v1.7.2
Panda Plan now makes requirement traceability, execution readiness, and future release evidence explicit.
- Tasks link requirements, dependencies, interfaces, acceptance checks, and expected evidence. Small changes can keep these links inline in one task.
- Verification identifies plausible incorrect implementations and the checks that reject them, including relevant failure and compatibility conditions.
- Plans report execution-ready, partially ready, or blocked scope and identify the first unblocked task.
- A conditional reference defines release, deployment, migration, and compatibility gates. Gates start with evidence pending. Scope revisions also reconcile their gates and traceability.
- The saved-plan handoff explicitly requests review and an execution mode in the user-facing response when applicable.
Three scenarios, synthetic faulty/correct controls, a blind-preparation regression, and frozen instruction/trial records accompany the refinement. The usage guide, routing documentation, English/Thai release references, and GitHub Pages are updated. See the refinement audit and trial results.
Verification
Fresh local checks for this release cover the Suite validator and upstream hashes, tooling tests, reviewer controls, saved execution artifacts, Codex plugin/standalone discovery, and documentation generation and links. The site build produced 50 pages with 25 skill guides; all 3,168 internal links passed target and fragment checks. Diff whitespace checks passed.
The 32 tooling tests pass. All 25 skills are validated and discovered in both Codex modes. The 14 recorded hashes and final instruction snapshot match. The saved campaign and scope-execution suites each pass six tests, and the saved resume probe passes 30 checks including its 15-test suite. The reviewer probe rejects the intentionally faulty implementation with the expected assertion and recovery failures.
Evidence limits
Recorded model trials have mixed results. Both original and candidate handled core correctness well in paired cases. Early candidates exposed readiness and handoff omissions, and the final matched handoff pair passed in both arms. Conciseness and ambiguous-policy handling remain inconsistent. The trial report preserves these observations and the conditions under which they were obtained.
Fresh model comparisons, fresh subagent execution, Python 3.9, and live-provider checks were not run for this release. General behavioral superiority, token savings, and deployment safety are unmeasured. Synthetic provider gates describe the evaluation fixture.
Compatibility and recovery
The catalog remains at 25 skills. Default and explicit plan paths, Markdown task checkboxes, chat-only output, native/subagent execution choices, and existing Execute handoffs remain supported. There are no runtime migrations or new required runtime dependencies.
Update or reinstall from tag v1.7.2. To recover a previous installation, reinstall from v1.7.1. GitHub Pages publication is checked through the Documentation workflow and the live homepage, Plan guide, and release page.