plan-runner（计划跑者）

Plan Runner

Executes a wave-grouped markdown plan file via parallel subagents. The parent agent orchestrates — reading the plan, dispatching subagents, running review gates, and committing. All code work happens in subagents.

When to use

When you have a plan file (.md) with wave-grouped [] checkbox items (produced by plan-builder) and want to execute them with parallel implementation, review gates, and checkpointing. Resumable across conversations.

Protocol

Input

• Plan file path: the markdown file containing the wave-grouped plan

Execution model

Parent agent (orchestrator):

• Reads the plan file and orchestrates execution
• Never reads source code files or writes code itself
• Launches subagents for all code work
• Invokes review/fix protocols after each wave
• Commits after each wave

Subagents (implementers):

• Receive: plan file path (to read their assigned items + structure outline), criteria from the plan header
• Read existing code in affected areas
• Implement the assigned work
• Return structured summary: {files_changed: [paths], summary: string, deviations: [string] | null, discoveries: [string] | null}
• If a subagent needs to modify a file assigned to another subagent in the same wave, it must NOT edit it. Instead return: {needs_scope_expansion: true, additional_files: [paths], justification: string}. The parent reassigns and re-dispatches.
• No file contents in return — paths and summaries only

Step 1 — Wave execution loop

1. Read the plan file. Extract PLAN_SLUG from the filename (e.g., plan-010-tanstack-query-migration.md → 010-tanstack-query-migration). On first wave, record PLAN_BASE_SHA=$(git rev-parse HEAD) and write **Base SHA**: <sha> into the plan file header (after the **Created** line) for final review diff range.
2. Find the next ## Wave N section with any [] items. If resuming mid-wave (some items [x], some []), dispatch only the remaining unchecked items.
3. If no waves with unchecked items: proceed to final review (Step 4).
4. Launch up to 3 Sonnet subagents in parallel for the wave's items (plan-builder caps waves at 3 items).

- Each subagent receives: plan file path, its assigned item ID(s), criteria from plan header. - If a wave has items marked "must land together", assign them to the same subagent.

1. Collect results from all subagents. If a subagent crashes or times out (no result returned):

- Surface immediately via AskUserQuestion with options: "Retry this item", "Skip and mark dependents blocked", "Abort plan". Recommended: "Retry this item". - If other subagents in the wave succeeded: do NOT commit the partial wave. Resolve the crashed item first.

1. If any subagent reports a blocking issue (returned a result but flagged a problem): pause, present to user via the AskUserQuestion tool with options: "Resolve and retry", "Skip and mark dependents blocked", "Override and proceed", "Abort plan". Recommended: "Resolve and retry".
2. Stage and commit the wave: git add [wave files] && git commit -m "plan(<PLAN_SLUG>): Wave N complete — [brief summary]"
3. Mark all wave items [x], append discoveries inline if any.
4. Per-wave review (Step 2).
5. Return to step 1.

Step 2 — Per-wave review

After each wave commit, invoke the code-review skill against the wave diff. In addition to its standard analysis (correctness, security, edge cases, bugs), evaluate the diff against the plan's criteria scoped to this wave's items.

• Artifact: wave diff (git diff HEAD~1..HEAD)
• Criteria: the plan's criteria from the header, scoped to this wave's items
• Scope: files changed in this wave only

Single review pass, no verifier. Findings have verdict: null.

After review completes, append a review summary as an inline note under the wave's last item: - Review: N findings (M fixed, K escalated, D dropped by pre-gate, E demoted). This captures the review outcome for resumability without a separate log section.

Flow control: Only pause for user input when the action table explicitly requires it (P0/P1 → fix-verify-loop escalation, or AskUserQuestion calls elsewhere). Clean reviews and P2/P3-only reviews do not pause — log the outcome and continue to the next wave immediately.

Step 3 — Per-wave fix-verify-loop

If the per-wave review produces P0/P1 findings, invoke the fix-verify-loop skill:

• Findings: P0/P1 findings with verdict: "confirmed" and evidence populated (set by orchestrator — per-wave review has no verifier pass)
• Artifact paths: files changed in this wave
• Criteria: the plan's criteria

Fix-verify-loop runs max 2 attempts per finding. After 2 failed rounds, escalate to user via the AskUserQuestion tool with options: "Retry with guidance", "Accept current state and defer", "Skip finding", "Abort plan". Recommended: "Retry with guidance".

