Panda Suite Docs
v1.9.2GitHub ↗
Panda Suite / Release v1.1.1

v1.1.1 — Contract verification and recommendation discipline

This patch refines six existing skills. The catalog remains at 19 skills; no runtime, dependency, specialist, or mandatory workflow was added.

Changes

  • Lead: distinguish exploration from artifact creation, check existing coverage before recommending additions, and reassess challenged recommendations.
  • Execute: derive expected outcomes from requirements, probe uncovered acceptance rules, and distinguish green tests from complete acceptance evidence.
  • Review: check changed behavior against the contract and whether existing assertions could detect a plausible violation, while retaining no-finding controls.
  • Brainstorm: consider existing coverage and retaining the current approach; reconsider the premise before replacing a challenged suggestion.
  • Plan: distinguish omitted, null, empty, and invalid inputs when their outcomes differ, and check whether the proposed boundary can deliver each error.
  • Skill Lab: inspect existing corrections before adding more instructions; validate green-suite defect cases and correct controls without calling fixture checks model evaluations.

Evidence and evaluation scope

Historical evaluation found implementations that passed supplied tests while violating a stated requirement, an initial review that missed the same defect, and an inconsistent proposed signature/error contract. Those original artifacts remain unchanged in the audit. Current Execute and Plan already contained related corrections; this patch clarifies the verification mechanism rather than presenting those historical defects as fresh failures of current instructions.

The motivating conversation also showed premature drafting and unsupported replacement recommendations. These observations justify narrow routing and recommendation refinements; they do not establish causation or comparative gains.

Eleven new scenarios cover Lead overlap, exploration, challenged recommendations, explicit creation, Execute acceptance coverage, Review defect detection and a correct control, Brainstorm coverage and reassessment, Plan error boundaries, and Skill Lab's handling of existing corrections. Reviewer expectations remain outside task fixtures. Existing nonactivation cases remain in the catalog.

The new label-normalization fixture deliberately has a green supplied suite but violates its duplicate-preservation requirement. Separate correct implementations provide a review control. This is synthetic case validation, not a model trial. The execution case receives only requirements, the faulty service, and supplied tests; correct reference implementations are confined to the review fixture.

Fresh baseline and revised-skill model comparisons: Not Run. Token consumption, time savings, and general behavioral improvement are unmeasured. No independent task results are simulated or inferred from structural checks.

Compatibility and recovery

Skill names, discovery descriptions, catalog size, invocation routes, and package layout are unchanged. Portable and Claude manifests advance to 1.1.1. The changes are instruction clarifications; they add no required infrastructure or data migration. Existing users can restore the prior instructions by selecting v1.1.0. An installed plugin cache is not modified by this source release.

The upstream Panda Debugger license uncertainty documented in source notices remains unresolved and unchanged.

Verification

Local checks on 2026-11-09:

Check Actual result
.venv/bin/python scripts/validate-skills.py 19 skills, references, manifests, and scenario coverage passed
.venv/bin/python -m unittest discover -s tests -p 'test_tooling.py' -v All 27 tooling tests passed
python3 scripts/vendor-upstreams.py Upstream hashes match pinned snapshots
.venv/bin/python scripts/build-docs.py and python3 scripts/check-docs.py 31 pages built; 1,327 internal links checked without broken targets or fragments
python3 tests/check_codex_discovery.py Codex CLI 0.159.2 read all 19 plugin skills and discovered all 19 standalone copies
claude plugin validate for plugin and marketplace manifests Both passed
Temporary all-case preparation check 58 scenarios × 3 arms = 174 workspaces; only task prompts and fixture files copied, no reviewer catalog
Contract-fixture checks Three supplied tests passed against the deliberately faulty service; requirement-derived duplicate check rejected both faulty variants and accepted both correct controls
git diff --check Passed

These checks establish structural correctness, discovery, and fixture behavior. They do not establish a fresh behavioral improvement from the revised skills.