tasks skill: require an integration matrix for boundary-touching tasks

[calvin-archastro]

Review on ArchCode

Problem

Plans pass review with one canonical e2e that proves a single path through a boundary. Example: a relay accept socket routed correctly only on the pod that owned it. The single-VM test shared the registry, the post-deploy soak gated nothing, and the bug was found in production. The tasks skill gave no way to state which entry-point/variation cells the e2e leaves unproven, or to make the task that proves them gate rollout.

What changed

Source: tasks/SKILL.md. Plugin copies regenerated with tools/plugin-gen (npm run build), not hand-edited.

1. Drafting step 3: integration matrix rule. For tasks touching a boundary (process, replica, service, network, queue, DB, external API, version, OS), verification carries Integration matrix: (entry points x placement/concurrency/ordering/failure/version skew), Fidelity: (real vs fake, what fakes share, production code at the boundary), and Not proven here: (uncovered cell -> proving task, which must be a depends_on of rollout/enablement or the epic-closing node). One acceptance bullet per matrix cell that matters. One generic worked example (webhook consumer, 2 workers).
2. Optional-fields paragraph: a human/post_deploy node holding deferred proof must be a depends_on of the rollout/enablement or epic-closing node.
3. Gap-check subagent prompt: asks to list every entry point, uncovered cells, shared fixture state, and whether the proving task gates rollout.
4. Resume section: the split-off human task gates rollout/epic close.

Companion: firstlanding PR adds lints integration-matrix-missing and ungated-deferred-proof plus schema docs.

Scope

Skill prose only (plus generated plugin copies for 8 harnesses). No CLI or code changes.

Risk

Low. Longer plans for boundary-touching tasks; the skill text is guidance, and nothing parses it.

User impact

Agents authoring plans add matrix/fidelity/not-proven entries and gating edges for boundary-touching tasks.

Testing

No executable e2e exists for skill prose. Checks run in tools/plugin-gen: npm run typecheck, npm test (115 pass, including the skill description length test), npm run check (140 generated files match the catalog), the same as the Harness plugins CI job.

🤖 Generated with Claude Code