After fix-verify-loop completes, commit the fixes separately: git add [fixed files] && git commit -m "plan(<PLAN_SLUG>): Wave N fixes — [summary]"

Step 3.5 — Review fixes commit (regression check)

If Step 3 produced a fixes commit, run a focused review to catch regressions introduced by fix-verify-loop. fix-verify-loop's bounded-resolver contract verifies per-finding resolution only, not regressions — that's the caller's job.

Spawn a code-review pass scoped to the fixes commit's diff (git diff HEAD~1..HEAD for the fixes commit).

If Step 3.5's fix-verify-loop also produces escalations, follow the same per-wave escalation handling as Step 3.

No fixes commit (Step 3 found nothing to fix) → skip Step 3.5.

Step 4 — Final review

After all waves complete, invoke the two-pass-review skill:

• Artifact: full diff from plan start to current HEAD (git diff $PLAN_BASE_SHA..HEAD)
• Criteria: complete criteria list from plan header
• Scope: all files changed across all waves

two-pass-review runs the reviewer agent (Pass 1) and, if any P0/P1 findings, auto-progresses to the verifier agent (Pass 2). Receive a ReviewOutput with verdicts populated.

If confirmed P0/P1 findings: invoke fix-verify-loop with those findings. Fix subagents use Opus for cross-file fixes at this stage.

Present: "Final review: N criteria checked. K findings fixed."

Step 5 — Discovery triage

After final review, review all discoveries logged during execution. For each:

Step 6 — Completion summary

After discovery triage, append a ## Completion Summary section to the plan file:

## Completion Summary
**Status**: Complete | Partial — [date]

### Criteria
| Criterion | Result |
|---|---|
| [from plan header] | PASS / FAIL / PARTIAL — [1-line evidence] |

### Filtered out by pre-gate
- [Findings dropped by fix-verify-loop's pre-gate as not-real, aggregated across all waves]
- [or "None"]

### Deferred
- [P2/P3 findings from reviews that weren't fixed]
- [Findings demoted by fix-verify-loop to P2/P3, with note about new severity]
- [Unresolved discoveries from triage]
- [or "None"]

Evaluate each criterion from the plan header against the final state. Be honest — mark FAIL or PARTIAL when warranted, not just PASS. The Deferred section collects P2/P3 findings, demoted findings (note their new severity), and unresolved discoveries into one place so they don't silently disappear. The "Filtered out by pre-gate" section captures dropped findings from fix-verify-loop across all waves so the user can see what the pre-gate rejected.

Coupling detection

Items within a wave are committed together — wave grouping from plan-builder handles most coupling. Cross-wave coupling should not exist if plan-builder did its job correctly.

If a subagent returns {needs_scope_expansion: true,...}, the parent reassigns overlapping items to a single subagent and re-dispatches.

Resumability

Resumable via [x] checks at wave granularity. On resume: find the first wave with any [] items, dispatch only those unchecked items.

If resuming means PLAN_BASE_SHA is lost, read it from the **Base SHA** line in the plan file header. Fallback: recover from the first wave's commit parent via git log --format=%H --grep="plan(<PLAN_SLUG>): Wave" --reverse | head -1, then git rev-parse <that-commit>~1.

Rules

• Parent NEVER reads source code or writes code. All code work via subagents.
• Always read the plan file fresh before each wave. It may have been modified by fix-verify-loop or externally.
• PLAN_BASE_SHA recorded before the first wave — used for final review diff range.
• Execute ONE wave per cycle. Don't batch waves — each wave needs its own commit and review gate.
• If an item is blocked or unclear, DON'T skip it. Use AskUserQuestion with options: "Clarify and proceed", "Skip this item", "Reorder plan", "Abort plan". Recommended: "Clarify and proceed".
• If an item requires a decision the plan doesn't specify, use AskUserQuestion with the enumerated options and a recommended choice. Do not proceed on assumptions.
• Discoveries go inline with the item that found them, not in a separate section.