cuioss/phase-5-execute

Phase Execute Skill

Role: DUMB TASK RUNNER that executes tasks from TASK-*.json files sequentially.

Execution Pattern: Locate current task → Execute steps → Mark progress → Next task

Phase Handled: execute

Foundational Practices

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

Enforcement

Shared lifecycle patterns: See phase-lifecycle.md for entry protocol, completion protocol, and error handling convention.

Execution mode: DUMB TASK RUNNER — locate task, execute steps, mark progress, next task. Follow workflow steps sequentially.

Prohibited actions:

• Never access .plan/ files directly — use manage-* scripts via Bash (Edit/Write tools trigger permission prompts on .plan/ directories)
• Never skip the phase transition — use manage-status transition
• Never improvise script subcommands — use only those documented below
• Never target file paths outside the active git worktree.
• Never stop to ask the user "should I run the integration/e2e tests?" — every entry in manifest.phase_5.verification_steps runs automatically. The manifest is the single authority for which verification steps fire; the composer already footprint-gated and resolvability-skipped each entry at compose time, so there is no per-run human decision about whether to run a given canonical verify step.

Constraints:

• Strictly comply with all rules from persona-plan-marshall-agent, especially tool usage and workflow step discipline

Exit-code convention for manage-* script calls

Every manage-* script call in this document carries the following exit-code contract unless a step explicitly states otherwise:

• exit_code == 0: parse the returned TOON and use the value as the step describes.
• exit_code != 0: STOP and return an error TOON to the orchestrator carrying the script's stderr verbatim. Non-zero exits include argparse_rejection (exit 2) — silent swallowing of wrong_parameters rejections is the prohibited anti-pattern; "log and continue" is equally forbidden.

Step-level exceptions — calls whose non-zero exit is itself the signal (e.g., manage-files exists returning exists: false, manage-status get-worktree-path returning an empty worktree_path) — are documented inline in the step that issues them.

• On phase entry (Step 4), resolve the active worktree absolute path and surface it as a [STATUS] work-log line so it stays visible in model context throughout the run.
• Every subagent dispatch (Task / Skill / execution-context invocation) MUST pass plan_id as an input parameter to satisfy the subagent's Input Contract (e.g., execute-task, execution-context). Under the cwd-pinned model the worktree binding is inherited via the pinned cwd, not forwarded per call — subagents resolve .plan/ cwd-relatively. A short reminder header MAY be embedded (see Dispatch Protocol below); it is a salience reminder, not a path-routing mechanism.

See workflow-integration-git/standards/worktree-handling.md for the worktree-specific application of this rule (path convention, never-edit-main-checkout invariant, cwd-pinned dispatch inheritance, --plan-id/--project-dir escape-hatch contract).

Phase-Entry Worktree Assertion

The Phase Entry Protocol's phase_handshake verify --phase {previous_phase_key} --strict call (see ref-workflow-architecture/standards/phase-lifecycle.md) asserts the worktree-resolution contract before any phase-5-execute work begins: when metadata.use_worktree==true, metadata.worktree_path MUST be non-empty AND filesystem-resolvable (the directory exists AND git -C {path} rev-parse --show-toplevel returns the same canonical path). When the assertion fails, the script returns status: error, error: worktree_unresolved and (under --strict) exits 1 — phase entry refuses to advance until the persisted metadata is repaired. Plans with metadata.use_worktree==false skip the assertion (main-checkout flow). The assertion fires uniformly at every phase boundary; see deliverable 8 in the originating lesson plan for the full contract.

