility does not activate production defaults.
7.13.1 - 2026-09-24¶
- Verify ordinary managed-file recovery backups before deletion; missing or corrupt backups preserve the current file and pending transaction.
- Retain existing ERROR classification and fixed exception types for non-assertion subtest exceptions, without expanding the report schema or recording private bodies.
- Use a 120-second process deadline for installer integration fixtures that perform multiple host probes and file transactions; functional assertions, production command limits, and Hook deadlines remain unchanged.
- Continue GPT-6 compatibility support with frozen V3 as default. Native dispatch, model qualification, and default activation still require separate acceptance. Preserve V7.13.0 as a candidate without completed release acceptance; full patch validation, artifacts, and publication are read back separately.
7.13.0 - 2026-09-24¶
- Add six models and eighteen Low/Medium/High evaluation combinations. V4 production candidates are the nine GPT-6 Luna, Sol and Astra combinations, gated by scenario qualification and gain evidence. Frozen V3 remains the default and existing tasks retain their original policy.
- Add scenario-bound qualification, gain and cost cards to compare higher reasoning effort with a different model. Protect later review and repair budgets, and separately constrain cumulative Astra attempts and inflight concurrency.
- Add the V4 ledger, V9/V6 review, version-isolated observations, frozen paired experiment suites and bounded Desktop task-name/lifecycle identity linkage. Unknown receipts never imply completion or refunds.
- Target Codex Desktop only. Management-component compatibility and real native acceptance remain separate. Preserve historical readers and recovery formats without a standalone CLI support track.
- Normalize staged Windows launchers to CRLF so different Git checkout settings cannot change archive contents or checksums. Source files and frozen ledgers retain their original bytes.
- This version delivers compatibility support: frozen V3 remains the default, and new models are available only for explicitly scoped evaluation. Code, cross-platform CI, candidate archives and installation are verified separately. Native budget admission and model qualification are not accepted; GPT-6 production defaults and enforced-budget effectiveness are not enabled or claimed.
7.12.0 - 2026-09-23¶
- The candidate adds phase-aware SessionEnd permission fallback, execution-result-driven validation evidence, positive/negative governance-Skill and phase-routing regression cases, and explicit
status --quickdiagnostics; defaultstatusretains the complete check. - The complete validation report no longer asserts capability success from filenames or fixed strings: missing required tests, all skips, failures, errors, subtest failures, unexpected successes, timeouts, or corrupted reports remain non-
PASS. - Update the package, Plugin, compatibility registry, bilingual documentation, and candidate release material to V7.12.0 while advancing the Codex CLI anchor to 0.156.0. The fixed eleven-release window now spans 0.156.0 through 0.152.0; 0.151.0 leaves the active window.
- Local package/runtime validation, SessionEnd recovery regressions and quick-status measurements are recorded in the release reports. The experimental LIGHT shortcut is withheld after the comparison; original LIGHT instructions remain. The 20 quick-status pairs belong to historical payload
948c; the current payload has not been remeasured. - Fixed-profile host-routing combined coverage retains 13 original evidence records and adds three targeted follow-ups:
HOST_FINAL_REPORTrecords 16 observed actual independent tasks, 16 PASS and none missing. It is neither a single-cohort pass-rate claim nor an independent router trace. The Windows PowerShell Hook launcher fix uses a pinnedcmd.exe /D /V:OFF /S /Cwrapper; managed installation exited 0 while native lifecycle results and execution conditions are recorded in the validation report. - The installer updates only
hooks.jsonand preserves trusted hashes inhooks.state; the new wrapper makes the current ten managed Hooks appear modified. After upgrading, review and trust them in the normal Codex/hooksUI without a bypass; passing full status does not mean trusted or lifecycle-passed. - Final host lifecycle, remote CI, artifact provenance and publication require separate readback.
- Install and verify the account-level
seal_worker.pyas a managed sibling; reject a missing entry before process creation. A fresh native task passes all five event types and current-head sealing, with the local elevated-setup limitation and temporary official fallback explicitly retained.
