Path Resolution

• <skill-root> means the directory containing this SKILL.md.
• Resolve scripts/... and references/... relative to <skill-root>, not the caller's current working directory.
• When executing local helpers, use explicit paths such as <skill-root>/scripts/....

Mode Dispatch

Determine the execution mode from $ARGUMENTS:

• If $ARGUMENTS includes a standalone token dod-recheck, --dod-recheck, or mode=dod-recheck → DoD Recheck Mode: read references/dod-recheck-mode.md and follow its instructions.
• Otherwise → Implement Mode (default; continue to # Plan Execution)

Plan Execution

Execute one selected task from an approved plan bundle.
Do not continue to other tasks unless the user explicitly asks in a new instruction.

When to Use

• You have an approved plan bundle from decompose-plan.
• decompose-plan review has produced .../plan.review.md with PASS verdict.
• The user explicitly asks to execute a specific task (for example, Task 7).
• Input includes:
  • docs/plans/<topic>/plan.md (primary)
  • docs/plans/<topic>/plan.trace.md (on-demand traceability evidence)
  • docs/plans/<topic>/plan.compose.md (on-demand reconstruction evidence)
  • docs/plans/<topic>/plan.review.md (required readiness gate)

Implement Mode Gate (Required before execution)

Before executing any task in implement mode, verify the plan review gate:

1. Run skit gate-check <review-file> <source-file>.
2. The review file must exist, contain Overall Verdict: PASS, and the Source Digest must match the current plan file.
3. If the gate check fails, stop as BLOCKED and request the user to run decompose-plan review first.

<HARD-GATE: TASK SELECTION>

Never choose a task implicitly.
Reason: implicit selection hides unvalidated dependency assumptions and removes the user's control over execution rhythm.

• Require an explicit task ID before any implementation step.
• The task ID may be specified by the user directly, or by a calling agent (e.g., ralph runtime's Task Selection Algorithm).
• If task ID is missing, ask for it and stop.
• Do not infer from dependency order, priority, first pending task, or any heuristic on your own.

Process

Step 1: Load Bundle and Select One Task

1. Read plan.md.
2. Resolve Trace Pack and Compose Pack paths from the plan header.
3. Confirm the target task ID:
   • Use only the task ID explicitly specified by the user.
   • If not explicitly specified, ask which task ID to execute and stop.
   • Never auto-select a task.
4. Validate bundle consistency:
   • Run skit bundle-validate-check <plan-file>. If it reports FAIL, stop and request plan correction.
   • plan.md exists and includes the selected task ID.
   • Referenced sidecars exist.
   • Checkpoint Summary exists and uses required keys:
     • Alignment Verdict
     • Forward Fidelity
     • Reverse Fidelity
     • Non-Goal Guard
     • Behavioral Lock Guard
     • Temporal Completeness Guard
     • Quality Gate Guard
     • Integration Coverage Guard
     • Risk Classification Guard
     • Trace Pack
     • Compose Pack
     • Updated At
   • Checkpoint Summary has Alignment Verdict: PASS.
   • Checkpoint Summary Trace Pack and Compose Pack values match header links.
   • Review report exists at .../plan.review.md (replace plan.md with plan.review.md) and has Overall Verdict: PASS. All sub-verdicts are validated programmatically by skit gate-check.
5. Check task dependencies:
   • Treat dependency status as satisfied only when the user explicitly confirms prerequisites are already satisfied.
   • If explicit user confirmation is missing, stop and ask for confirmation.
6. Read source design doc linked in header for acceptance context.
7. Read sidecars only when needed:
   • plan.trace.md for requirement/anchor ambiguity.
   • plan.compose.md for scope ambiguity.
8. Register only the selected task in TodoWrite (pending -> in_progress -> completed).

If validation fails, stop and ask for plan correction before implementation.

Step 2: Execute Selected Task

1. Execute the selected task exactly per task contract:
   • RED: create failing test. If referenced implementation symbols/files do not exist, add minimal scaffolding first so the test compiles and runs. RED is valid only when the test runner executes the test and fails (assertion/runtime); compilation/import/module errors are not valid RED.
   • If direct unit-level RED is technically difficult, apply this recovery loop in order: (1) add minimal seams/scaffolding, (2) retry unit-level RED, (3) move to nearest executable boundary test (integration/contract/e2e) while preserving fail-first.
   • Do not abandon implementation solely because testing is difficult. Continue until executable RED is established or a true blocker is reached.
   • GREEN: implement minimal code to pass RED.
   • REFACTOR: perform safe cleanup while keeping tests green.
   • DoD: treat all DoD items as AND conditions. Task completion requires every DoD item to pass.
     • When a DoD item reads Run: all commands in \## Quality Gates`, resolve the command list from the ## Quality Gatessection inplan.md` and run each command individually.
   • If the task has deps and the DoD includes boundary-level tests (integration/contract/e2e), confirm those tests exercise the actual dependency implementations — not mock substitutes. If boundary tests run against mocks only, follow Stop Conditions.
2. Verify expected outputs after each step before proceeding.
3. Mark the task as completed in TodoWrite.