Phase 5 is the materialization phase. Phases 1–4 only declare the worktree intent (metadata.use_worktree, written by phase-1-init); the feature branch is not recorded early — Step 2.5 below is the single point where the worktree directory and feature branch are actually created on disk, deriving the branch as feature/{plan_id} and persisting metadata.worktree_branch then. Step 2.5 is unconditional and runs BEFORE the early_terminate short-circuit evaluation (Step 2.6 below). Hoisting Step 2.5 above the short-circuit guarantees that metadata.worktree_path is always backfilled regardless of the manifest's early_terminate flag — otherwise an analysis-only plan that the composer marks early_terminate=true would transition to finalize without ever populating the worktree path, and the phase_handshake verify assertion at the 5→6 boundary would fail with worktree_unresolved. This ordering rules out an early-terminate path that transitions to finalize without ever populating the worktree path. Re-entry semantics: when phase-4-plan's capture ran without a populated metadata.worktree_path (because Step 2.5 had not yet executed), the phase_handshake verify --phase 4-plan --strict call MUST tolerate the still-empty value at phase-5 entry, then Step 2.5 populates worktree_path in both references.json and status.metadata before any task dispatch. On every subsequent phase-5 re-entry (orchestrator re-dispatch), Step 2.5's idempotence guard observes the populated worktree_path and short-circuits — no re-creation, no duplicate git checkout -b.

Dispatch Protocol (cwd-Pinned Inheritance)

Under the move-based, cwd-pinned model (ADR-002), the worktree binding is carried by the pinned current working directory, not by per-call path forwarding. Step 2.5 pins the orchestrator's cwd to the worktree root after prepare_execute.py moves the plan directory and the executor in; every subprocess spawned and every subagent dispatched inherits that cwd, and .plan/ resolution is cwd-relative (file_ops.get_base_dir() walks up to the nearest ancestor containing .plan/local). A dispatched subagent therefore resolves the worktree-resident state without being told a path. See ../tools-script-executor/standards/cwd-policy.md for the single cwd-unchanged invariant — it is not restated here.

Consequence for dispatch: subagents do NOT forward --plan-id or --project-dir to working-tree-touching scripts for path resolution — the inherited pinned cwd binds them. The only structured input a dispatched workflow takes is plan_id as the plan-identifier prompt-body field (selecting which plan's metadata to operate on; unrelated to cwd binding), per the dispatched workflow's own Input Contract (execute-task, the per-phase workflow doc loaded by execution-context-{level}).

When the plan runs in an isolated worktree (see the [STATUS] Active worktree work-log line from Step 4), a short reminder header MAY be embedded as the first lines of the dispatch prompt to keep the never-edit-main-checkout invariant salient through free-form delegation:

WORKTREE: cwd is pinned to this plan's worktree (ADR-002 cwd-pinned model).
Resolution is cwd-relative — do NOT forward a worktree path; do NOT pass --project-dir.
All Edit/Write/Read tool calls and tool invocations (git, mvn, npm) operate against the pinned cwd.
NEVER edit the main checkout. See tools-script-executor/standards/cwd-policy.md.

The header is a reminder, not a path-routing mechanism — the binding holds whether or not the header is present, because cwd-pinning carries it. The worktree absolute path MUST NOT appear in dispatch prompts. --project-dir <abs> survives only as the escape hatch for callers invoked outside a pinned-cwd context (post-worktree-removal cleanup, fixture-driven test invocations); it MUST NOT be forwarded inside a phase-5+ pinned-cwd dispatch. The complete contract — cwd-pinned inheritance, the merge-lock exception, and the --plan-id/--project-dir three-state escape-hatch binding — is documented in workflow-integration-git/standards/worktree-handling.md § Dispatch Protocol and ../tools-script-executor/standards/cwd-policy.md.

The [STATUS] Active worktree: ... work-log line is the observability signal that the worktree was materialized and cwd was pinned. Child agents inherit the same pinned cwd and MUST NOT re-derive or forward a worktree path into any further dispatch.

This applies to every dispatch in the execution loop, including (but not limited to) Step 6 (Execute Steps) task dispatches and Step 9 (Independent Change Verification) subagent invocations.

See standards/operations.md for the complete set of dispatch pattern templates and workflow-integration-git/standards/worktree-handling.md for the worktree-specific application of this rule.

Common anti-patterns to avoid (mirrored from persona-plan-marshall-agent)

Each Bash tool call dispatched during execute must contain exactly ONE command. Never combine with newlines, &, &&, ;, or inline env-var assignment of the form VAR=val cmd. The VAR=val cmd shape combines the assignment and the command into one shell argument, which trips the host platform's permission UI and obscures the env-var contract by hiding the variable inside the command line rather than declaring it explicitly.