7.11.2 - 2026-09-20¶
- Advance the frozen compatibility window to official stable Codex CLI 0.155.1, retain the ten preceding stable releases from 0.155.0 through 0.151.0, and remove 0.150.1 from the active window.
- Freeze npm integrity and tarball SHA-256, the official tag commit, and Hook/apply_patch source digests. Windows CLI, isolated Plugin, and synthetic Hook contracts pass; the stable compatibility matrix validates Ubuntu independently.
- Synchronize the package version, bilingual release material, builder, verifier, installation/recovery documentation, and tag artifact links to V7.11.2. Future, prerelease, and out-of-window hosts continue to fail closed.
7.11.1 - 2026-09-18¶
- Fix the 17-to-9 score regression after adding valid evidence. reviewer-matrix-v3 / luna-evidence-v2 jointly optimizes root-group and dimension constraints with deterministic whole-solution ties.
-
Preserve old policy resources, evaluators, ledger and state replay. Initialization accepts both policies and verifies root bindings. Model tuples, weights, limits, and highest-affordable selection remain unchanged. The local account upgrade and native read-only loading passed; this candidate is not publicly released.
-
Adapt to official stable Codex CLI 0.155.0 while retaining the current-plus-ten stable window. Verify CLI, Plugin, Hook and apply_patch contracts separately; prereleases still fail closed.
7.11.0 - 2026-09-17¶
Local candidate; no public release has been performed in this work.
-
Bind native tool calls, agent creation receipts, and lifecycle callbacks. Calls without task names carry a random single-use dispatch reference; only its hash is retained. Implicit matching, incorrect references, and repeated consumption reject. Missing agent_id is not proof of root-caller identity. Direct start/complete cannot stand in for V3 native completion; unknown no-start causes do not trigger refunds.
-
Add Sol/Astra Low, Medium, and High for ten registered Reviewer combinations. Start from Luna Low, score valid evidence and review budget, and dispatch once within explicit quality constraints.
- Add frozen policies, V3 root accounting, V8/V5 review contracts, and version-isolated calibration while retaining V2/V7/V4 compatibility and preventing permit reuse, root mismatches, cost mixing, and silent downgrades.
7.10.0 - 2026-09-15¶
- Correct diagnostic aggregation, installed-version and payload readback, sample stability, and explicit project control prerequisites. Unreadable or corrupt state returns structured results and scoped recovery guidance.
- Recovery validates actual checkpoint references and identity with bounded reads, process output, deadlines, and descendant cleanup. Legacy Markdown fingerprints or sampled contents never prove current validation.
- The unified entry preserves the caller directory and avoids redundant Python startup. Native help remains available without Python. Benchmarks recompute statistics and reject mixed identity, invalid timings, and overwriting prior results.
- Incomplete-transaction recovery takes priority over control-prerequisite guidance. Diagnostics resample changing read errors and journals and reject oversized or corrupt journals. Unrelated Git text converters no longer make recovery evidence unknown; missing fixture directories stop benchmark collection before commands run.
- A complete recovery query returns exit code 0 while retaining its
STALEwarning; partial or unknown queries still return 1. The POSIX entry rejects an invalid explicitly selected interpreter and validates each automatically discovered interpreter, matching the Windows entry. -
Preserve stable Windows/Ubuntu required-check names for the protected branch. Aggregate checks depend on all four Python validations and block the build if the matrix is unsuccessful.
-
Complete the P1-P2 experience improvements: separate base and optional-enhancement availability, capability-level status and structured next actions, four daily paths, a unified source/archive entry, a read-only
resumeview, and explicit UX benchmark collection/import/comparison. - Fix Windows PowerShell 5.1 base installation confusing hidden directories with files during payload copying; use controlled recursive copying so native base Marketplace and ten-Skill registration can be read back.
- Preserve the legacy installer, runtime, and recovery contracts; queries do not repair state or grant write authority. Add the boundary cases, bilingual documentation, isolated fixtures, and regression tests to the release gates.
7.9.2 - 2026-09-14¶
- Complete release build metadata and synchronize the compatibility registry digest; fix the English release-notes heading. V7.9.1 remains an unpublished failed candidate; V7.9.2 delivers the registry fix.
