Skip to content

Fix stack 13: correct doc drift + add API stability policy#37

Merged
ebarti merged 6 commits into
review-fixes/12-packaging-cifrom
review-fixes/13-docs-policy
Jul 2, 2026
Merged

Fix stack 13: correct doc drift + add API stability policy#37
ebarti merged 6 commits into
review-fixes/12-packaging-cifrom
review-fixes/13-docs-policy

Conversation

@ebarti

@ebarti ebarti commented Jul 2, 2026

Copy link
Copy Markdown
Owner

Stacked review-fix PR 13/N (base: review-fixes/12-packaging-ci).

REVIEW.md docs findings + §5 policy:

  • capability-matrix credential row implied Claude/Codex report missing-credential diagnostics; only Antigravity does. Clarified (Claude/Codex report available with an auth_source label; auth failures surface at run()). Streaming/tool-event rows updated to match the adapters (Codex now emits tool events; vendor.turn is Antigravity-only).
  • providers.md Antigravity policy table claimed a READ_ONLY filesystem drops the allow_all policy; the code drops it only for STRICT. Corrected, plus the READ_ONLY + explicit allowed_tools rejection.
  • README artifacts was advertised as returned but is never populated — marked reserved (empty today). Documented the new first-class model/reasoning_effort fields and FinishReason.
  • providers.md now covers system-prompt mapping across all three and Claude's READ_ONLY→plan + metadata["permission_mode"].
  • New docs/api-stability.md (public surface, 0.x policy, vendor-pin policy, deprecation), shipped in the sdist.

Docs-only + sdist include. Suite 177 passed / 12 skipped.


Re-review amendments.

  • CHANGELOG now records the whole review-fix series under Unreleased (the stack shipped breaking changes with no changelog entries despite the Keep-a-Changelog policy), including the fail-closed permission-drift behavior.
  • docs/api-stability.md now states what the capped weekly canary can and cannot catch (a release above a cap is invisible until the cap is raised) and describes the read-only mappings accurately (dict subclass — asdict/pickle/deepcopy/JSON keep working).

Round 2 (re-review follow-up)

  • Scoped the mapping-immutability claims to the top level (shallow freeze) in the CHANGELOG and docs/api-stability.md.
  • docs/providers.md no longer claims all three adapters map reasoning_effort: Claude/Codex map it to effort=, Antigravity rejects it; effort vocabularies are vendor-defined pass-throughs.
  • Documented the per-mode subtractive deny-list semantics and STRICT allow-list rejection in the capability matrix, and added a reasoning_effort row.
  • Round-2 CHANGELOG entries for the fixes in this stack (codex fail-closed statuses, cached-token accounting, smashed-case redaction, budget fail-closed, reasoning_effort wiring).

@ebarti ebarti force-pushed the review-fixes/12-packaging-ci branch from 72f6288 to fe2d61a Compare July 2, 2026 10:00
@ebarti ebarti force-pushed the review-fixes/13-docs-policy branch from 1e865fb to 9a106ca Compare July 2, 2026 10:00
@ebarti ebarti force-pushed the review-fixes/12-packaging-ci branch from fe2d61a to 0ff966a Compare July 2, 2026 10:26
@ebarti ebarti force-pushed the review-fixes/13-docs-policy branch 2 times, most recently from 22b065c to e1fcdb3 Compare July 2, 2026 10:32
@ebarti ebarti force-pushed the review-fixes/12-packaging-ci branch from 6cc291c to 516a45f Compare July 2, 2026 10:41
@ebarti ebarti force-pushed the review-fixes/13-docs-policy branch from e1fcdb3 to 162349f Compare July 2, 2026 10:41
ebarti added 6 commits July 2, 2026 17:02
…ty policy

- capability-matrix credential-diagnostics row implied Claude/Codex report
  missing-credential diagnostics; only Antigravity does. Clarified (Claude/Codex
  report available with an auth_source label and defer auth to run()). Also fixed
  the streaming/tool-event rows (Codex now emits tool events; vendor.turn is
  Antigravity-only) to match the adapters.
- providers.md Antigravity policy table claimed a READ_ONLY filesystem drops the
  allow_all policy; the code drops it only for STRICT. Corrected, and documented
  the READ_ONLY + explicit allowed_tools rejection.
- README advertised AgentResult.artifacts as a returned value; it is never
  populated. Marked reserved (always empty today). Documented the new first-class
  model/reasoning_effort fields and FinishReason.
- providers.md now documents system-prompt mapping across all three adapters and
  Claude's READ_ONLY -> plan mapping + metadata['permission_mode'].
- Add docs/api-stability.md (public surface, 0.x policy, vendor pin policy,
  deprecation) and ship it in the sdist.

Refs REVIEW.md docs findings, §5 policy.
…emantics; round-2 CHANGELOG

- Scope the mapping-immutability claims to the top level everywhere they
  are made (CHANGELOG, api-stability): the freeze is shallow and nested
  containers stay mutable, which the test suite now pins.
- providers.md no longer claims all three adapters map reasoning_effort:
  Claude/Codex map it to effort=, Antigravity rejects it with a typed
  error; effort vocabularies are vendor-defined pass-throughs.
- Document the per-mode subtractive deny-list semantics and the STRICT
  allow-list rejection in the capability matrix, and add a
  reasoning_effort row to the rejected-inputs table.
- CHANGELOG entries for the round-2 fixes: codex unknown-status
  fail-closed, codex cached-token accounting, smashed-case redaction,
  claude budget fail-closed, reasoning_effort wiring + kw_only fields.
@ebarti ebarti force-pushed the review-fixes/12-packaging-ci branch from 516a45f to 099a00e Compare July 2, 2026 15:13
@ebarti ebarti force-pushed the review-fixes/13-docs-policy branch from 162349f to 7c28b8a Compare July 2, 2026 15:13
@ebarti ebarti merged commit 7c28b8a into review-fixes/12-packaging-ci Jul 2, 2026
12 of 22 checks passed
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