Anti-pattern: MY_VAR=value python3 some_command.py ...

Safe alternative (option A) — Pass the value as a flag arg:

python3 some_command.py ... --my-var value

Safe alternative (option B) — Set the env var in the command's invocation header (e.g., a separate env MY_VAR=… line, NOT inline) before launching the bash command, or define the value as a Python module-level constant lookup inside the script itself.

See persona-plan-marshall-agent Hard Rules for the authoritative source.

cwd for .plan/execute-script.py calls

manage-* scripts (Bucket A) resolve .plan/ via git rev-parse --git-common-dir and work from any cwd. Build / CI / Sonar scripts (Bucket B) bind to a working tree via --plan-id when a worktree is active. See plan-marshall:tools-script-executor/standards/cwd-policy.md for the Bucket A/B split and workflow-integration-git/standards/worktree-handling.md for the worktree-specific application of this rule.

---

Standards (Load On-Demand)

Workflow

Read standards/workflow.md

Contains: Task execution pattern, phase transition, auto-continue behavior

Operations

Read standards/operations.md

Contains: Delegation patterns for builds, quality checks, PR creation

Recovery Patterns

Read standards/recovery.md

Contains: First-line response to mid-plan origin/main advances — stash + merge + pop, with works/does-not-work conditions and rationale vs rebase.

Coverage Contract (per-task scope × thoroughness)

Read ../persona-plan-marshall-agent/standards/thoroughness.md

Contains: the scope × thoroughness coverage contract each task body honors at its execution coverage point — the thoroughness ladder (T1–T5), the scope ladder, the grade-to-the-floor rule, and the coupling constraint reject thoroughness ≥ T4 ∧ scope < component. Per-task work covers its declared cell; the floor-graded self-report states asked-for vs achieved.

---

Dispatched workflows vs inline steps

This document IS the phase-5 envelope workflow. The orchestrator's Execute-phase dispatch (see ../plan-marshall/workflow/execution.md § "Execute Phase") MUST wire its workflow: field to plan-marshall:phase-5-execute/SKILL.md — THIS runner, which drives the envelope task loop and LOADS execute-task in-context per task. The execution-context dispatcher executes the workflow doc; the skills[] list is loaded only for context. Wiring the workflow: field to plan-marshall:execute-task/SKILL.md (the single-task runner that returns next_action: task_complete after ONE task) is the dispatch defeat: the leaf then runs exactly one task and echoes task_complete, forcing the orchestrator into a wasteful per-task re-dispatch loop. execute-task appears in skills[] because THIS workflow loads it in-context per task — never as the dispatched workflow.

One-context-per-phase invariant (durable contract). A core phase completes in exactly ONE execution-context dispatch per envelope group — precisely envelope_count dispatches (typically 1, and NEVER one dispatch per task). The bin-packer pre-computed the envelope grouping at plan time; the executor drives its whole assigned envelope_id group to exhaustion inside a single dispatch, yielding only at the three sanctioned TASK-boundary points (budget_yield / triage_required / baseline_drift). The only sanctioned sibling dispatches OUT of this envelope are the ones the main-context orchestrator owns (the verification-feedback triage and the adversarial validators) — the leaf returns a signal, it never dispatches. Returning a bare task_complete echo while pending same-envelope tasks remain is the task_complete_returned_verbatim defect (see § "Forbidden: agent-initiated checkpoints"): it collapses the envelope loop into a per-task re-dispatch and is a control-flow violation, NOT a legitimate yield.