7.9.1 - 2026-09-14¶
-
Fix UNREADABLE from the default capability registry command after Plugin installation: ship an authority-matched registry in the managed runtime and resolve source and installed layouts separately. Add isolated payload, explicit missing-path, and corrupt-file regression coverage.
-
Synchronize current documentation with V7.9.0: navigation, installation readback, base/enhancement boundaries, Hook event and registration counts, and current validation links. Preserve historical release evidence and regenerate English and compatibility copies.
7.9.0 - 2026-09-12¶
- Add a base-install entry point requiring neither this package's Python runtime nor an API key. The base Plugin loads the ten Skills only and starts no Hook, worker, Profile, index, or budget ledger.
- Move Hooks, account runtime, global rules, Reviewers, recoverable state, budgets, and controlled writes into the optional enhancement transaction. Isolated regressions cover base-to-enhancement and enhancement-to-base restoration while unknown files remain preserved.
- Make
doctorandstatusreport usable work, affected capability, cause, and next action by default while--jsonretains the full machine format. Host drift and a missing enhancement can appear together. - Keep simple local work serial in the main agent by default. Strict budgets enforce only with a verifiable real-host binding, ledger, and permit.
7.8.1 - 2026-09-10¶
- Fix account
inventorycomparing the wholeAGENTS.mdfile against the managed-block hash in install state, which incorrectly reportedDRIFTafter a valid installation. - Add unique managed-block extraction and verification. User text outside the markers remains independent, while managed-content changes, duplicate or missing markers, path escapes, and reparse points still fail closed.
- Retain V7.8.0 AUTO, onboarding, scoped review, installation recovery, and Codex 0.154.0 compatibility unchanged. The V7.8.0 tag and Release remain immutable history.
7.8.0 - 2026-09-10¶
- Add the exact C01-C25
capability-registry/1, per-feature AUTO/BASIC/ASSISTED/FULL selection, and external explicit preferences while keeping levels, risk, and external authority independent. - Add
onboarding/1and recoverable scan jobs with one offer, nonce/revision CAS, renewal, cancel epochs, lease fencing, cursor, and partial-coverage readback. - Add scoped-review dependency fingerprints. Unknown dynamic dependencies are
INCOMPLETE, related changes areSTALE, and release-wide audit remains separate. - Add actual Python 3.11+ preflight, per-feature doctor, read-only inventory, zero-delete state-less preview, non-Git BASIC guidance, and lazy preference migration.
- Fix Windows device-prefix, short-path, and case aliases incorrectly classifying managed files as
UNSAFE_PATH; containment now normalizes aliases while out-of-root paths and reparse points still fail closed. - Add exact contracts and discoverable tests for U02-U13, UX01-UX40, M01-M12, and T25-T28 while retaining V7.6.2 message-boundary and V7.7.0 Operation v2 regression coverage.
7.7.1 - 2026-09-10¶
- Advance the closed compatibility registry to OpenAI Codex CLI 0.154.0, freezing
0.154.0,0.153.4,0.153.3,0.153.2,0.153.1,0.153.0,0.152.1,0.152.0,0.151.0,0.150.1, and0.150.0, and removing 0.149.1 from the active window. - Pin the official 0.154.0 Git tag/commit, Hook discovery, PreToolUse/PostToolUse schema, successful
apply_patchsources, npm SRI, and SHA-256, with a 0.154.0-onlyresult-v154profile. - Fix a configured-but-disabled capability gate incorrectly blocking PostToolUse when no operation is in flight; an already-dispatched B still converges after disablement.
- Windows CLI, account Plugin 7.7.1, three-way payload digest, and a real fresh task were read back successfully; full evidence remains in the V7.7.1 release record.
7.7.0 - 2026-09-10¶
- Add repository-external Operation v2: the first
apply_patchattempt A only creates an origin and is denied; after preparation, a different B atomically claims READY, and only its real PostToolUse receipt can proceed to completion verification. - Add the
--operation-refpath tocapability-task-prepare/finish/check/cancel. The CLI creates neither host origins nor dispatch IDs, while legacy GateTask schema 1 remains isolated and compatible. - Cover Add/Delete/Update/Move, multiple targets, UTF-8/path/file budgets, concurrent claims, cancellation, expiry, policy changes, missing or invalid receipts, scoped evidence, and index maintenance.
