plan-founder-review（计划创始人审查）

Founder Review

Inputs

• $request: Optional plan name, direct plan path, or mode override such as auth-system, .ulpi/plans/foo.md, --quick, or foo --full.

Goal

Catch plan defects before agents execute them: phantom file paths, markdown/JSON drift, hidden prerequisites, bad task boundaries, unsafe rewrites, missing tests, and overconfident execution graphs.

Step 0: Resolve the target plan and review mode

Locate the plan to review:

• if $request names a plan, resolve .ulpi/plans/<name>.md
• if $request provides a direct path, use it
• if no target is provided, select the most recently modified .md plan under .ulpi/plans/

Then:

• read the full markdown plan
• read the companion JSON file if present
• extract title, mode, task count, task IDs, file paths, and dependency summary
• determine whether review mode is FULL or QUICK

Load references/review-rubric.md for:

• mode selection rules
• section requirements per mode
• verdict criteria

Stop early if:

• no plan exists
• the markdown plan cannot be read
• the review target is too small for formal founder review and should just be executed directly

Success criteria: The target plan and review mode are explicit before detailed review starts.

Step 1: Run the codebase reality check

Verify the plan against the actual repository:

• every filesToModify path exists
• every filesToCreate path has a valid parent path or prior creator task
• markdown and JSON agree on task IDs, file paths, and execution meaning
• claimed reuse actually exists locally
• the plan is not building functionality that already exists
• if the plan introduces a public SQL, API, or CLI surface, the plan pins the real signature or example rather than hand-waving it

Load references/review-checklist.md and complete Section 1 in full.

Success criteria: The plan's file paths, reuse claims, and artifact consistency are reality-checked against the repo.

Step 2: Challenge scope and strategy

Review:

• whether plan mode matches actual scope
• whether every task serves the stated goal
• whether non-goals are credible
• whether the ship cut is believable
• whether hidden prerequisites or scope creep exist

Load references/review-checklist.md and complete Section 2.

Success criteria: Scope, mode, and minimum shippable cut are defensible.

Step 3: Audit architecture, contracts, and dependencies

Review:

• architecture diagram completeness
• task-to-diagram consistency
• producer/consumer contract clarity
• dependency JSON realism
• execution-summary consistency
• ownership of shared wiring points such as routers, registries, manifests, startup hooks, or export barrels
• semantic safety for rewrites, planners, optimizers, filters, or query composition

Load references/review-checklist.md and complete Section 3.

If review mode is QUICK, skip directly to Step 6 after this step.

Success criteria: The plan's architecture and dependency story is coherent and executable.

Step 4: Audit risk, recovery, and degraded-mode safety

For FULL review only, challenge:

• failure-mode coverage
• mitigation quality
• rollback or containment strategy
• degraded-mode semantics
• malformed-output or upstream-failure handling

Load references/review-checklist.md and complete Section 4.

Success criteria: Critical risks are covered by actionable mitigations rather than vague warnings.

Step 5: Audit test coverage and execution feasibility

For FULL review only, challenge:

• completeness of the test coverage map
• appropriateness of test types
• edge-case coverage in acceptance criteria
• security-sensitive path coverage
• DAG efficiency
• agent matching
• effort calibration
• P0 foundation honesty
• write-scope collisions or hidden overlap

Load references/review-checklist.md and complete Sections 5 and 6.

Success criteria: The plan is testable, executable, and not obviously over-constrained or under-specified.

Step 6: Score sections and deliver the verdict

Use references/review-rubric.md to:

• score each reviewed section as PASS, WARN, or FAIL
• classify findings as BLOCK, CONCERN, or OBSERVATION
• determine APPROVE, REVISE, or REJECT

Then render a concise founder-review report that includes:

• plan title and mode
• verdict
• section summary table
• blocking issues
• concerns
• observations
• recommended plan changes when revision is required

Do not ask the user procedural questions at the end of the review. Deliver the audit and let the main thread decide what happens next.

Success criteria: The verdict follows directly from evidence-backed findings and the rubric.

Guardrails

• Do not modify the plan file as part of this skill.
• Do not rubber-stamp a non-trivial plan.
• Do not invent findings; a clean section can legitimately pass.
• Do not add disable-model-invocation; this skill should remain callable when the user asks for plan review.
• Do not add paths:; this is a generic review workflow.
• Do not keep giant examples, quality scorecards, or gate tables inline in SKILL.md.
• Do not use AskUserQuestion as a substitute for a verdict.
• Do not forget that this skill runs best as a forked review workflow with separate reasoning budget.

When To Load References

• references/personality.md Use at session start for review persona, traits, and communication style.
• references/review-checklist.md Use for section-by-section audit checks and evidence requirements.
• references/review-rubric.md Use for mode selection, verdict rules, section scoring, and finding classification (BLOCK/CONCERN/OBSERVATION gate tables).

Output Contract

Report:

1. reviewed plan and selected mode
2. verdict
3. section status summary
4. blocking issues
5. concerns
6. observations
7. required plan changes when applicable