SuperAPI AI Directory
Partner with us中文AI newsFavorites
Skillscodex-bridge

codex-bridge

Agents & models

View repository

Updated 10-16 05:01

About this skill

Codex Bridge

Package context, run Codex with a fixed sandbox, validate its response, and synthesize verified findings.

Replace <skill-dir> with this skill's literal directory. On Windows, py -3 may replace python.

Scenarios

Scenario Purpose Sandbox Default model
plan-review Review an explicit plan read-only gpt-5.6-sol
codify Implement approved work workspace-write gpt-5.6-sol
review-iteration Apply focused review; round 2 is final workspace-write gpt-5.6-terra
verification-round Verify extrapolated instances read-only gpt-5.6-sol

Sandbox is code-owned. Only model and effort accept project or CLI overrides.

Workflow

  1. Confirm explicit Codex intent and select one scenario.
  2. Create a round: python "<skill-dir>/scripts/create_bundle.py" <scenario> "<project-root>".
  3. Distill decisions, rejected approaches, unknowns, and session provenance into conversation.md.
  4. Copy required sources into files/ as a file-backed fixture.
  5. Fill every placeholder in request.md. Do not omit referenced plans, conventions, or reviewed files.
  6. Set manifest session, previous-round, and verification-purpose fields.
  7. Run python "<skill-dir>/scripts/validate_bundle.py" "<bundle>" --phase preflight; fix every failure.
  8. For workspace-write, preserve unrelated changes and enforce the authorized scope.
  9. Run python "<skill-dir>/scripts/run_bundle.py" "<bundle>". Add --output-schema only when the active Codex endpoint supports protocol-level schema validation.
  10. Run python "<skill-dir>/scripts/validate_bundle.py" "<bundle>" --phase post-response.
  11. Compare files_changed with the working tree; synthesize accepted, rejected, and uncertain points.
  12. Record pattern extraction; verify only concrete lateral candidates. Never recurse or create round 3.

Read workflow detail, bundle conventions, and the scenario prompt-notes.md before filling a bundle. Use checklist.md before execution and jsonl-guide.md when locating source conversations. Routing review cases live under evals/ and load only for route evaluation.

Governance

  • owner: lyh; review cadence: quarterly and whenever the Codex model family changes.
  • input_files: user-authorized plans, source files, project rules, and distilled conversation fixtures.
  • output contract: a retained bundle, validated response.json, and a primary-agent synthesis with real working-tree verification.
  • rollback boundary: remove this skill directory, then run just docs-sync to remove generated catalog entries.
  • trust report: reports/trust-report.md; output evidence: reports/output_quality_scorecard.md. Unrun telemetry, live Codex benchmarks, and approval metrics remain missing evidence.

Model names evolve. Verify current availability against the official Codex models page before changing models.json; do not silently fall back to an obsolete model.