- Bind the eleven frozen Codex versions to official PreToolUse/PostToolUse schema tags, commits, paths, and SHA-256 digests. The installer registers the enforcing Hook only when that capability is SUPPORTED.
7.6.2 - 2026-09-09¶
- Fix asynchronous UserPromptSubmit Hook registration and require explicit compatibility evidence; optional feedback no longer blocks ordinary prompts.
- Isolate legacy capability-gate state: native writes fail closed when an old policy is enabled, while unconfigured or disabled policies remain neutral; Stop/UserPrompt no longer read or mutate legacy task state.
- Focused checks cover Hook registration, legacy-gate isolation, fail-closed behavior, and state immutability. Full release validation, independent review, commit, push, publication, installation, and effective-state readback remain separate.
7.6.1 - 2026-09-09¶
- Correct Hook, migration, budget-owner and format guidance; unify English sources and deterministic fact checks.
- Add stable documentation entries while retaining old paths and anchors; archive historical outputs with manifests.
- Capture tracked release sources or explicit hashed snapshots; reject missing, tampered, conflicting or linked inputs and use captured locale overlays.
- See the V7.6.1 validation record for current evidence; earlier results do not validate the final candidate.
7.6.0 - 2026-09-09¶
-
Windows CI exposed short-path aliases being compared with canonical paths. Normalize ReadBudget through the existing link-rejecting safe_path helper, compare installed-entry tests against canonical paths, and add a real Windows short-path regression. Repository isolation and read/time budgets remain unchanged.
-
Add bounded, identity-bound external capability indexing with incremental maintenance, exact-revision updates, and recovery.
- Connect reuse discovery, multidimensional suitability and maintenance-cost decisions to frontend/backend Skills and existing review roles.
- Add project-opt-in workflow gates with host-bound starts, preparation, current finish receipts, cancellation and at most two repairs. Unenabled projects retain existing behavior.
- Restrict the cold-index exemption to one existing local file; scope expansion requires a pre-edit scan. Keep temporary decision JSON outside the repository.
- Candidate behavior, host and installed-task acceptance completed with explicit limits; net cost savings are not established. See validation and capability index.
7.5.1 - 2026-09-06¶
- Correct the concurrent acceptance timing assumption: retry a bounded two-second lock timeout after the other call finishes, and verify a single committed transaction. Add forced-contention coverage for unchanged watermarks and idempotent worker recovery. The lock bound and runtime watermark protocol retain their existing contract.
7.5.0 - 2026-09-06¶
-
Windows CI exposed mixed short-path aliases (such as RUNNER~1) and full paths in relative references. Feedback and regression enumeration now use fixed references beneath checked directories, and project identity uses the canonical name of the verified Profile path. A real Windows short-path regression passed. Post-fix focused tests and installation readback are recorded separately; the first complete package run belongs to the preceding candidate, and final full-suite evidence comes from CI for the fix commit.
-
Fix Profile/Hook repository fingerprint disagreement and unify default/explicit policy.
- Add task binding, structured validation evidence and parent-finalized feedback; Stop links reports while preserving host terminal fields.
- Check identity, policy, seals, coverage and freshness before analysis. Explicit opt-in enables incremental, cooled-down analysis and idempotent recovery.
- Freeze testable proposal hypotheses. Separate implementation validation from observed benefit, revalidate references on replay, and cover cancellation, rollback and terminal states.
- Verify each calibration sample against its own ledger; compare matching scenarios within one project/repository using independent-task weights and uncertainty.
- Generate pending regression candidates and recurrence follow-ups from confirmed causes, retaining execution_authorization=NONE.
The frozen Codex CLI 0.153.4 plus ten preceding stable-release window is retained; no broader host support is claimed. Automation defaults to disabled. Historical event/proposal hashes remain unchanged. Analysis without valid project identity now returns a blocked state.
See the validation report and audit report for delivery evidence.
7.4.6 - 2026-09-05¶
Changed¶
- Advanced the closed compatibility registry to Codex CLI 0.153.4, retained an eleven-stable-release window, and removed 0.149.0 from the active set. Pinned the official npm artifact, SRI, SHA-256, and normalized CLI/Plugin evidence.
- Updated the account installer, compatibility matrix, release validation, bilingual documentation, and site indexes. Codex CLI 0.153.4 fixes Astra visibility in the bundled model picker, makes it the bundled default when no model is selected explicitly, and limits async-question guidance to hosts where the tool is available. Plugin, Marketplace, and Hook contracts remain unchanged; automatic subagents remain limited to Luna/Terra profiles.
- Strengthened the semantic release gate to require the V7.4.5-to-V7.4.6 upgrade declaration and added a regression assertion.
Validation¶
- The native Windows CLI was updated through the same global npm channel from 0.153.3 to 0.153.4. Official-artifact and isolated 0.153.4 checks pass, as do the V7.4.6 account-level transaction, Plugin activation,
HOST_COMPATIBLE, and three-way 182-file payload readback. - Actual uninstall/rollback and a parent/child Agent lifecycle journey were not exercised. The complete eleven-version matrix, remote CI, tag, six-asset provenance, and public Release require delivery-stage readback.
7.4.5 - 2026-09-05¶
Changed¶
- Advanced the closed compatibility registry to Codex CLI 0.153.3, retained an eleven-stable-release window, and removed 0.148.0 from the active set. Pinned the official npm artifact, SRI, SHA-256, and normalized CLI/Plugin evidence.
- Updated the installer, compatibility matrix, release validation, bilingual documentation, and site indexes. The upstream Amazon Bedrock/Astra changes in 0.153.3 do not alter Plugin/Hook contracts or expand automatic subagent routing beyond Luna/Terra profiles.
Validation¶
- The native Windows CLI was updated through the same global npm channel from 0.153.2 to 0.153.3. The isolated 0.153.3 compatibility cell and V7.4.5 account-level transactional install, Plugin activation,
HOST_COMPATIBLE, and three-way payload readback pass. - Actual uninstall/rollback and a parent/child Agent lifecycle journey were not exercised. Remote CI, tag, asset provenance, and public Release require post-delivery readback.
7.4.4 - 2026-09-04¶
Added¶
- Added versioned GitHub Release titles: Chinese comes from
manifest.json.release_nameand English fromlocales/en/manifest-localization.json.release_name; invalid titles fail closed, and the workflow transports the title through a job output and environment variable. - Added documentation for Draft-only creation, existing-Release no-overwrite, bilingual assets/provenance, and the independent online metadata boundary for historical title backfill.
Validation¶
- V7.4.4 local package-only validation and logical-readonly review passed; remote CI, tag, Draft, and public Release readback remain pending. Historical title backfill for v7.3.0 through v7.4.3 is complete with individual online readback, while v7.2.0 and earlier were unchanged. The Codex CLI 0.153.2 compatibility boundary is unchanged.
7.4.3 - 2026-09-04¶
Changed¶
- Model governance now uses only the pre-dispatch approved profile, permit reference, reserved units, and outcome attribution. Host runtime model identity and reasoning effort are no longer read, inferred, stored, attested, billed, or used by release gates.
- TaskOutcomeEvent advances to V3, DelegationBudget to V2, and Reviewer results to V4. Calibration and Evolution now compare outcome value per reserved unit between approved profiles.
Fixed¶
- Legacy Event V2 and Budget V1 chains are verified against their original contracts before recursive safe projection. Legacy and current schemas use separate chains and cannot be mixed.
- Reviewer state migration no longer reserializes historical runtime model fields, and release tooling adds privacy-boundary and abstract dispatch-policy gates.
- Account reinstall no longer copies Python bytecode; Hooks, the worker, and account tools no longer mutate the managed Plugin cache; SessionEnd signed-job creation, scanning, and movement support Windows long paths; and the Windows Hook is normalized to CRLF so checkout conversion cannot change the payload digest.