This phase dispatches under one role key: phase-5-execute (resolves through phase-5-execute.default). The dispatch unit is envelope-bounded — explicitly NEITHER per-task NOR per-deliverable. The continue-vs-yield decision is pre-computed at plan time: phase-4-plan's bin-packer (manage-tasks pack-envelopes) groups tasks into execution envelopes in depends_on order under per_envelope_budget_tokens, stamps each task with an envelope_id, and records envelope_count in the manifest. The orchestrator then dispatches ONE execution-context envelope per envelope_id group (exactly envelope_count dispatches), passing each its assigned envelope_id; the dispatched executor runs only the tasks whose envelope_id matches its assigned group and yields when its group is exhausted — a trivial countable "is the next pending task in MY envelope?" equality check with NO runtime cost-summing, NO threshold evaluation, NO self-measurement (the subagent cannot measure its own mid-turn context usage; only the orchestrator sees post-return <usage>, which it feeds back to recalibrate the size→token table). Per task the envelope LOADS the unified execute-task skill in-context as a Skill: with the task-declared skill list as runtime input — leaf-legal in-context skill loading per persona-plan-marshall-agent, explicitly NOT a per-task Task: subagent dispatch. The envelope yields to the orchestrator — which then re-dispatches the next envelope group — only at one of three TASK-boundary re-dispatch points: (a) envelope-group exhaustion (the next pending task's envelope_id differs from the assigned group → a budget_yield decision is logged and a wrapped terminal TOON is returned); (b) triage_required (Step 11/11b verify / quality-gate failure); (c) baseline_drift. It is NOT one envelope per task and NOT one envelope per deliverable. Deliverable boundaries govern the COMMIT + FOCUSED-BUILD points only — the Step 10 Per-Deliverable Chain-Tail (an unconditional per-deliverable commit, plus a focused per-module build) is a sub-event that fires within OR across envelopes, decoupled from where the budget sentinel yields, and is NOT a dispatch boundary; because the sentinel yields at TASK boundaries (between finalize-step and the next manage-tasks next), a mid-deliverable yield is normal and lossless (finalize-step persists in-flight task state, and the Step 10 chain-tail commit still fires whenever the run reaches it, regardless of which envelope gets there). This per-task body runs as a leaf inside the execution-context envelope — it cannot itself issue a Task: dispatch (see ref-workflow-architecture/standards/agents.md, the canonical leaf/dispatch-topology contract). The built-in verification steps (the single parameterized default:verify:{canonical} step) stay inline as pure build invocations — no LLM judgement, no envelope. Step 9 independent change verification stays inline (three deterministic re-checks: git-diff empty-test, obfuscation-pattern grep, exit-code compare). Steps 11 and 11b detect the test-failure / lint-issue, persist each finding to the per-plan Q-Gate store (manage-findings qgate add — a script call, legal inside a leaf), then return a triage_required signal to the main-context orchestrator; the orchestrator owns the verification-feedback dispatch (--phase phase-5-execute --role verification-feedback, producer=build-runner) and consumes its return to drive the fix-task / suppress / accept branch. The leaf never dispatches verification-feedback itself. For the rationale see dispatch-granularity.md § 2, § 4 (Heuristic 3 — per-iteration Task: dispatch only when models differ OR iterations parallelise; the budget-bounded task loop is the iterate-in-context application), and § 5.1 (script over dispatch; phase-scoped resolution + producer-mode bundling).

Execution Loop

Step 1: Get Routing Context (Once at start)

Get current phase, skill routing, and progress in a single call:

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

Returns:

status: success
plan_id: {plan_id}
current_phase: 5-execute
skill: plan-marshall:phase-5-execute
skill_description: Execute phase skill for task implementation
total_phases: 4
completed_phases: 2
phases:
- init: complete
- refine: complete
- execute: in_progress
- finalize: pending

Use current_phase for logging, skill for dynamic routing, and completed_phases/total_phases for progress display.

Step 2: Read Execution Manifest (Once at start)

Read the execution manifest — the manifest is the single source of truth for which Phase 5 verification steps fire. It is composed by phase-4-plan Step 8b and stored at .plan/local/plans/{plan_id}/execution.toon:

python3 .plan/execute-script.py plan-marshall:manage-execution-manifest:manage-execution-manifest \
  read --plan-id {plan_id}

Extract phase_5.early_terminate (bool), phase_5.verification_steps (list[string]), and phase_5.step_execution_tier (list of {step_id, tier} records) from the output. Do NOT evaluate early_terminate yet — Step 2 only reads the manifest and caches the values. Step 2.5 (worktree materialization) MUST run before the early_terminate short-circuit fires, so the short-circuit evaluation is deferred to Step 2.6 below.