If expected results are not met, stop and follow Stop Conditions.

Step 3: Report and Stop

1. Summarize what was implemented for this single task.
2. Provide verification evidence (commands, key outputs, exit codes).
3. Output a Recheck Input block for the DoD Recheck Mode:

   ## Recheck Input
   - **Task ID**: Task N
   - **Risk Tier**: [Standard | Sensitive | Critical]
   - **DoD Commands**: [DoD verification command list]
   - **Expected Outcomes**: [Expected result for each command]
   - **Quality Gates**: [Resolved commands from `## Quality Gates`, if applicable]
   - **Risk Flags**: [Any deviations recorded during implementation]

4. If the task's Risk Tier is Sensitive or Critical, or if the task's Risk Tier is Standard and its Files contain implementation files (paths not matching *test*, *spec*, *.md, docs/*, *.txt), output an Adversarial Verify Input block:

   ## Adversarial Verify Input
   - **Task ID**: Task N
   - **Risk Tier**: [Standard (impl) | Sensitive | Critical]
   - **Minimum Probes**: [1 for Standard (impl) | 2 for Sensitive | 3 for Critical]
   - **Required Scope**: [Standard (impl): most relevant 1 category | Sensitive: Category 1 + most relevant 1 category | Critical: all applicable categories]
   - **Change Areas**: [Areas from Risk Classification that this task touches]
   - **Change Rationale**: [From design doc Risk Classification]
   - **Implementation Files**: [Files created/modified in this task]
   - **DoD Evidence**: [Summary of verification results]

   Use repository-relative paths in both handoff blocks; do not emit absolute filesystem paths.
5. Explicitly stop after this task.
6. If the user wants another task, ask them to specify the next task ID.

Stop Conditions

Stop immediately and ask user guidance when:

• Plan bundle validation fails (missing sidecar, invalid summary, broken links).
• Review report (plan.review.md) is missing, malformed, or has any verdict that is not PASS.
• Target task ID is missing, ambiguous, or not explicitly provided by the user.
• Dependency satisfaction was not explicitly confirmed by the user.
• RED cannot reach executable failing state after applying the RED recovery loop.
• GREEN cannot reach pass state.
• Any DoD item fails verification (DoD is AND, not OR).
• The task has deps, the DoD includes boundary-level tests, and those tests exercise only mock substitutes instead of actual dependency implementations.
• A task anchor/requirement mapping is unclear and cannot be resolved from plan.md.
• Sidecar evidence contradicts plan.md.
• Verification fails 3 or more times for the same step.

When stopped:

• Report command, output, expected vs actual.
• Do not guess a fix.
• Do not modify the plan bundle without user approval.
• Wait for user decision.

On-Demand Read Policy

• Default to plan.md during execution.
• Read plan.trace.md only to resolve mapping/coverage ambiguity.
• Read plan.compose.md only to resolve scope ambiguity.
• After resolving, return to plan.md.
• Do not repeatedly re-read full sidecars unless a new ambiguity appears.

Verification Principle

Before claiming task completion:

1. Identify the exact verification command.
2. Run it fresh.
3. Read full output and exit code.
4. Confirm evidence matches the claim.
5. State claim with evidence.

Avoid unevidenced completion claims.

Adapting to Plan Deviations

Situation	Action
Test command in task is invalid	Apply minimal command correction, record deviation
Minor implementation adjustment needed	Make minimal adjustment, record deviation
Task is materially larger than planned scope	STOP; request plan reslicing
Missing prerequisite not captured by task	STOP; report dependency gap
Dependency completion is uncertain	Ask user for explicit confirmation and STOP
Test passes before intended change	STOP; plan may be stale
Sidecar and plan.md conflict	STOP; request plan bundle correction
User asks to continue immediately to another task	STOP current turn and request explicit next task ID
User asks to run execute-plan without task ID	Ask for explicit task ID and STOP