Execute-Task Skill

Role: Unified, domain-agnostic workflow skill for executing tasks during phase-5-execute. Handles all profiles: implementation, module_testing, and verification. Loaded in-context as a Skill: by the single plan-marshall:phase-5-execute envelope, once per task — this is leaf-legal in-context skill loading, NOT a per-task Task: subagent dispatch.

Key Pattern: The phase-5-execute envelope loads this unified skill in-context (a Skill: load within the single envelope, not a subagent dispatch). The skill reads the task's profile and dispatches internally to the appropriate profile workflow. Domain-specific knowledge comes from task.skills (loaded in-context alongside this skill).

Base Contract: This skill follows the execute-task skill contract defined in execute-task-skills.md for input/output contracts, error handling, and script notations.

Foundational Practices

Skill: plan-marshall:persona-plan-marshall-agent

Enforcement

Prohibited actions:

• Never target file paths outside the active git worktree. The authoritative source for the worktree root is manage-status get-worktree-path --plan-id {plan_id} (resolved internally from the Input Contract below); every Edit/Write/Read tool call during task execution MUST resolve against the returned path.
• Never run git as a cd {worktree_path} && git ... compound. All git commands during task execution MUST use the git -C {resolved_worktree_path} <subcommand> form, where {resolved_worktree_path} is the value returned by manage-status get-worktree-path --plan-id {plan_id}.
• Never combine Bash commands with &&, ;, &, or newlines in a single Bash tool call. Each Bash tool call MUST contain exactly ONE command. The compound form trips the host platform's permission UI and produces silent swallowing of intermediate exit codes — both are load-bearing failure modes during task execution.
• Never run polling loops (for/while/until loops, $() substitution, subshells, heredocs with # lines) inside a Bash tool call. Poll conditions belong in a background-completion watch, or are eliminated by running commands synchronously with an explicit timeout. Polling loops trip the host platform's security heuristics and are a structural signal that the verification step is wrong.
• Never use background-wait patterns (command &, then wait or sleep loops) to track a running step. Run verification commands synchronously via Bash with timeout set high enough; use run_in_background: true only when the task description explicitly requires a background process AND the step does not need to read the result.
• Never suppress errors in Bash tool calls (2>/dev/null, || true, || :, -q flags that hide exit codes). Every Bash tool call MUST surface its exit code cleanly so the verification loop can detect failures.
• Never dispatch a build-class command with run_in_background: true from a task body. This skill runs as a leaf and cannot reap a backgrounded build — resolve every build/verify command through the architecture-resolved envelope and branch on execution_tier: run a per_task build inline (synchronously, timeout: bash_timeout_seconds * 1000) and hand an orchestrator-tier build back to the orchestrator's await-long-running seam. This is the leaf-no-background-build invariant; the canonical rule and rationale live in the leaf/dispatch-topology SSOT at `../ref-workflow-architecture/standards/agents.md` § "Leaf cannot reap a backgrounded build" — do not restate them here.
• Never commit, and never defer a commit as this skill's own obligation. Every per-deliverable commit belongs to the phase-5-execute envelope's Step 10a chain-tail (see `../phase-5-execute/SKILL.md` § Step 10a "Commit-ownership contract") — the leaf leaves its edits in the working tree for the envelope to commit; returning done while treating the commit as pending leaf work is a contract violation the boundary clean-tree guard converts into a worktree_dirty_at_boundary refusal.

Constraints:

• Strictly comply with all rules from persona-plan-marshall-agent, especially tool usage and workflow step discipline
• Never bypass the manage-tasks next/finalize-step loop — if parallelization is needed, it must happen at the TASK level, not at the STEP level within a task
• A task in implementation or module_testing profile MUST NOT be marked done until the resolved canonical command (quality-gate or verify respectively) exits cleanly. Module-tests passing alone is necessary but not sufficient — mypy and ruff must also pass.
• Before every Bash tool call, emit an [ATTEMPT] work-log line that names the command being run. Use: python3 .plan/execute-script.py plan-marshall:manage-logging:manage-logging work --plan-id {plan_id} --level INFO --message "[ATTEMPT] (plan-marshall:execute-task) {short description of command}". This provides an auditable trail when a Bash call hangs or produces unexpected output.