The verification steps to execute at end of phase come from phase_5.verification_steps — this replaces today's lookup of marshal.json's phase-5-execute.steps. The list is consumed by Step 11b (Final Quality Sweep) and the verification dispatch loop. See Verification Step Types below for dispatch rules. Each verification step's runnability is governed by the execution_tier the leaf resolves LIVE when it runs that step; the cached phase_5.step_execution_tier entry is the advisory compose-time expectation — see Per-step execution_tier (advisory compose-time stamp; live resolve routes) below.

The step IDs in the manifest are bare — cmd_compose's boundary normalization strips only the leading default: prefix, so a built-in canonical-verify step appears as verify:{canonical} (e.g., verify:quality-gate, verify:module-tests, verify:coverage). Translate them to the default: prefixed names used by the Built-in Step Dispatch Table by prepending default: for built-in steps — e.g., verify:quality-gate → default:verify:quality-gate. Steps that already contain : beyond the verify: segment for project/skill steps (a project: or bundle:skill prefix) are passed through verbatim.

---

Verification Step Types

The phase_5.verification_steps list from the manifest contains verification step references. Three step types are supported, distinguished by prefix notation (same model as phase-6-finalize):

| Type | Notation | Resolution | |------|----------|------------| | built-in | default: prefix (e.g., default:verify:quality-gate) | Execute built-in verification command (see dispatch table) | | project | project: prefix (e.g., project:verify-step-lint) | Skill: {notation} with interface contract | | skill | fully-qualified bundle:skill (e.g., my-bundle:my-verify-step) | Skill: {notation} with interface contract |

Type detection logic:

• Starts with default: -> built-in type (strip prefix, execute built-in command)
• Starts with project: -> project type
• Contains : (other) -> fully-qualified skill type

Each verify step declares an order: <int> value in its authoritative source — frontmatter on built-in standards docs (standards/{name}.md), frontmatter on project-local SKILL.md for project: steps, and the return-dict order field for extension-contributed skills. marshall-steward sorts the steps list by this value when writing it to marshal.json. This skill iterates the list as written and does NOT re-sort or validate order at runtime — the persisted order is the runtime order.

Built-in Step Dispatch Table

| Step Name | Action | Description | |-----------|--------|-------------| | default:verify:{canonical} | Resolve the trailing {canonical} via architecture resolve --command {canonical} and run the resolved executable | Single parameterized verify step backing every canonical (quality-gate, module-tests/verify, coverage, integration-tests, e2e, …); the canonical is a parameter, not a hardcoded branch — see standards/canonical_verify.md |

Dispatch detection: a step ID starting with the default:verify: prefix routes to the single parameterized canonical-verify step. The SKILL strips the default: prefix and feeds the trailing {canonical} segment to architecture resolve --command {canonical}. The full step body — canonical resolution, execution_tier/bash_timeout_seconds handling, the unresolved-canonical skip, and the module-scoped vs whole-tree invocation contract — lives in standards/canonical_verify.md; do NOT restate it here. Threshold enforcement for coverage is native to the resolved build command (pytest --cov-fail-under, JaCoCo build-tool config) — no secondary parse-and-check call is required.

Footprint gating (build-decision consult) for the end-of-phase default:verify:{canonical} loop

The end-of-phase whole-tree verification dispatch — Step 11b Final Quality Sweep and the loop that runs each default:verify:{canonical} step from phase_5.verification_steps — is footprint-gated at execution time via the manage-config build-decision verb, the sole build/no-build authority (see ADR-004 § "Amendment: build-decision is the sole build/no-build authority"). The consult call, the not_necessary skip path, and the decision != not_necessary run path are the same mechanics Step 11b spells out below for quality-gate; do NOT restate them here for other canonicals — apply the identical shape per-canonical. This gate applies ONLY to the whole-tree end-of-phase surface — the one footprint-blind phase-5 build surface. Steps 10b (focused per-deliverable build) and 11c (Execute-Exit Verify Gate) are already footprint-aware — they derive the live footprint on demand and already skip when no buildable module is present — and are UNCHANGED by this gate.