Validation¶
- V7.4.3 reports local unit, bilingual, privacy, lifecycle, compatibility-chain, reproducible-build, independent-review, and Windows account-reinstall evidence separately. Remote CI, push, tag, and public Release still require explicit readback.
7.4.2 - 2026-09-04¶
Changed¶
- Advanced the closed registry anchor to Codex CLI 0.153.2 and updated the eleven-stable-release window to 0.153.2 through 0.148.0.
- Updated Plugin metadata, installer, bilingual builds, validators, documentation site, and the Windows/Ubuntu compatibility matrix while preserving V7.4.1 historical evidence.
Fixed¶
- Pinned official 0.153.1/0.153.2 artifacts and normalized CLI/Plugin digests while retaining fail-closed behavior for future, prerelease, and exited versions.
- Added regressions preventing stale untrusted runtime evidence after a model switch and preventing
unified_execfrom consuming sub-agent dispatch permits or budget.
Validation¶
- Windows isolated CLI, Plugin round-trip, synthetic Hook, and artifact checks pass for 0.153.1/0.153.2. Full-window, real-account, CI, review, and public-release evidence is reported separately in the V7.4.2 reports.
7.4.1 - 2026-09-03¶
Added¶
- Added a closed Codex compatibility registry for 0.153.0 and the ten preceding stable releases, binding official npm artifacts, capability profiles, and layered evidence states.
- Added a 22-cell Windows/Ubuntu release matrix, isolated
CODEX_HOMEPlugin preflight, and schema-3 CLI/registry/capability/payload host snapshots.
Changed¶
- Expanded Plugin mode from only 0.153.0 to eleven pinned stable releases while keeping future, prerelease, and out-of-window versions fail-closed.
- Reduced Marketplace ownership to managed fields while preserving unknown top-level data, nested
interfacedata,owner, and other Plugin entries. - Registered Hook snake_case, camelCase, and compatibility aliases consistently. Security conflicts deny dispatch, observation conflicts remain unavailable, and Stop/SubagentStop return neutral JSON.
Fixed¶
verify,status, anddoctornow detect CLI file, registry, profile, and capability drift instead of comparing version alone.- The Windows matrix can bind an explicit Codex executable so PATH resolution cannot select the global version accidentally.
- Release builds and runtime-text audits exclude repository-external
project-contextevidence so temporary official CLIs cannot enter an archive.
Validation¶
- All eleven official Windows CLIs passed CLI, isolated Plugin, and synthetic Hook cells under registry digest
1c204bd34cc355d5771376278c6251a5e133b7db09a7613b5c35d5c7bcdcbdd8. GitHub Ubuntu, real-account, and public-release states remain separate evidence.
7.4.0 - 2026-09-03¶
Added¶
- Added repository-external DelegationBudget V1, one append-only hashed root-task budget for Reviewer, Explorer, and Worker.
- Added controlled route reasons, explicit dispatch permits, idempotent reservations, trusted actual-profile top-ups, host-proven not-started release, and budget closure.
- Added role-specific scenario calibration and adjacent-tier offline replay. Only the parent coordinator can finalize samples, and proposals retain
execution_authorization=NONE.
Changed¶
- Upgraded Task Envelope to schema 3, execution state to schema 4, and review state to schema 6 while keeping V7.3 state readable.
- The Reviewer controller now owns rounds and findings only; DelegationBudget exclusively owns total charging.
- Updated the Plugin target to Codex CLI 0.153.0 and generate its required Marketplace
interface.displayName. The ten-preceding-stable compatibility window is deferred to V7.4.1.
Fixed¶
- PreToolUse fails closed on missing stable dispatch identity, unknown roles, permit mismatch, exhaustion, or chain corruption, while Hook replay remains idempotent.
- Start/Stop without a reservation correlation remains unavailable instead of being guessed; ordinary Hook model fields are never reported as trusted runtime evidence.
- Started agents are never refunded. Only explicit host proof of not starting releases a reservation.
Validation¶
- Added regressions for unified accounting, concurrency, nesting, Hook permits, single-owner Reviewer charging, privacy, tamper detection, parent finalization, and insufficient-data no-change behavior. Full package, real 0.153.0 installation, independent review, and public artifact evidence are reported separately.
7.3.0 - 2026-09-02¶
Added¶
- Added
minimum_acceptable_profileto Reviewer dispatch and an append-onlyINLINE/DELEGATEdecision gate;INLINEcreates no round and consumes no Reviewer budget. - Added Reviewer result schema v3 with task difficulty, duration, pending attribution, finding disposition, and
profile-weight-v1estimated cost, projected by the controller into a deduplicated calibration ledger; Reviewers cannot finalize attribution themselves.
Changed¶
- Reviewer declarations now produce only
declared_match/fallback_acceptable/underpowered/unverified/mismatch. Results below the minimum acceptable tier may only beincompleteand cannot be merged or closed normally. - Evolution keeps missing cost unknown, segments metrics by Reviewer, model tier, and task difficulty, and excludes unfinalized attribution from low-yield classification; default routing stays unchanged when real evidence is insufficient.
- Aligned the bilingual site's current V7.3 system architecture, domain routing, configuration, controlled-evolution, and authoritative-source documentation. Earlier versions now appear only for migration or historical evidence.
- Added prominent bilingual notices to historical detail pages during site generation and excluded those pages from default site search while keeping current and release-archive indexes searchable.
Fixed¶
- Fixed Plugin-mode launchers selecting a stale standalone runtime ahead of the installed versioned cache; Plugin mode now uses only its state-bound cache and fails closed when that cache is missing.
- Removed historical-version labels mixed into current navigation and specifications, and replaced a dead command that referenced a removed validator.
- Reconciled the previously divergent Chinese and English Codex configuration guides and added regression gates for the current-document inventory, historical isolation, and referenced public scripts.
- Fixed self-referential Chinese links on current English pages, Unicode heading anchors, and a stale table label; version changes now trigger Pages and fail closed when site entry points drift from the manifest.
Validation¶
- Added regressions for minimum tiers, INLINE redecision, legacy v2 result compatibility, calibration projection, unknown cost, and unfinalized attribution. Local installation and real-data readback are reported separately at task delivery.
7.2.0 - 2026-09-02¶
Added¶
- Added eleven real Codex host-routing acceptance scenarios, schema-2 observations, and SHA-256-bound final-report evidence.
- Added a persisted controlled-evolution end-to-end case from observation, snapshot, and assessment to an
execution_authorization=NONEproposal registry.
Changed¶
- Set Python 3.11 as the minimum and cover both 3.11 and 3.13 on Windows and Ubuntu CI.
- Evaluate controlled-evolution evidence per signal and calculate coverage over unique
task_idvalues; model escalation and negative outcomes consume only their required evidence. - Removed duplicated controlled-evolution guidance from
long-running-task-memoryandmulti-agent-independent-review, routing both tocontrolled-evolution-governance.
Fixed¶
- Complete validation now binds the Git index, tracked and untracked content digests, deletions, and link states, runs the post-snapshot even after interrupts, and refuses repository-local output.
- Host-routing acceptance rejects unknown Skills, duplicate tasks or reports, non-finite pass rates, hash/byte-count mismatches, and report-field drift while separating host final reports from internal router traces.
- The installer removes local Marketplace top-level
owner/interfacefields rejected by Codex CLI 0.152.1 while preserving unknown external metadata.
7.1.0 - 2026-09-02¶
Changed¶
- Updated the current Codex CLI release baseline to 0.152.1, retained verified 0.150.1 compatibility, and kept fail-closed behavior for other versions.
- Synchronized Manifest, Plugin, bilingual builds, release verification, attestations, documentation site, and current operating guides to 7.1.0, with a declared 7.0.0 upgrade path.
Fixed¶
- Plugin mode now transactionally installs and verifies the account-level
cp-runtime.pyandevolution.pylaunchers. When the account runtime is unreadable, both launchers resolve the exact active Plugin-cache version from installation state instead of reporting a missingcp_runtimemodule. - Retained Marketplace/Plugin command and core
plugin list --jsoncontract checks on native Windows Codex CLI 0.152.1. - Increased the complete-package command timeout from 300 to 600 seconds so slower GitHub Windows runners are not terminated before the test suite finishes.
