codex/gate-hardening/oracle-findings #1

Merged

pdurlej merged 4 commits from codex/gate-hardening/oracle-findings into main

2026-05-01 10:17:41 +02:00

pdurlej commented

2026-05-01 10:17:20 +02:00

Owner

No description provided.

pdurlej added 4 commits

2026-05-01 10:17:20 +02:00

fix(schema): unblock v2 allOf composition ff40edd117

Remove additionalProperties:false on metadata and spec.intent so v2
schema can extend them via allOf without rejection.

Background: JSON Schema 2020-12 nuance — additionalProperties (and
unevaluatedProperties) are scoped to the schema where they are declared,
not to composed parent schemas. With allOf [v1, v2_extensions], v1's
metadata.additionalProperties:false rejects v2's serves_areas because
v1 cannot see v2's evaluation.

Trade-off: extra/typo'd fields in metadata or intent now slip past schema.
Compensated by tools/validate_module.py + negative-control fixtures
landing in next commit, plus the runtime token budget cap.

feat(gates): strict YAML validator + negative-control fixtures 41cd9d2e69

Adds platformctl/tools/validate_module.py which:
- Parses YAML with StrictLoader (rejects duplicate keys at parse layer;
  catches the codex exec --full-auto apply_patch double-apply bug we
  observed on 2026-05-01)
- Validates against module.schema.v2.json with proper allOf reference
  resolution via referencing.Registry

Adds 3 negative-control fixtures + test_negative_controls.py:
- v2-valid-traefik.yaml         → must PASS
- v2-invalid-duplicate.yaml     → must FAIL at YAML parse layer
- v2-invalid-missing-required.yaml → must FAIL at schema layer
- test_gates_distinguish_clean_vs_bad → asserts gates produce
  different outcomes (catches the case where gates are 'fake')

Per Oracle 2026-05-01: 'If all real PRs pass forever but bad fixtures
also pass, the gates are fake.' Negative-control fixtures are the
defense against that.

All 4 tests green: pytest platformctl/tests/test_negative_controls.py

docs: apply Oracle architecture review findings (2026-05-01) c263039359

Restructures Phase 02 sequencing per state/L0/oracle-arch-5layer-review.md:

- Gate hardening PR ships FIRST (this PR) before any manifest migrations
- Max 2 review PRs open at any time (down from 5 parallel)
- Phase 02 PR sequence: gate hardening → canary → integration → batch → semantics
- Operator review note structured (what changed / what could break / what artifact I checked)
- AI review = 'objection generator', not approver
- GLM/Claude reviewer constraints: max 3 blockers, 5-bucket categorization
- Codex CLI: use --sandbox workspace-write (not --full-auto, which is deprecated)
- Health metrics (DORA-adapted, single-operator)

OPERATOR_ACTIONS.md gets a short appendix linking to full Oracle review.

chore(state): archive Oracle review artifacts b1d0738493

- state/L0/oracle-arch-5layer-review.md: GPT-5.5 Pro architecture review
  of 5-layer pipeline (Q1-Q7 + Top 3 pushbacks), 307 lines, basis for
  the prompts/02-catalog.md restructuring landed in previous commit.

- state/oracle-3-hermes-preview-questions.md: questions for Oracle #3
  (Hermes preview-app edge cases, PreviewApp schema, idle reaper logic).
  Response was retrieved separately via oracle-local browser run.

pdurlej merged commit c4848d2179 into main

2026-05-01 10:17:41 +02:00

pdurlej referenced this pull request from a commit

2026-05-01 10:17:42 +02:00

Merge pull request 'codex/gate-hardening/oracle-findings' (#1) from codex/gate-hardening/oracle-findings into main

pdurlej referenced this pull request from a commit

2026-05-01 10:35:06 +02:00

spec(review): 3+3+consolidator review pipeline pattern

pdurlej referenced this pull request

2026-05-01 10:36:20 +02:00

spec(review): 3+3+consolidator review pipeline pattern #2

pdurlej referenced this pull request from a commit