See workflow-integration-git/standards/worktree-handling.md for the worktree-specific application of this rule (never-edit-main-checkout invariant, git -C rule, dispatch header propagation).

Infeasible-deliverable contract

When a planned deliverable turns out to be infeasible during execution — the target cannot be cleanly built as the task specifies (the required surface does not exist, an assumed precondition is false, or building the named artifact is structurally impossible as scoped) — the skill MUST treat this as a first-class terminal outcome, NOT improvise a weaker substitute:

• (a) Stop, mark, and report. Stop work on the task, mark it infeasible via manage-tasks update --status infeasible, and return status: infeasible with a populated infeasibility_reason field naming why the deliverable cannot be built as scoped. The task flows into the phase-5-execute Step 11 triage surface (the same path a blocked task takes) so a gate-level planning decision — drop / re-scope into a new task / abort — resolves it. infeasible is terminal: it is never resolved by resuming the same task.
• (b) No silent substitution. Narrowing the deliverable into a buildable-but-valueless artifact under the original name — so the step "passes" while delivering none of the declared value — is the PROHIBITED anti-pattern. It is a contract violation equivalent to silently abandoning a blocked task, NOT a permitted pivot. The infeasibility is a real failure and belongs on the triage surface, never hidden behind a substituted deliverable.
• (c) Exact-pattern-spec rule. When a step specifies an explicit literal — a regex, an enum value, an exact constant, a literal string — that literal is a HARD copy-target. Copy it verbatim from the step; never approximate or reconstruct it from general knowledge. A literal that "looks right" but does not match the spec character-for-character is a defect, not an acceptable rendering.
• (d) Update-tests-not-implementation rule. In a breaking redesign, failing tests are MIGRATION WORK: the design specification is the authority, and the tests must be brought into line with it. Preserving the old form to keep tests green inverts the redesign and is prohibited. Update the tests to assert the new specified behavior; do not soften the implementation back toward the legacy form merely to satisfy a stale test. (This is the per-task application of the Scope-Deviation Escalation guard below — when satisfying the spec as written feels structurally riskier than estimated, escalate via AskUserQuestion rather than silently keeping both surfaces.)

---

Common Workflow

Input Contract

Every invocation of this skill MUST provide the following inputs:

Parameter	Type	Required	Description
plan_id	string	Yes	Plan identifier. Used by all manage-* script calls AND as the worktree-binding token: the skill resolves the active worktree internally via plan-marshall:manage-status:manage-status get-worktree-path --plan-id {plan_id}. The resolved path is the mandatory root for every Edit/Write/Read tool call during this task. When the plan runs against the main checkout (metadata.use_worktree == false), the resolved path is empty and operations target the main checkout directly. See workflow-integration-git/standards/worktree-handling.md for the canonical --plan-id two-state binding.
task_number	number	Yes	Numeric task id to execute
worktree_path	string	Deprecated	Deprecated — kept only for backward compatibility with callers that still pass an absolute path. New callers MUST forward only plan_id. When supplied, the value MUST agree with the path resolved from plan_id; treat any disagreement as fail-loud.

Callers (the phase-5-execute envelope, itself dispatched as execution-context-{level} under the phase-5-execute role key, which then LOADS this skill in-context once per task — "dispatching this skill" here means in-context Skill: loading within that single envelope, never a per-task Task: subagent dispatch) MUST forward plan_id verbatim — no absolute-path forwarding is required. This skill runs as a leaf and issues no Task: subagent dispatches of its own; the path-free worktree binding is inherited via the pinned cwd (ADR-002), not echoed into any further dispatch.

All profiles share the steps below. Profile-specific steps are documented in each profile section.

Step: Resolve Stale Targets

Check if a rename mapping exists and rewrite step targets before loading the task. This handles cases where earlier tasks renamed directories, making subsequent step targets stale.

python3 .plan/execute-script.py plan-marshall:manage-files:manage-files exists \
  --plan-id {plan_id} --file work/rename_mapping.toon

If exists: true, the rename mapping has already been applied to task step targets at recording time (by the rename-path subcommand). No further action needed — proceed to Load Task Context. The mapping file serves as an audit trail of path changes during the plan.