Per-step execution_tier (advisory compose-time stamp; live resolve routes)

Every entry of phase_5.verification_steps carries a stamped execution_tier in the manifest's phase_5.step_execution_tier record list ({step_id, tier} per step; tier ∈ {per_task, orchestrator} — see manage-execution-manifest/SKILL.md § "Per-step execution_tier stamping").

The stamp is ADVISORY; the live re-resolve is the routing authority. The stamped tier derives from the adaptive learned build duration, which every intervening build updates, so a ceiling-adjacent step's tier legitimately changes between compose and execute — the same plan has been composed with verify:coverage=per_task and then, after a single whole-tree build, verify:coverage=orchestrator. The leaf therefore re-resolves the tier live as it runs each step, per standards/canonical_verify.md § Workflow steps 1-2, and routes on THAT verdict. Read the stamp for planning — how many orchestrator-tier steps this plan expects — never as the sole routing input. A stamp that disagrees with the live tier is expected behaviour, not a defect: the live tier wins, always.

This dispatched phase-5 leaf branches on the live-resolved tier for every whole-tree verification step it would run (Step 11b Final Quality Sweep and Step 11c Execute-Exit Verify Gate) and for each default:verify:{canonical} end-of-phase step:

• tier == per_task — the step is in the leaf's runnable slice. Run the resolved executable inline (synchronously) per standards/canonical_verify.md, honouring the resolved bash_timeout_seconds. Read the result TOON.
• tier == orchestrator — the step is NOT in the leaf's runnable slice. The leaf MUST NOT run it (inline OR backgrounded). Return an orchestrator-tier yield signal (status: blocked, voluntary_checkpoint) naming the step so the main-context orchestrator runs it through the await-long-running detach-and-notify seam (the only component permitted to background a build).

A step whose stamped tier is absent (a manifest composed before this field existed) reads as per_task. per_task is the permissive default, not a safe floor — it is the value that would put a long build inline, where the host platform auto-backgrounds it past the Bash ceiling and this leaf cannot reap it. That default is safe only because the leaf re-resolves live before running: the live resolve, not the stamp, is what keeps a long build off the leaf.

Interface Contract for External Steps

Project and skill steps receive these parameters:

Skill: {step_reference}
  Arguments: --plan-id {plan_id}

Input contract: --plan-id only. Retry logic is managed by the task runner (Step 11 triage loop with max_iterations), not by the step itself.

Return Contract (required TOON output from external steps):

status: passed|failed
message: "Human-readable summary"

# Optional — only when status: failed
findings[N]{file,line,message,severity}:
src/Foo.java,42,Unused import,warning
src/Bar.java,10,Missing null check,error

• status: passed → step complete, continue to next step
• status: failed + findings[] → findings fed into Step 11 triage (fix task creation, suppress, or accept)
• status: failed without findings[] → treated as single unstructured failure, triaged as one finding

---

Step 2.5: Materialize Worktree and Feature Branch (Once per phase, idempotent)

Phase 5 is the materialization phase for the worktree. Earlier phases only persisted the intent (metadata.use_worktree, written by phase-1-init); this step derives the feature branch feature/{plan_id}, creates the worktree directory and that branch on disk, and propagates the resolved path to both references.json and status.metadata.worktree_path BEFORE Step 3 reads them. The derived branch is persisted to metadata.worktree_branch by prepare_execute's bookkeeping loop (via the --branch argument below), so every phase-6 reader of metadata.worktree_branch stays valid.

Idempotence guard (must run first): read metadata.worktree_path and short-circuit when it is already populated — Step 2.5 has already executed on a prior phase-5 entry, the directory exists on disk, and no re-creation is needed.

