v1.2.0 implementation and qualification¶
Branch: codex-audit. Manifest version: 1.2.0. Released 2026-11-21 (pull request #22). Audit evidence against a996d59 remains unchanged in docs/codex/.
The implementation tracker and
project status distinguish completed work from open gates.
Implemented behavior¶
- YAML scalar coverage, stable redaction and hostile Profile A framing regressions.
- Exact lexical postings augment approximate retrieval; snapshots 04 persist postings and read 03.
- Finite/range/cache consistency validation, checked bus sequences and overflow-safe chunk sizes.
- Automatic bounded icx-managed prompt/window and structured recall frames; unsent bus stays queued.
- Daemon saves before success; SQLite WAL FULL. Failed saves return errors.
- Strict MCP envelope/argument/canonical-tool contracts. Codex capture does not promise replacement.
- Exact-SHA CI/REUSE release gates, independent measured 1k/10k scans, JSON build/fixture provenance.
- Current compatibility decisions in ADR 0006.
Evidence and pending qualification¶
The 27 frozen corpus rows produce 100% recall@1/@5 for their synthetic unique identifiers.
The actual >10k turn ingestion/budget test passes; old content still dereferences exactly.
Neither is a live model-reasoning or whole-provider-request guarantee. Account external user
input/tool schemas and the model's completion reserve when configuring icx context-budget.
Three initial parser fuzz campaigns completed 900+ seconds with normal sanitizer shutdown. After the retained counterexamples were fixed, final redactor qualification passed 859,078 executions/901 seconds and seeded 03/04 snapshot qualification passed 2,231,269/901 seconds. Logs retain failures instead of relabeling them. Initial mutation samples covered all eight crates; full generated mutation space is not exhausted. Post-snapshot targeted proofs remain separate from sample counts. Codex's portable/native strict benchmark series ran beside unrelated compilation and failed. The pre-release review diagnosed the one reproducible failure (the 10k exhaustive scan is memory-bandwidth bound and its ceiling was an unmeasured extrapolation), added an exact parallel scan and a measured ceiling, and the final quiet-host native series passed all gates in 10 of 10 runs; see performance.
The qualification ledger recorded 24 PASS, 2 FAIL and 4 WARN checks at the end of Codex's work (the 2 FAILs were the strict performance series). The pre-release review resolved the performance failure as above and, from v1.2.0 on, finding 071 by owner decision (publications replaced by a source manifest; the v1.1.0 tag, its source archives and git history still contain them, and rewriting published history is a separate owner decision); finding 074 stays open by owner decision pending counsel. The release is qualified by the exact-commit CI run of the pull request. The review's own fixes are listed under "Release-review fixes" in the CHANGELOG. Codex's last full workspace run passed 224 tests with 0 ignored, including long-horizon qualification, at 19,571/20,559 = 95.1943% line coverage. Strict Clippy passes after test-formatting correction; the earlier failure remains recorded. Dependency/advisory, license, documentation, secret-history and workflow-policy checks pass. The final nonignored working-tree source scan passes with narrowly scoped exceptions for two reviewed public source hashes and one exact synthetic audit fixture. Six positive/negative controls verify that other credentials remain detectable; the gate also runs in CI. Initial scan failures and individual adjudication remain visible. A full-directory scan additionally flags ignored generated documentation/fuzzer inputs; its separate adjudication explains scope. Initial local package checks exposed incorrect musl linkage and host GNU symbols newer than the advertised minimum. Corrected builds of all four variants now pass static/symbol checks, checksums, deterministic archives and Ubuntu 22.04/glibc 2.35 smoke tests. The initial negatives remain recorded. See local package qualification; this working-tree result does not certify a committed/tagged release. The governance review packet records owner/qualified-review decisions that tests cannot resolve. No commit, merge, tag, publication, active agent configuration or user-memory change is part of this local implementation.
Upgrade and downgrade¶
Back up a workspace before upgrading. 03 derived snapshots are read and rebuilt from preserved CAS events, then saved as 04. Older executables cannot read 04: restore the pre-upgrade backup for downgrade, rather than reinterpreting or deleting data. Unversioned Profile A fc remains Q15 scale 32,767. Existing stored secrets are not silently scrubbed or user credentials rotated.
icx context-budget --dir PATH reports the persisted policy and estimator. For example:
icx context-budget --dir PATH --context-tokens 32768 --completion-reserve 4096 --external-reserve 4096
The one-UTF8-byte-per-token allowance is conservative estimation, identified as
utf8-bytes-upper-bound-v1; it is not a downloaded model tokenizer or provider measurement.
Changing the model/window policy can change prompt prefixes and invalidate cached history.