Step: Load Task Context

python3 .plan/execute-script.py plan-marshall:manage-tasks:manage-tasks read \
  --plan-id {plan_id} --task-number {task_number}

Extract key fields: domain, profile, skills, description, steps, verification, depends_on. Verify profile matches the expected profile for this execution.

Step: Mark Step Complete

After completing each step:

python3 .plan/execute-script.py plan-marshall:manage-tasks:manage-tasks finalize-step \
  --plan-id {plan_id} --task-number {task_number} --step {N} --outcome done

Step: Run Verification

After all steps complete, run task verification using commands from task.verification.commands.

Sub-step: Auto-inject --plan-id for Bucket B commands

When the plan resolves to an active worktree, before executing any task.verification.commands[N], route the command through the injection helper, passing --plan-id {plan_id} directly:

python3 .plan/execute-script.py plan-marshall:execute-task:inject_project_dir \
  run --command "{verification_command}" --plan-id {plan_id}

Injecting --plan-id (rather than --project-dir {worktree_path}) lets the Bucket B script auto-resolve the worktree path itself via its --plan-id/--project-dir two-state contract. No separate get-worktree-path resolution is required.

Parse the TOON output from the script's stdout. Use the rewritten_command value as the command to execute. When injected is true, log:

python3 .plan/execute-script.py plan-marshall:manage-logging:manage-logging \
  work --plan-id {plan_id} --level INFO \
  --message "[VERIFY] (plan-marshall:execute-task) Auto-injected --plan-id for {notation} (auto-resolves worktree path for the Bucket B script)"

The helper whitelists the eight Bucket B notations from plan-marshall:tools-script-executor/standards/cwd-policy.md; Bucket A manage-* notations and unknown notations pass through unchanged. The helper skips injection when the command already supplies --plan-id (no double injection) and when it already supplies an explicit --project-dir (a legacy override is respected untouched). See scripts/inject_project_dir.py for the authoritative whitelist.

Safety net (should not trigger in normal operation): If verification commands are missing, log a WARNING and resolve from architecture:

python3 .plan/execute-script.py plan-marshall:manage-logging:manage-logging \
  work --plan-id {plan_id} --level WARNING --message "[VERIFY] (plan-marshall:execute-task) TASK-{N} missing verification — falling back to architecture resolve"

python3 .plan/execute-script.py plan-marshall:manage-architecture:architecture \
  resolve --command {resolve_command} --module {module} \
  --audit-plan-id {plan_id}

Where {resolve_command} depends on profile: implementation → quality-gate, module_testing → verify. Verification profile uses commands from task steps directly.

Step: Handle Verification Results

If verification passes: Mark task done via manage-tasks update --status done.

If verification fails:

1. Analyze error output and identify failing component
2. Fix the issue (see profile-specific scope below)
3. Re-run verification
4. Iterate until pass (max max_iterations from config, default 5)

If still failing after max iterations: mark task as blocked and record details in work.log.

Scope-Deviation Escalation (per-task guard)

This skill runs as a dispatched leaf — it CANNOT fire AskUserQuestion (operator input is unreachable inside a dispatched execution-context envelope; see `../ref-workflow-architecture/standards/agents.md`). When a per-task verification deviation would soften a request-level hard requirement, the leaf MUST NOT raise the prompt in-leaf and MUST NOT log-and-continue. Instead it detects the softening, leaves the task not-done, and surfaces the deviation as a prompt_options[] entry handed back to the phase-5-execute loop, which batches it into an escalate_ask envelope for the main-context orchestrator to fire. The canonical deviation taxonomy, the three-option shape (Hold / Accept-with-rationale / Split), and the prohibited "log-and-continue" anti-pattern live in `../ref-workflow-architecture/standards/scope-deviation-escalation.md` — the single source of truth.