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

Extract metadata.use_worktree, metadata.worktree_path, and the plan's base_branch (from references.json via manage-references get). Derive the feature branch deterministically as worktree_branch = feature/{plan_id} — it is NOT read from metadata, since phase-1-init no longer records it. If worktree_path is non-empty, log the short-circuit and proceed to Step 3:

python3 .plan/execute-script.py plan-marshall:manage-logging:manage-logging \
  work --plan-id {plan_id} --level INFO \
  --message "[STATUS] (plan-marshall:phase-5-execute) Step 2.5 short-circuit: worktree_path already populated ({worktree_path}) — skipping materialization"

Materialization branch (when worktree_path is empty): branch on metadata.use_worktree.

Case A — use_worktree == true: CALL the atomic move-in script prepare_execute. In one call it materializes the worktree + feature branch (delegating to worktree-create), MOVES the plan directory and the executor into the worktree-resident .plan/, and RETURNS the canonical worktree_path:

python3 .plan/execute-script.py plan-marshall:workflow-integration-git:prepare_execute prepare \
  --plan-id {plan_id} --branch {worktree_branch} --base {base_branch}

Parse status and worktree_path from the returned TOON. On status: success, capture worktree_path. The move is atomic-with-rollback and idempotent — a re-entry returns action: noop carrying the same path (the idempotence guard above normally short-circuits before this CALL, but the script's own guard is the structural backstop). See workflow-integration-git Canonical invocations → prepare_execute for the full contract.

Pin cwd to the returned path: immediately after a successful CALL, the phase-5 orchestrator pins ITS OWN cwd to worktree_path for the remainder of phase-5+ (the script does NOT change the caller's cwd — a subprocess cannot mutate its parent's cwd). Under the uniform cwd rule (ADR-002) every subsequent .plan/ resolution then targets the worktree-resident state moved in by this CALL. The single execution-time invariant is that cwd is never changed away from the worktree until finalize moves the plan dir back. D8 owns the caller-side cwd-unchanged guard the lifecycle scripts assert.

Case B — use_worktree == false: the plan runs against the main checkout — there is NO worktree to materialize and NO move-in. Create the feature branch in place via git -C .:

git -C . checkout -b {worktree_branch}

Set worktree_path to the empty string (the main-checkout flow uses . everywhere worktree_path would otherwise apply; see Step 3's worktree_path absent → substitute . rule). Do NOT call prepare_execute in Case B — cwd stays on main and resolution is already main-relative.

Fatal-error contract: if either branch fails, abort the phase fail-loud and do NOT silently proceed to the task loop. Emit the canonical [ERROR] line per the Error Handling section and return the structured error TOON; the orchestrator surfaces the failure for human repair. The failure driver differs by case:

• Case A failure (prepare_execute returns status: error — worktree create failed, the plan dir is missing on main, or a move-in step failed and rolled back): the script's atomic-with-rollback contract guarantees the plan state is left WHOLLY on main, never half-moved. phase-1's expectation has already committed downstream consumers to the worktree path. Do NOT silently fall back to the main checkout — that would orphan every subsequent --plan-id-resolved Bucket B call. Do NOT pin cwd, and do NOT proceed to the task loop.
• Case B failure (git checkout -b {worktree_branch} exits non-zero, branch already exists with divergent history on the main checkout): the plan is already bound to the main checkout; the failure is the inability to create the feature branch in place. Do NOT silently fall back to main or --no-branch — phase-1 committed downstream consumers to a dedicated feature branch.

python3 .plan/execute-script.py plan-marshall:manage-logging:manage-logging \
  work --plan-id {plan_id} --level ERROR \
  --message "[ERROR] (plan-marshall:phase-5-execute) Worktree materialization failed for branch {worktree_branch} on base {base_branch}: {error_context}"

Persist worktree_path to both stores on success (skip when empty in Case B — the absence already signals the main-checkout flow):

1. Write to references.json via the manage-references typed setter:

   python3 .plan/execute-script.py plan-marshall:manage-references:manage-references set \
     --plan-id {plan_id} --field worktree_path --value {worktree_path}
   

2. Write to status.metadata.worktree_path via the metadata setter:

   python3 .plan/execute-script.py plan-marshall:manage-status:manage-status metadata \
     --plan-id {plan_id} --set --field worktree_path --value {worktree_path}
   

Both writes are required: references.json is the canonical artifact Step 3 reads to resolve worktree_path; status.metadata.worktree_path is the value the phase_handshake verify assertion checks on every subsequent phase boundary, and the value the idempotence guard above reads on phase-5 re-entry.

