Skip to content

Integrate OCR 1.10.0 for toolkit 0.8.2 - #134

Merged
xeonvs merged 12 commits into
mainfrom
codex/v0.8.2-ocr-1.10.0
Aug 25, 2026
Merged

Integrate OCR 1.10.0 for toolkit 0.8.2#134
xeonvs merged 12 commits into
mainfrom
codex/v0.8.2-ocr-1.10.0

Conversation

@xeonvs

@xeonvs xeonvs commented Aug 25, 2026

Copy link
Copy Markdown
Owner

Current state

Implementation is complete, locally qualified, pushed, and green across all 13 exact-head hosted checks. The user explicitly waived local LLM execution because this environment has no provider access. This is recorded as a non-claim rather than substituting Codex credentials or treating deterministic evidence as a production-model review.

Release identity

  • toolkit target: 0.8.2
  • OCR target: exact 1.10.0
  • OCR 1.9.10: inherited predecessor for toolkit 0.8.0/0.8.1, not an intermediate installation
  • exact release-candidate head: 11b58547d9cba9523cc646c51f934c0b904dfe23
  • exact release-candidate tree: 56b9286585862a071f90c5df889c0240dbb91a46
  • Linux amd64 OCR SHA-256: f8f99ea071bed77dbcaa15fdd2083287bb8ae408d5928b3943ebe0788d191b6b
  • Darwin arm64 OCR SHA-256: c8f51b17c2be193ca178ecce6b5bcc1e38a5614629fbe81c6e1c95af5ede12e4

Added and changed

  • Qualified OCR 1.10.0 path-aware comments, semantic grouping/filtering, bounded multi-round review, budgets, and result consumers.
  • Added OCR_REVIEW_EFFORT=low|medium|high; exact toolkit default is medium (up to two rounds), while caller --effort remains the per-run override.
  • Updated the inherited OpenAI completion-cap contract from OCR 1.9.10's observed 58888 to OCR 1.10.0's 16384; toolkit override remains unset by default and explicit 4096 remains supported for gateways that reserve spending against the requested cap.
  • Rejected caller-owned OCR --output, --output=..., -o, and attached short forms so toolkit-owned result descriptor, DLP, cleanup, receipt, and posting boundaries remain authoritative.
  • Made failed compatibility qualification retain a closed schema-versioned issue/artifact while the qualification job remains red and promotion remains blocked.
  • Completed BL-017 with a no-new-layer result: OCR owns provider/review telemetry; toolkit owns deterministic context/DLP/receipt/posting/approval state.
  • Documented OCR 1.9.10 and 1.10.0 separately, made deployment/migration steps agent-readable, and bolded required configuration-table inputs according to their scoped Required contract.
  • Updated the PATH-effective local Darwin arm64 OCR atomically to checksum-verified 1.10.0 without changing user config, credentials, or HOME.

Trust boundaries

  • OCR group labels, paths, membership, and round diagnostics are untrusted private result data.
  • They do not enter findings, severity, fingerprints, lifecycle commands, receipt v5, toolkit telemetry, GitLab text, or automatic approval.
  • Safe MR title/description/discussion/adapter context continues through normalization, DLP, admission, store/MCP, receipt, and approval without false auto-approval blocking; rejected, mutated, required-degraded, or partial state remains fail-closed.
  • GitLab acquisition remains provider-specific; policy, DLP, store, result, receipt, and approval layers remain provider-neutral.

Completed gates

  • Compatibility workflow failure retention and hosted Linux qualification
  • OCR 1.10.0 manifest/runtime/example promotion
  • Medium effort, output ownership, group/DLP/approval contracts
  • Documentation, required-variable emphasis, backlog reconciliation, and categorized changelog
  • Checksum-verified local OCR 1.10.0 no-LLM qualification
  • Holistic requirement/privacy/architecture/data-flow/telemetry self-review
  • Local quality: 1,256 tests + 310 subtests, 86.14% combined branch coverage; risk groups 84%/82%/85%/87%
  • Ruff, MyPy, Bandit, pinned Gitleaks, lock/manifest, Towncrier, diff, and plan lifecycle gates
  • Exact release-candidate hosted PR checks: all 13 successful
  • Local LLM qualification explicitly waived by the owner; no production-provider behavior claimed
  • Protected merge, TestPyPI development verification, and stable 0.8.2 release reconciliation

Release resume point

Reconcile this plan-only waiver commit through exact-head hosted checks, ready and merge the protected feature PR, verify the TestPyPI development publication, and continue the protected release/v0.8.2 flow. Do not add a model-qualification claim during release reconciliation.

