Historical version
This page records design, operation, or validation facts from an earlier release. It is not current V7.17.3 guidance. Start from the documentation hub for current instructions.
V7.12.0 Release Notes¶
V7.12.0 is a candidate for reliability, validation evidence, and day-to-day diagnostics improvements. It advances the compatibility anchor to Codex CLI 0.156.0 and retains the eleven-release stable window. This record covers a local integrated candidate and local installation readback; commit, CI, artifacts, public publication, and effectiveness in already-open tasks remain separate gates.
Candidate changes and executed evidence¶
- F01 adds phase-aware SessionEnd permission fallback. Fifty targeted regressions and 16 concurrent stress repetitions passed, including named-root overlap rejection, authenticated recovery receipts, tampering rejection, and keyless-diagnostic isolation.
- F02 completed three-scenario evaluation. The candidate shortcut is not delivered and the original LIGHT rule is restored. The simple tool median changed from 6 to 5 (-16.7%, below the suggested 20% target), duration did not improve, and resume baseline passed 5/5 while candidate passed 4/5. These results do not attribute quality loss or failure to the rule and do not support stable cost, general-efficiency, or causal claims.
- F03 retains complete-history validation for every append and reuses only same-execution identity reads. Twenty samples each at 100, 1,000, and 10,000 scale were executed; this is not a claim of improved asymptotic complexity.
- F04 adds explicit
status --quickwhile defaultstatuskeeps the complete check. Fifty targeted tests passed. In 20 paired, same-account concurrent native samples from historical payload948c, full status had a 2924.6905 ms median and 3521.541 ms p95; quick status had a 110.248 ms median and 115.587 ms p95, a local 96.23% reduction. The current final payload has not been remeasured. Quick status does not provide complete validation or authorization, and this result does not generalize to other accounts, hosts, or a public release. - F05 adds positive/negative controlled-evolution governance and phase-routing regression coverage. Fifty-two cases passed. The earlier
0.155.1cohort of 16 independent host tasks had 13 PASS and three FAILED results, which remain retained. With the fixed profile, SHA-256-bound unchanged bytes for 13 original cases, and three targeted follow-up tasks, combined coverage records 16 observed actual independent tasks, 16 PASS and none missing asHOST_FINAL_REPORT. It does not claim a single-cohort pass rate or an independent router trace, and only covers this profile. - The Windows native Hook launcher is fixed: with Codex
environment.shell=WindowsPowerShell, the old two-quoted-path command parsed asUnexpectedToken; the static Hook now uses a pinnedcmd.exe /D /V:OFF /S /Cwrapper, PreToolUse has explicit argv, unsafe shell characters fail closed, and POSIX remains unchanged. Managed installation completed; a fresh0.156.0parent/child task emitted all five real event types and sealed the current head under the conditions recorded in the validation report. - The installer updates
hooks.jsonand preserves existing trusted hashes inhooks.state; changed definitions no longer match their previous trusted hashes and appear modified. After upgrading, use the normal Codex/hooksUI to review and trust those ten Hooks without a bypass. A passing formal full status does not establish that trust step or an actual Hook lifecycle. - Fix the omitted account-level sealing entry: install
seal_worker.pyas a managed sibling of the Hook, detect absence/content drift, and reject a missing entry before process creation. Installed append/seal, missing/tamper detection and removal regressions pass. - This host's
0.156.0elevated sandbox fails while pinning a directory. Acceptance processes use the documentedunelevatedfallback with a read-only policy and leave global configuration unchanged; this is not a claim that elevated setup was repaired. - F06 binds complete-validator capability conclusions to discovered and executed unittest identifiers and results. Missing, all-skipped, failed, errored, subtest-failed, unexpected-success, timed-out, or corrupted evidence never projects to
PASS. Nineteen capability groups passed reprojection after the complete run, followed by nine focused evidence tests.
Evidence boundary and remaining items¶
- Local complete validation ran 617 package tests: 616 passed and one platform-conditional test skipped; all 231 runtime tests passed. These results apply to the uncommitted integrated candidate based on
6641ed24626925b3445a88014621ed530f8c2ee8, not to a future commit's CI or artifact provenance. - Current account readback proves installed, enabled and HOST_COMPATIBLE. Source, marketplace and cache have the same 245-file digest
4d16bf233697500fd2f4087f356037dbe61aa93af61670b289b44cd0b1a0abc7, and the backup exists. This evidence remains separate from complete CI and publication. - In F02 ordinary samples, baseline had three PASS and two timeout
UNKNOWNresults; candidate had five PASS results. Successful tool counts were baseline[8,13,11]and candidate[7,13,10,7,7]. Original resume samples were 0/5 on both sides because synthetic external-context reads were denied; those failures remain. The completed parent-workspace read-only comparison passed 5/5 baseline and 4/5 candidate tasks. - Two bilingual ZIP candidate builds were reproducible, but documentation changes require rebuilding them; the old ZIP is not a final public asset. The final commit, remote matrix, CI, tag, GitHub Release, and actual native Hook lifecycle each require separate confirmation.
Upgrades support V7.11.2 and the older versions declared by the manifest. Commit, tag, public release, installed state, and effectiveness require separate readback.
- New
0.156.0host acceptance retains 13 original passing independent tasks and the host execution error that stopped that cohort. A targeted follow-up covers the interrupted case and two unstarted cases. Verified raw metadata, projections, report hashes and actual task IDs produce 13+3 combined coverage of the unchanged 16-case profile, all passing with no missing cases. Execution uses temporaryunelevatedplus read-only policy. Evidence remainsHOST_FINAL_REPORT, not one-shot success or an independent router trace.