Detection: A deviation softens a hard requirement when, mid-fix-iteration, the implementor concludes that satisfying the request as written is structurally riskier than estimated and the conservative response would be to keep both the old and new surfaces. Concrete signals: verification command reports a non-zero hit count against a "zero-hit grep" gate; tests pass against a legacy code path the deliverable was meant to delete; the implementor is about to add a transition hedge ("until X has fully landed", "callers may still see Y") to satisfy the test gate.

Envelope shape: assemble a prompt_options[] entry carrying the canonical three-option shape from scope-deviation-escalation.md, each option's label naming the branch the orchestrator takes when it is chosen (Hold the line / Accept with rationale / Split into follow-up plan). When BOTH this gate and the smart_and_ask Compatibility Strategy gate (Step: Compatibility Strategy above) fire within one task, batch BOTH into ONE envelope rather than two. The leaf does not pause on the prompt — it returns the envelope (see § Return Results) and lets the phase-5-execute loop yield to the orchestrator.

Resolution (orchestrator-owned): the main-context orchestrator fires the batched AskUserQuestion and applies each option's side effect post-return per scope-deviation-escalation.md — Hold the line → resume the fix loop with the requirement intact (do NOT mark the task blocked merely because the softening was refused); Accept with rationale → persist the rationale to decision.log AND the PR body; Split into follow-up plan → seed a successor lesson. A [VERIFY] / [STATUS] / [OUTCOME] work-log line confirming the user's chosen option is allowed only AFTER the orchestrator resolves the prompt, never as a stand-in for it.

Anti-pattern: never batch a destructive checkout to re-baseline

Prohibition: When a fix iteration mishandled a too-narrow mechanical transform and the instinct is to "get back to a clean baseline" before re-applying it, NEVER batch git checkout -- <files> or git restore <files> across the modified files. These commands are destructive of uncommitted working-tree content with no undo — content that was never staged is not recoverable via git fsck. The risk is amplified by the per-deliverable-commit model this phase runs under: a file that shows as modified-vs-HEAD may carry another deliverable's uncommitted output, so a batched revert can silently roll a completed deliverable's files back to pristine while the plan still believes that work is done. The plan's task ledger then disagrees with the worktree, and the loss surfaces only after the missing changes are noticed downstream.

Safe alternatives (in preference order):

1. Operate forward-only. Do not revert at all. Broaden the mechanical transform so it is idempotent and re-apply it on top of the current tree — dry-run it against a synthetic fixture first to confirm the broadened form converges, then apply it to the real files. A forward-only re-apply never discards working-tree content.
2. If a revert is genuinely required, scope it to your own files only. Revert ONLY the specific files YOU modified in THIS task, one path at a time, and only after confirming each carries no other uncommitted content. Per-path proof before each revert: git status --short <path> shows only your expected change, and git diff --quiet HEAD -- <path> (or an explicit review of git diff HEAD -- <path>) confirms there is no foreign work mixed in. A batched, whole-directory checkout never satisfies this proof — it cannot distinguish your changes from a sibling deliverable's.

When batching IS permissible: a batched checkout/restore is only safe when every target path is known to hold no uncommitted work — e.g. the files are unmodified-vs-HEAD, or the worktree was committed clean immediately before. Absent that guarantee, treat the batched destructive checkout as forbidden and fall back to alternative 1.

Step: Record Lessons

On issues or unexpected patterns, first run the canonical three-gate lesson-creation policy in `../manage-lessons/standards/lesson-creation-policy.md` — Gate 1 (dedup), Gate 2 (active-plan check), Gate 3 (create). The two-step path-allocate flow below is Gate 3, reached only when Gates 1 and 2 both clear; when Gate 1 returns merge_into / already_closed or Gate 2 finds a covering active plan, extend the existing lesson or fold into the plan instead of allocating a new one. Do not restate the gate mechanics — follow the standard.

When the gates clear, use the two-step path-allocate flow:

1. Allocate a lesson file and capture the returned path:

python3 .plan/execute-script.py plan-marshall:manage-lessons:manage-lessons add \
  --component "plan-marshall:execute-task" --category improvement \
  --title "{issue summary}"

2. Parse path from the output, then write the lesson body directly to that path via the Write tool. Markdown sections with ## headings, code fences, and multiple paragraphs are all safe because the body never passes through a shell argument.

Step: Return Results

Base output contract (profile-specific extensions noted in each section):

status: success | error | infeasible
plan_id: {echo}
task_number: {echo}
execution_summary:
  steps_completed: N
  steps_total: M
  files_modified: [paths]
verification:
  passed: true | false
  command: "{cmd}"
next_action: task_complete | requires_attention
message: {error message if status=error}
infeasibility_reason: {required when status=infeasible — why the declared deliverable cannot be built as scoped}
escalate_ask: {true when a scope-deviation or smart_and_ask gate fired — the task is left not-done and the phase-5-execute loop batches prompt_options[] into an escalate_ask envelope for the orchestrator to fire}
prompt_options[N]{id,question,header,options,recommended}: {the batched deviation / compatibility questions — present only when escalate_ask is true}

When escalate_ask is true, the leaf leaves the task not-done, sets next_action: requires_attention, and populates prompt_options[] with the canonical three-option deviation shape (Scope-Deviation Escalation) and/or the uncertain-change compatibility questions (smart_and_ask). Both gates firing in one task batch into ONE prompt_options[] list. The leaf performs no operator-facing interaction — the main-context orchestrator fires the batched AskUserQuestion and applies each resolution's side effect post-return (see `../ref-workflow-architecture/standards/scope-deviation-escalation.md` and plan-marshall/workflow/execution.md § "Post-return escalate_ask batched deviation dispatch").

next_action: task_complete is a per-task signal, not the phase's terminal payload. It reports single-task completion to the enclosing phase-5-execute envelope loop, which consumes it to advance to the next same-envelope_id task. Echoing a bare task_complete to the orchestrator while pending same-envelope tasks remain is the task_complete_returned_verbatim defect — see `../phase-5-execute/SKILL.md` § "Forbidden: agent-initiated checkpoints".

status: infeasible is the terminal return for a deliverable that cannot be built as the task specifies (the required surface does not exist, an assumed precondition is false, or building the named artifact is structurally impossible as scoped). It is distinct from error (an execution failure that may be retried) — infeasible is a planning-level verdict resolved by a gate decision (drop / re-scope into a new task / abort), never by re-running the same task. When status: infeasible, the skill MUST have already marked the task infeasible via manage-tasks update --status infeasible and MUST populate infeasibility_reason. See the Infeasible-deliverable contract in the Enforcement section.

---

Profile: implementation

Production code creation and modification.

Path Resolution

This skill runs as a leaf inside the execution-context envelope — it issues no Task: subagent dispatches. When the plan runs in an isolated worktree, cwd is pinned to the worktree root by ADR-002, so every Edit/Write/Read tool call in this profile uses the path verbatim (relative to the pinned cwd). Bucket A manage-* scripts remain cwd-agnostic (they take --plan-id). See workflow-integration-git/standards/worktree-handling.md for the canonical --plan-id two-state binding and plan-marshall:tools-script-executor/standards/cwd-policy.md for the Bucket A/B split.

Compatibility Strategy

Before implementing, read the compatibility approach:

python3 .plan/execute-script.py plan-marshall:manage-config:manage-config \
  plan phase-2-refine get --field compatibility --audit-plan-id {plan_id}

No fallback — if field not found, fail with error and abort task.

Apply throughout all subsequent steps:

• breaking: Make changes directly. Remove old code, rename freely, no backward compatibility.
• deprecation: Keep old APIs/methods with @Deprecated markers. Add new code alongside old. Provide migration notes in commit messages.
• smart_and_ask: For each change that could break consumers, evaluate impact. If uncertain, the leaf MUST NOT ask the user in-leaf — a dispatched leaf cannot reach the operator (see `../ref-workflow-architecture/standards/agents.md`). Instead it records the uncertain change as a prompt_options[] entry and hands it back to the phase-5-execute loop for the orchestrator to fire, batched into ONE escalate_ask envelope together with any scope-deviation entry that fired in the same task. See § Return Results and `../ref-workflow-architecture/standards/scope-deviation-escalation.md`.

Workflow

1. Understand Context: Read affected files (step.target) if they exist. Use architecture files --module {module} to enumerate the module's components and architecture which-module --path P / architecture find --pattern P for module-spanning lookups; fall back to Grep for content-level searches inside known files and Glob for sub-module path patterns or when the architecture verb returns elision. Apply domain knowledge from loaded skills. Opt-in LSP lookup: when a language server is configured and reachable for the task's language — plan-marshall:lsp-client's preflight returns state: ready — you MAY locate symbols by coordinate with its lookup verb (definition / references / document-symbol / workspace-symbol) instead of reading whole files; a provider_count: 1 result with an empty locations[] is a real "found nothing", distinct from a degraded (no-server) return. When preflight returns not_configured or unreachable, use the architecture / Grep / Glob / Read ladder above — behaviour is byte-identical to a project that never configured a server. See `../lsp-client/SKILL.md` § Canonical invocations. How completely the in-radius items are read and how deeply their relations are traced is the task's declared thoroughness over its declared scope — see the two-dial coverage contract (thoroughness ladder T1–T5, scope ladder, grade-to-the-floor, coupling constraint reject thoroughness ≥ T4 ∧ scope < component) in `persona-plan-marshall-agent/standards/thoroughness.md`.

2. Plan Implementation: For each step, determine changes needed, domain skill patterns to apply, modification order, and integration considerations.

3. Implement Changes: For each step — create new files with Write, modify existing files with Edit. Apply domain patterns and maintain existing code style. Opt-in LSP rename: for a multi-file symbol rename with a configured, reachable server, plan-marshall:lsp-client's edit verb produces a WorkspaceEdit, captures its footprint from the edit, applies it, and re-runs diagnostics — a worsened diagnostic set fails the step and rolls the change back. Treat a failed return as a rejected edit (investigate the reported diagnostics; do not force it) and a degraded return as "unavailable" (fall back to Edit). This does not replace Step 5 verification — lsp-client's diagnose is a pre-build signal that supplements, never replaces, the quality gate.

4. Mark Step Complete (common step)

5. Run Verification — the per-task gate is two complementary checks: (a) the static-analysis quality-gate canonical, then (b) a task-scoped breakable-test module-tests run over exactly the tests the task's own change could break.

   (a) Static-analysis gate — resolve command: quality-gate (mypy + ruff on the touched production sources). Resolve it through the architecture-resolved envelope and honour execution_tier (run per_task inline, hand orchestrator-tier back — the leaf-no-background-build invariant in the Enforcement block).

   (b) Task-scoped breakable-test gate — run the tests the task's own change could break, scoped to the task's changed files (NOT the whole tree):

   1. Resolve the task's changed paths from git — the files this task modified in the worktree — and join them comma-separated as {changed_paths}.

   2. Resolve the scoped test target via the task-scoped footprint:

      python3 .plan/execute-script.py plan-marshall:build-pyproject:pyproject_build resolve-test-scope \
        --changed-paths "{changed_paths}" --plan-id "{plan_id}"

      Parse scoped_modules[], recommended_target, and divergence_possible from the TOON.

   3. Docs-only short-circuit: when scoped_modules is empty (recommended_target: None) the change resolves to no buildable module — run NO pytest for this task; the breakable-test gate is a no-op. Proceed to Step 6.

   4. Otherwise run the scoped module-tests build: scope it to recommended_target when set, or run whole-tree when divergence_possible: true (a multi-module / shared-infra change whose scoped run could diverge from a whole-tree run). Resolve module-tests through the architecture-resolved envelope and honour the stamped execution_tier — run a per_task build inline (synchronously, timeout: bash_timeout_seconds * 1000), hand an orchestrator-tier build back to the orchestrator's await-long-running seam (the same leaf-no-background-build invariant). After the build, read the result TOON status / errors[] — the wrapper exits 0 even on failure.

   Seam reconciliation (supersedes the prior "per-task gate resolves quality-gate — not module-tests — by design" narrative for the breakable-test scope): the per-task gate now runs BOTH static analysis AND the task-scoped breakable-test slice — the narrow slice of tests the task's own change could break, kept lean by scoping to the changed files only. This is complementary to, not a replacement for, the per-deliverable Step 10b module-tests chain-tail in phase-5-execute: the per-task gate is the breakable slice (caught at the leaf, immediately, per task), while Step 10b is the deliverable backstop (module-scoped over the changed module once all the deliverable's tasks settle, correctly resolved after the D1 which-module containment fix). The two seams read as a narrow-then-wider ladder, not as duplicate or contradictory gates.

6. Handle Verification Results — fix scope: production code (a static-analysis or breakable-test failure is a real regression the task must fix before it is marked done)

7. Record Lessons, Return Results

Error Handling

Failure	Action
Conflicting changes	Analyze conflict, prefer preserving existing behavior, ask for clarification if needed

---

Profile: module_testing

Unit and module test creation.

Path Resolution

This skill runs as a leaf inside the execution-context envelope — it issues no Task: subagent dispatches. When the plan runs in an isolated worktree, cwd is pinned to the worktree root by ADR-002, so every Edit/Write/Read tool call in this profile uses the path verbatim (relative to the pinned cwd). Bucket A manage-* scripts remain cwd-agnostic (they take --plan-id). See workflow-integration-git/standards/worktree-handling.md for the canonical --plan-id two-state binding and plan-marshall:tools-script-executor/standards/cwd-policy.md for the Bucket A/B split.

Workflow

1. Understand Implementation Context: Use architecture files --module {module} to enumerate the module's components and architecture find --pattern '*{name}*' for module-spanning lookups; fall back to Grep for content-level searches inside known files and Glob for sub-module path patterns or when the architecture verb returns elision. Read and examine the matched implementation files. Identify testable elements: public methods, edge cases, error conditions, input validation, integration points. (Opt-in: when a language server is configured and reachable, plan-marshall:lsp-client's lookup verb locates symbols by coordinate without reading whole files; on not_configured / unreachable the ladder above is byte-identical — see `../lsp-client/SKILL.md`.)

2. Plan Test Implementation: For each step, determine test scenarios, test structure (unit vs integration per domain skills), assertions needed, setup/teardown requirements.

3. Implement Tests: For each step — create new test files with Write, modify existing test files with Edit. Follow the AAA pattern (Arrange-Act-Assert). Include positive and negative test cases with descriptive names.

4. Mark Step Complete (common step)

5. Run Verification — resolve command: verify (full verify pipeline for the module — quality-gate + module-tests)

6. Handle Verification Results:

   Sub-step: Diff written test identifiers against the module-test log

   After a green module-tests run that produced new test files during step 3 (Implement Tests):

   1. Collect pytest nodeids for every newly-written test file: {rel_path}::{test_function} for each test function. Write them to a temp file under .plan/temp/ (one identifier per line).
   2. Run the diff assertion helper:

   python3 .plan/execute-script.py plan-marshall:execute-task:assert_test_identifiers \
     run --identifiers-file "{temp_identifiers_file}" --log "{module_test_log_path}"

   3. On passed: true: proceed to the standard "Mark task done" path.
   4. On passed: false: do NOT mark the task done — the run was silently incomplete. Log, mark the task requires_attention, and surface the mismatch in the return value:

   python3 .plan/execute-script.py plan-marshall:manage-logging:manage-logging \
     work --plan-id {plan_id} --level WARNING \
     --message "[VERIFY] (plan-marshall:execute-task) Diff assertion failed: {missing_count} written test identifiers absent from module-test log — {missing}"

   On test failure: determine if test logic is wrong or implementation has a bug. If test logic → fix test. If implementation bug → fix production code AND the test. Adapting production code to make tests pass is expected within this profile.

7. Record Lessons, Return Results

Output Extensions

execution_summary:
  tests_written: N
  coverage_impact: {if available}
verification:
  tests_passed: N
  tests_failed: N
  diff_assertion:
    passed: true | false
    missing_count: N
    missing[]: [identifier, ...]

Note: diff_assertion.passed: false overrides tests_passed — a green test count does not imply a successful run if written identifiers are absent from the log.

Error Handling

Failure	Action
Implementation not found	Check if implementation task is in dependencies. If yes → mark task as blocked. If no → note in lessons learned

---

Profile: verification

Run verification commands without modifying files.

Note: The verification profile is distinct from the verification change-type (see change-types.md). The profile determines HOW a task executes; the change-type describes WHY a request was made.

Workflow

1. Resolve Worktree Path: Before executing any verification step, resolve the active worktree path from plan_id:

   python3 .plan/execute-script.py plan-marshall:manage-status:manage-status get-worktree-path \
     --plan-id {plan_id}

   Capture the returned worktree_path. When metadata.use_worktree == false the returned path is empty — skip the auto-injection sub-step below and execute each raw step.target directly against the main checkout.

2. Execute Verification Steps with --plan-id Auto-Injection: Steps contain verification commands (not file paths). Execute sequentially. For each step.target:

   a. Route the command through the injection helper, passing --plan-id {plan_id}:

   python3 .plan/execute-script.py plan-marshall:execute-task:inject_project_dir \
     run --command "{step.target}" --plan-id {plan_id}

   Parse the TOON output. Use the rewritten_command value as the command to execute. When the worktree path resolved in Step 1 is empty (main-checkout flow), skip the helper and execute the raw step.target. Injecting --plan-id lets the Bucket B script auto-resolve the worktree via its two-state contract.

   b. Execute the resulting command with a Bash timeout derived from the architecture-resolved canonical envelope. See plan-marshall:persona-plan-marshall-agent § "Bash: Timeout from architecture-resolved canonical command" for the authoritative rule: read bash_timeout_seconds and execution_tier from the resolved TOON, pass timeout: bash_timeout_seconds * 1000 when execution_tier=per_task, and hand off to the orchestrator when execution_tier=orchestrator. The 600000ms floor still applies to ad-hoc invocations that do not flow through architecture resolve, and matches CLAUDE.md § Build Commands.

   c. On injected: true, emit the standard auto-injection work-log entry:

   python3 .plan/execute-script.py plan-marshall:manage-logging:manage-logging \
     work --plan-id {plan_id} --level INFO \
     --message "[VERIFY] (plan-marshall:execute-task) Auto-injected --plan-id for {notation} (auto-resolves worktree path for the Bucket B script)"

   Check exit code and output of the executed command. This step mirrors the Common Workflow → Step: Run Verification sub-step; see that section for the authoritative whitelist of Bucket B notations, the no-inject pass-through rule for Bucket A manage-* notations, and the rationale for skipping injection when the command already supplies --plan-id or an explicit --project-dir.

3. Mark Step Complete (common step)

4. Handle Failures: This is a verification task — do NOT modify source files. Report failures with structured output for triage. If verification fails, mark task as blocked.

5. Record Lessons (on unexpected failures or environment issues), Return Results

No domain skills are needed for this profile.

Output Extensions

execution_summary:
  commands_run: [commands]
verification:
  exit_code: {exit_code}
  stderr: "{truncated stderr, max 2000 chars}"
  findings:
    - type: {compile-error|test-failure|lint-issue}
      file: {file_path}
      line: {line_number}
      message: "{error message}"

On success: next_action: task_complete.
On failure: status: error, next_action: requires_attention, plus the extension fields. The findings array is best-effort: parse compiler errors, test failures, or lint output into structured entries. If parsing fails, include the raw stderr.

Canonical invocations

The canonical argparse surface for the two entry-point scripts this skill registers: inject_project_dir.py and assert_test_identifiers.py. The plugin-doctor analyzer (_analyze_manage_invocation.py) reads this section as source-of-truth for the manage-invocation-invalid and missing-canonical-block rules. Consuming docs xref this section by name instead of restating the command inline. See `pm-plugin-development:plugin-script-architecture` cross-skill-integration.md § "Script invocation in documentation".

inject_project_dir — run

python3 .plan/execute-script.py plan-marshall:execute-task:inject_project_dir run \
  --command COMMAND --plan-id PLAN_ID

assert_test_identifiers — run

python3 .plan/execute-script.py plan-marshall:execute-task:assert_test_identifiers run \
  --identifiers-file IDENTIFIERS_FILE --log LOG