relay-plan（接力计划）

Relay Plan

Build a scoring rubric from task Acceptance Criteria (AC), then generate a dispatch prompt that drives autonomous iteration until convergence.

Process

1. Read the task

Read the normalized task source (try in order, use first that succeeds):

• Relay-ready handoff brief from relay-intake: ~/.relay/requests/<repo-slug>/<request-id>/relay-ready/<leaf-id>.md
• Local task file: backlog/tasks/{PREFIX}-{N} - {Title}.md
• GitHub: gh issue view <N>
• User-provided description

If relay-intake already produced a handoff brief, treat that file as the source of truth instead of re-reading the raw request.

1.5 Read historical signal

Before designing the rubric, read relay reliability history:

node ${CLAUDE_SKILL_DIR}/../relay-dispatch/scripts/reliability-report.js --repo . --json

Use historical_signal.stuck_factors, divergence_hotspots, and avg_rounds to tighten factor wording and calibration. The signal does not gate dispatch or alter state. Empty/failure cases render as no historical data available; details: references/signals.md.

1.6 Read probe quality signals

Before designing the rubric, read repo-local quality signals:

node ${CLAUDE_SKILL_DIR}/scripts/probe-executor-env.js . --project-only --json

Use probe_signal.test_infra, lint_format, type_check, ci, and scripts to inform rubric design, prerequisites, and Available Tools. The signal exposes data; it does not pick. No-signal/failure cases render as no quality infra detected; details: references/signals.md.

2. Build the rubric

Use the guided interview (references/rubric-design-guide.md) to derive factors from AC, or convert directly:

rubric:
  prerequisites:
    - command: "npm test"
      target: "exit 0"
  factors:
    - name: API returns cursor-paginated response
      tier: contract
      type: automated
      command: "curl -s localhost:3000/api/items?limit=10 | jq '.next_cursor'"
      target: "non-null cursor string"
      weight: required
    - name: Pagination robustness
      tier: quality
      type: evaluated
      criteria: "Last page works; cursor is opaque and stable under writes."
      scoring_guide: { low: "happy path only", mid: "last page handled", high: "opaque stable cursor" }
      target: ">= 8/10"
      weight: required

Tier classification, type, weight, setup/baseline, criteria, and scoring_guide: see references/rubric-design-guide.md.

Domain references

Consult references/rubric-*.md for frontend, backend, security, refactoring, documentation, and design thinking. Design factors from AC, informed by references.

Trust-model audit factor (auth-boundary tasks)

If the task crosses an auth boundary (trust root, anchor, invariant, validate, forge, bypass, gate-check, auth-boundary, or validateTransition* / validateManifest* / evaluateReviewGate), follow references/rubric-trust-model.md. Each question becomes a named factor. Record answers under ### Trust-model audit in the PR body before dispatch.

3. Validate the rubric

Quick gate before dispatch:

• Prerequisites hold repo-wide hygiene only; factors stay substantive (tier test)
• Contract/Quality tier minimums met for task size (S/M/L/XL)
• ≥ 1 automated check across prerequisites + factors
• Every evaluated factor has scoring_guide with low/mid/high anchors
• Criteria are specific and reference discoverable artifacts; targets are concrete

Full checklist, factor counts, grading, and risk signals: references/rubric-validation.md. Grade D = revise; Grade C = warn and state the tradeoff.

3.4 Simplify the rubric

Before persisting the draft rubric, apply the 6 heuristics in references/rubric-simplification.md.

Apply to all task sizes: rewrite HOW into observable WHAT, merge overlaps, remove unsupported defensive clauses, and verify weights.

3.45 Optional isolated planner draft

For standalone opt-in planner isolation, generate draft artifacts without changing the default /relay flow:

node ${CLAUDE_SKILL_DIR}/scripts/plan-runner.js \
  --issue 42 --planner codex --repo . --out-dir /tmp/relay-plan-42 --json

This writes rubric.yaml, dispatch-prompt.md, and planner-notes.md. The orchestrator reviews and may edit before dispatch.

Persisting Phase 1 deviations as anchor

Use this when operator planning rejects or narrows the issue body AC. Persist the operator-authored Phase 1 decision before dispatch so fresh-context review uses the same scope anchor.

RUN_ID=issue-42-$(date -u +%Y%m%d%H%M%S000)-deadbeef
node ${CLAUDE_SKILL_DIR}/scripts/persist-done-criteria.js \
  --repo . --run-id "$RUN_ID" --file /tmp/done-criteria-42.md --json

node ${CLAUDE_SKILL_DIR}/../relay-dispatch/scripts/dispatch.js . \
  --run-id "$RUN_ID" -b issue-42 --prompt-file /tmp/dispatch-42.md \
  --rubric-file /tmp/rubric-42.yaml \
  --done-criteria-file ~/.relay/runs/<repo-slug>/$RUN_ID/done-criteria.md

The helper writes ~/.relay/runs/<repo-slug>/<run-id>/done-criteria.md with source planner_decision. Dispatch picks it up via --done-criteria-file when the same run id is used. Canonical filename is always done-criteria.md; ad-hoc file paths remain source file.

3.5 Review the rubric (L/XL tasks)

S/M skips. L does one stress-test round. XL adds calibration simulation. Skip re-dispatches with iteration history and all-automated rubrics. Protocol: references/rubric-stress-test.md.

4. Generate dispatch prompt

Take the base template (../relay/references/prompt-template.md) and append Setup, Scoring Rubric, Iteration Protocol, and Score Log sections.

Full iteration-protocol text + Score Log format: references/iteration-protocol.md.

5. Dispatch

Write the rubric YAML to a temp file alongside the dispatch prompt. Every relay dispatch must pass --rubric-file so the rubric is persisted at anchor.rubric_path for review and merge gates.

node ${CLAUDE_SKILL_DIR}/../relay-dispatch/scripts/dispatch.js . \
  -b issue-42 --prompt-file /tmp/dispatch-42.md --rubric-file /tmp/rubric-42.yaml --timeout 3600

When to use

All tasks dispatched via relay. S/M = lightweight rubric (1-5 factors), skip stress-test. L/XL = detailed rubric with stress-test and calibration. Re-dispatches automatically prepend previous Score Log + reviewer feedback to the prompt (see relay-dispatch docs). Full rubric guide: references/rubric-design-guide.md.