7.0.0 - 2026-09-01¶
Added¶
- Added language-neutral
backend-engineeringfor Node.js, Go, .NET, Rust, Java, Python, and mixed-language services. - Added independent
ai-engineeringfor model integration, structured output, RAG, agents, evaluation, inference, GPU, and multimodal semantics. - Added a four-domain responsibility matrix and 45 positive/negative routing cases.
- Rebuilt the documentation root as a bilingual project landing page with project identity, language cards, capability metrics, release and source links, responsive layouts, light and dark themes, and keyboard access.
- Added repository-wide Markdown path, anchor, and same-repository URL checks, with scheduled external-link probing.
- Added separate Chinese and English documentation sites published by GitHub Actions, with navigation, search, theme switching, and versioned release evidence.
- Added fail-closed tag-version validation, reproducible bilingual artifacts, GitHub-signed provenance, and draft-only Release automation.
- Added a bilingual GitHub Release page index for V1.0.0 through V6.6.0 and documented the zero-asset policy for historical original ZIP files.
Changed¶
- Java and Python are now progressive backend specializations; the data domain is renamed to
data-middleware-infrastructureand no longer owns AI product semantics. - Manifest, AGENTS, bilingual documentation, recovery flows, release tooling, and package validation now target 7.0.0.
- The bilingual landing page, navigation, and security support matrix now target V7. The project preview no longer embeds a release number and uses a reusable 1280×640 JPEG below GitHub's 1 MB upload limit.
- Main CI now reads versioned archive and witness names from constrained release metadata instead of hard-coding an older release.
Fixed¶
- Fixed Material for MkDocs retaining stale repository release facts in browser session storage, which could leave
v6.6.1in the header. The site now repairs both the cached and visible version facts. - Upgrades remove only the four Manifest-declared legacy Skill directories: three V7 domain replacements and the previously deprecated Vue Skill. Unknown Skills and custom files remain untouched.
- Worktree link auditing ignores tracked paths already removed from the working tree, allowing consistent validation during renames.
- Fixed the documentation language-selector path difference between repository checks and the generated Pages tree.
- Fixed strict Pages builds by copying newly added sibling English documents into the English site source tree.
Validation¶
- 45 routing cases, 128 package tests, 6 runtime tests, strict bilingual audit, Markdown link audit, and strict MkDocs build pass.
- The source tree completed a native Windows Codex CLI 0.150.1
6.6.0 -> 7.0.0Plugin upgrade readback, and a fresh read-only task selected the general backend and data-infrastructure routes. Public ZIP artifacts remain independently built and attested by the tag workflow.
6.6.1 - 2026-08-31¶
Added¶
- Complete, independently installable, deterministic
zh-CNandenarchives. - Human-reviewed English counterparts for every natural-language document, all ten Skills and their References and templates, seven Reviewer definitions, examples, structured descriptions, and Python runtime messages.
- Locale binding, reproducible-build checks, full-project translation coverage, and fail-closed runtime-literal validation.
Fixed¶
- Bounded retry for transient Windows sharing failures during atomic file publication.
- Explicit validation-process waiting removes temporary-directory cleanup races while production SessionEnd remains asynchronously sealed outside the Hook budget.
- Windows batch launchers now use the UTF-8 code page and CRLF line endings. The neutral-language gate distinguishes the unshipped runtime source-string catalog from delivered prose.
Security¶
execution_authorization=NONE, dual project isolation, minimal metadata, and the Terra High automatic ceiling remain unchanged.- Reviewer TOML files contain no hard-coded model. Diagnostic observations are not promoted to actual runtime model proof.
- Both archives exclude unrelated brands, personal paths, nested ZIPs, Git metadata, caches, and localization-source overlays.
Earlier reconstructed release history remains preserved in the Chinese historical changelog and version tags rather than being presented as newly translated evidence.