@xeonvs
xeonvs marked this pull request as ready for review August 25, 2026 08:56
@xeonvs
xeonvs merged commit cdc4673 into main Aug 25, 2026
13 checks passed
@xeonvs
xeonvs deleted the codex/v0.8.2-ocr-1.10.0 branch August 25, 2026 08:56
@xeonvs xeonvs mentioned this pull request Aug 25, 2026
xeonvs added a commit that referenced this pull request Aug 25, 2026
## Stable release candidate

- stable version: `0.8.2`
- next development line: `0.8.3`
- exact reviewed head: `8198879399953fb3eca6ee9aebb5f172b6b360eb`
- exact tree: `ba935a1f0a7f0e6e6c899120ecdc08e3822097c4`
- protected base / feature squash merge:
`cdc46735ba4338906d2049ff29789c1c92295d7c`
- deterministic source epoch: `1787648187`
- tracked release issues: #135, #136, #137

## Included delta

- Promotes exact checksum-qualified OCR 1.10.0; OCR 1.9.10 remains the
0.8.0/0.8.1 predecessor and is not an intermediate deployment.
- Adds `OCR_REVIEW_EFFORT=low|medium|high` with exact default `medium`,
while explicit OCR `--effort` retains per-run precedence.
- Keeps semantic group/path/round diagnostics private and outside
findings, severity, fingerprints, lifecycle commands, GitLab text,
receipt v5, toolkit telemetry, and automatic approval.
- Rejects caller OCR `--output/-o` forms so `ocr-ci review --result`
retains descriptor, DLP, cleanup, receipt, and posting ownership.
- Retains bounded compatibility status/issue/artifact evidence after
failed qualification while preserving the red job and blocked promotion.
- Completes BL-017 with `no-new-layer`; OCR owns provider/review
telemetry and toolkit owns deterministic lifecycle signals.
- Keeps required configuration inputs visually bold according to their
scoped `Required` contract.

## Development artifact evidence

Feature PR #134 was squash-merged as
`cdc46735ba4338906d2049ff29789c1c92295d7c`. TestPyPI run 32829250700
published and read back `0.8.2.dev70` with verified provenance and exact
wheel/sdist installs:

- wheel SHA-256:
`8ac4fe8a1b04e2472e1f31b818b97d06f35425597a6d59f51e9f0c1d6e06a9d6`
- sdist SHA-256:
`fe04831bc1bb6db566ebb1131f787562150be0e7cdb5b96932fcb519aaf2e3a7`

## Release-candidate validation

- Feature head passed 1,256 tests + 310 subtests at 86.14% combined
branch coverage and risk groups 84/82/85/87, plus Ruff, MyPy, Bandit,
lock/manifest, Towncrier, pinned Gitleaks, diff, and plan gates.
- Exact feature/reconciliation head passed all 13 protected PR checks
across Linux/macOS Python 3.12-3.14, Build artifacts, Security, CodeQL,
and Dependency Review.
- Release-only changes passed 69 focused
release/config/example/authorization tests, Ruff, JSON metadata, plan
lifecycle, lock/OCR manifest, pinned Gitleaks, and diff/invariant
checks.
- The owner waived local LLM execution because this environment has no
configured provider access. The release does not claim a production
model/provider result; checksum-verified real-OCR deterministic wire and
result-consumer evidence remains the qualification boundary.
- The stable workflow is the sole owner of reproducible release builds,
Twine, registry publication/readback, provenance/attestations, and clean
Python 3.12-3.14 installs; those boundaries are not duplicated locally.

This is the final repository mutation for 0.8.2. After exact-head checks
pass, protected squash merge starts the unchanged stable publication,
provenance, registry readback, supported-Python install, immutable
Release, receipt, issue, and milestone reconciliation workflow.
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant