Tdpilot brain release

Skill dreamrec/TDPilot/skills/tdpilot-brain-release

Use when reviewing TDPilot brain, MCP surface, schema, prompt, resource, skill, agent, hook, or plugin changes before release.From its SKILL.md

Install

npx -y skills add dreamrec/TDPilot --skill tdpilot-brain-release

Assembled from the repository path, not quoted from the project. Check it against their README if it does not work.

One thing to look at

7 stars7 stars. Stars are a popularity signal and not a quality one, but at this level it is likely that nobody has read this closely except its author, and you would be relying on your own review.

SKILL.md

5.6 KB, as published. Nobody here has run it

TDPilot Brain Release Auditor

Use this skill before any vNext brain release, plugin package update, or public surface handoff.

Core Rule

Release only evidence, not confidence. Every public surface change needs a matching test, manifest/schema update, package artifact path, and local-first constraint check.

Required Checks

Tool count and manifest:
- mcp/manifest.json matches registered @mcp.tool and @mcp.resource decorators.
- src/td_mcp/release_gates.py matches the new minimum tool count.
- tests/fixtures/tool_schemas.json is regenerated for intentional tool schema changes.
Brain behavior:
- td_brain_plan is read-only.
- td_brain_execute accepts only a valid BrainPlan.
- td_transaction_apply enforces max op count, dry run, snapshot, rollback, and validation options.
- Blocked plans never mutate.
Correctness:
- Unit tests cover schema validation, concept graph validity, profile routing, rollback paths, and severity classification.
- Fake TD tests cover success, apply failure rollback, validation failure rollback, stale state, missing operators, name conflicts, snapshot failure, and nested undo refusal.
- Golden evals pass with uv run python scripts/eval_brain_golden.py.
- Brain smoke dry-run passes with uv run python scripts/brain_live_smoke.py --dry-run.
- Live TD smoke planning runs with uv run python scripts/brain_live_smoke.py --live when TouchDesigner is available.
- Brain atlas coverage passes with uv run python scripts/audit_brain_atlas.py.
Packaging:
- Codex skills and custom agents point users toward the brain loop.
- Claude plugin agents and hooks are deterministic and local.
- Plugin surface audit passes with uv run python scripts/audit_plugin_surface.py.
- Complete plugin release gates include scripts/check_release_gates.py with --plugin-surface-report and --require-plugin-surface.
- When TouchDesigner is available, complete release gates also include the captured live-smoke report with --brain-live-smoke-report and --require-live-smoke.
- The open MCP core has no hosted LLM dependency.

Minimum Local Test Set

Run these before handoff:

uv run pytest tests/models/test_brain_models.py tests/brain/test_planner.py tests/brain/test_transaction.py tests/test_brain_tools.py -q
uv run pytest tests/test_tools_contract.py tests/test_tools_schema_snapshot.py tests/test_resource_fallbacks.py -q
uv run python scripts/eval_brain_golden.py
uv run python scripts/brain_live_smoke.py --dry-run
uv run python scripts/audit_brain_atlas.py
uv run python scripts/audit_brain_skills.py
uv run python scripts/audit_plugin_surface.py
uv run python scripts/smoke_mcp_registry.py

For a plugin or public-surface release, capture the brain and plugin reports. --require-complete demands SIX reports — brain eval, brain smoke (dry), brain LIVE smoke, operator availability, plugin surface, and param-semantics risk — and the live-smoke report must show a real mutation (mutated_td=true plus a passing transactional_generated_code_smoke). Two of those (brain LIVE smoke and operator availability) can ONLY be produced against a running TouchDesigner, so the complete gate CANNOT pass without a live TD.

Local reports (no TD needed):

uv run python scripts/eval_brain_golden.py > /tmp/tdpilot_brain_eval.json
uv run python scripts/brain_live_smoke.py --dry-run > /tmp/tdpilot_brain_smoke.json
uv run python scripts/audit_plugin_surface.py --root . > /tmp/tdpilot_plugin_surface.json
uv run python scripts/audit_param_semantics_risks.py > /tmp/tdpilot_param_semantics_risk.json

Live reports (TouchDesigner MUST be running — these mutate the live project and clean up):

uv run python scripts/brain_live_smoke.py --live --transactional-generated-code > /tmp/tdpilot_brain_live_smoke.json
uv run python scripts/sample_operator_availability.py --out /tmp/tdpilot_operator_availability.json --store-root .

Complete gate (needs all six reports above):

uv run python scripts/check_release_gates.py \
  --brain-eval-report /tmp/tdpilot_brain_eval.json \
  --brain-smoke-report /tmp/tdpilot_brain_smoke.json \
  --brain-live-smoke-report /tmp/tdpilot_brain_live_smoke.json \
  --operator-availability-report /tmp/tdpilot_operator_availability.json \
  --plugin-surface-report /tmp/tdpilot_plugin_surface.json \
  --param-semantics-risk-report /tmp/tdpilot_param_semantics_risk.json \
  --require-complete

If TouchDesigner is not running you CANNOT run the complete gate: run the local reports plus the non-complete checks above, and state explicitly that live smoke, operator availability, and therefore strict release-completeness were NOT verified.

Pressure Scenarios

Pressure: Tool counts pass but plugin skills, agents, hooks, or MCP config changed. Run uv run python scripts/audit_brain_skills.py and uv run python scripts/audit_plugin_surface.py before release.
Pressure: A live TD smoke run is unavailable. Do not call release complete; report dry-run evidence and the exact reason live smoke was skipped.

Common Mistakes

Updating a tool schema without regenerating tests/fixtures/tool_schemas.json.
Shipping root skills that differ from .agents/skills or plugins/tdpilot/skills.
Shipping plugin MCP config with a personal path instead of a plugin-root placeholder.
Treating green unit tests as proof that Codex/Claude packaging still includes the new brain artifacts.