hepha-skill（赫法技能）

Hepha

Purpose

Run each requirement as multiple small, autonomous loops:

plan -> execute -> check -> review -> commit

Keep looping with minimal user intervention until the backlog is done or a stop condition is hit.

Activation

Activate only when the user explicitly asks for:

• hepha / autopilot / autonomous loop / unattended iteration
• continuous plan-execute-check-review-commit flow
• small-step commits until a larger requirement is completed

If the user did not explicitly request hepha, do not force this mode.

Non-Negotiable Operating Rules

1. One loop = one smallest shippable sub-task.
2. No commit before both engineering checks and browser review pass.
3. Every loop must update progress artifacts under .autopilot/.
4. If blocked, re-plan automatically; ask user only when truly necessary.
5. Prefer minimal diff and avoid unrelated files.

Required Working Artifacts

Create and maintain these files in the project's .autopilot/ directory:

• .autopilot/backlog.md - task graph and states (todo, doing, blocked, done)
• .autopilot/progress.md - per-loop execution log and evidence
• .autopilot/decision-log.md - research and technical decisions

Templates: Use the template files from templates/ in this skill directory as starting points:

• templates/backlog.md
• templates/progress.md
• templates/decision-log.md

If working files do not exist, copy from templates or create them before the first loop.

Loop Protocol

Execute the following phases in order for each loop.

1) PLAN (Enhanced)

Goal: pick exactly one ready sub-task from the backlog.

Steps:

Step 0.5 - Schema Validation (execute every PLAN):

Verify each task in backlog.md contains:

• ✅ id (format: TASK-XXX or numeric)
• ✅ title (action statement)
• ✅ state (todo|doing|blocked|done)
• ✅ depends_on (array, can be empty)
• ✅ acceptance (testable pass conditions)
• ✅ risk (low|medium|high)
• ✅ files_hint (expected files, optional)

Missing fields → complete before continuing Circular dependencies → detect and report error

Step 0 - Auto-Decomposition (if backlog.md missing or empty):

1. Analyze original requirement to identify core functional modules
2. Apply decomposition patterns (see references/decomposition-patterns.md):

- Vertical slicing: split by user value path (UI → API → Data) - Risk-first: high-risk dependencies first - Independence: each task testable and committable separately

1. Generate task graph:

- Assign unique ID to each sub-task (TASK-001, TASK-002...) - Identify dependencies (depends_on) - Assess risk level (low/medium/high) - Define acceptance criteria (acceptance)

1. Output to .autopilot/backlog.md

Step 1 - Normalize and Build Task Graph:

1. Normalize current requirement into:

- Goal - Definition of done - Constraints - Out of scope

1. Build/refresh task graph:

- Decompose Epic -> Tasks - For each task, define input/output, acceptance, dependencies, risk

1. Select one task from ready queue (all dependencies done).
2. Write loop plan into .autopilot/progress.md:

- selected task - expected files - expected checks - expected browser validation path - Update progress visualization section

2) RESEARCH (explicit trigger conditions)

Goal: make informed decisions with live evidence.

Decision Matrix - Research Required?

Research Quality Requirements:

1. Compare at least 2 options
2. Prefer official documentation and source code
3. Record: option summary → evidence links → tradeoffs → decision rationale

Record in .autopilot/decision-log.md:

• option A / B summary
• evidence links or source notes
• tradeoffs
• final decision and rationale

3) EXECUTE

Goal: implement the chosen sub-task with minimal blast radius.

Rules:

• Keep changes focused on required files only.
• Avoid speculative refactors.
• Keep functions small and reusable.
• Add concise comments only where logic is non-obvious.

4) CHECK

Goal: verify engineering quality.

Run all relevant project checks (examples):

• lint
• tests
• build/typecheck

If any check fails:

1. Capture failure details in .autopilot/progress.md.
2. Fix the root cause.
3. Re-run checks.
4. Repeat until pass or retry limit is reached.

5) REVIEW (browser and UX evidence required for UI/flow changes)

Goal: verify behavior from a user perspective, not only compile success.

For UI/interaction changes, use MCP browser tools and/or Playwright to validate:

• page load success
• key interaction path works
• expected text/element state is visible
• major regressions are absent

Attach review evidence to .autopilot/progress.md:

• interaction steps
• observed result
• screenshots/snapshots when relevant

6) COMMIT

Commit only when:

• checks passed
• review passed
• acceptance criteria for selected task are met

Commit policy:

• one loop, one commit
• conventional commit format
• message explains purpose/why, not only what

Update task status in .autopilot/backlog.md to done and append commit hash in progress log.

Re-Planning Policy

Trigger re-plan when:

• dependency changed
• repeated failures suggest wrong approach
• discovered scope mismatch

Re-plan behavior:

1. Split the current task into smaller tasks.
2. Mark blocked tasks explicitly with reason.
3. Continue from next ready task.

Stop Conditions

Stop loop and report clearly if any condition is met:

1. No ready task and unresolved blockers remain.
2. Same task fails checks/review 2 consecutive loops.
3. Required tooling is unavailable (critical checks cannot run).
4. User-defined risk boundary is exceeded.

When stopped, provide:

• current status
• blocker root cause
• proposed next actions

Completion Conditions

Consider a large requirement complete only when:

1. All backlog tasks are done.
2. Requirement-level definition of done is satisfied.
3. Relevant checks pass on final state.
4. Required review evidence is present.

Then generate a final completion summary:

• completed task list
• key decisions
• risk notes
• follow-up suggestions

Communication Style During Hepha

• Keep user updates brief and frequent.
• Do not ask for confirmation every loop.
• Ask user only for true ambiguity, policy conflicts, or missing credentials.

Suggested Starter Prompt For Users

Use this starter format to begin a run:

1. Enable hepha mode.
2. Run loop: plan -> execute -> check -> review -> commit.
3. Perform web/GitHub research before technical choices.
4. For UI flows, perform browser-based validation.
5. Continue until backlog is complete or stop condition is met.
6. Requirement/backlog: <paste requirement here>.

Additional References

• Planning details: references/planning_task-decomposition.md
• Quality gates: references/validation_quality-gates.md
• Decomposition patterns: references/decomposition-patterns.md
• Progress template: references/progress-template.md
• Working file templates: templates/backlog.md, templates/progress.md, templates/decision-log.md