Log the materialization outcome:

python3 .plan/execute-script.py plan-marshall:manage-logging:manage-logging \
  work --plan-id {plan_id} --level INFO \
  --message "[STATUS] (plan-marshall:phase-5-execute) Step 2.5 materialized worktree at {worktree_path} on branch {worktree_branch}"

Proceed to Step 2.6.

Step 2.6: Evaluate early_terminate Short-Circuit (Once at start)

This step evaluates the phase_5.early_terminate flag cached at Step 2 and is intentionally placed AFTER Step 2.5 so the worktree directory and metadata.worktree_path are always populated before any early-exit path runs. The manifest composer narrows early_terminate=true to plans where BOTH verification_steps == [] AND the task queue is empty (no pending or in-progress tasks).

Early-terminate decision: If phase_5.early_terminate == true, log the decision and transition directly to phase-6-finalize — skip the entire execute loop including Steps 3 through 12:

python3 .plan/execute-script.py plan-marshall:manage-logging:manage-logging \
  decision --plan-id {plan_id} --level INFO \
  --message "(plan-marshall:phase-5-execute) Early terminate — manifest.phase_5.early_terminate=true; skipping execute loop and transitioning directly to phase-6-finalize"

Then jump directly to Phase Transition (below) to advance to finalize. Do NOT execute Steps 3–12. Because Step 2.5 already ran unconditionally, metadata.worktree_path is populated and the 5→6 phase_handshake verify assertion will succeed.

Otherwise (early_terminate == false): proceed to Step 3.

Step 3: Baseline Fast-Path Check (Once per phase)

Substantive baseline reconciliation now happens at refine time — see phase-2-refine/standards/refine-workflow-detail.md § Step 3d. Phase-5-execute is a fast-path "still clean?" verification: if the worktree branch is still ahead of (or merged with) origin/{base_branch}, continue to the task loop; if upstream commits have landed since the refine baseline-reconciliation pass, error out with a clear redirect — re-running phase-2-refine is the documented path. Phase-5-execute MUST NOT perform substantive reconciliation (no merge, no rebase).

Full procedure, fast-path semantics, error contract, and main-checkout fallback are documented in standards/sync-with-main.md.

Inlined flow:

1. Resolve base_branch and worktree_path from references.json (written at phase-1-init Step 6):

   python3 .plan/execute-script.py plan-marshall:manage-files:manage-files read \
     --plan-id {plan_id} --file references.json
   

   Extract base_branch and worktree_path. If worktree_path is absent, the plan runs against the main checkout; substitute . for {worktree_path} in every git command below.

2. Entry guard — stale base_branch check: before fetching, verify that origin/{base_branch} still resolves on the remote. A merged-and-deleted feature branch produces an empty ls-remote result, and a downstream git fetch origin {base_branch} will fail with a misleading could not find remote ref error.

   git -C {worktree_path} ls-remote --heads origin {base_branch}
   

   When the output is empty, return a structured error and ABORT the phase:

   status: error
   error: base_branch_unresolvable
   base_branch: {value}
   suggested_fix: "Update references.json:base_branch to the repo default (main/master), then re-enter phase-5-execute. Run phase-2-refine to invoke baseline-reconcile auto-update."
   

   baseline-reconcile (invoked by phase-2-refine Step 3d) self-heals stale base-branch values by detecting the remote default and writing the new value to references.json; the canonical recovery is therefore to re-run phase-2-refine, not to manually fix-up references.json in-place. The entry guard here is purely a fail-loud surface so the orchestrator does not waste a fetch round-trip on a known-bad input.

3. Fetch base (read-only network round-trip):

   git -C {worktree_path} fetch origin {base_branch}
   

4. Fast-path check — verify the current branch tip already contains origin/{base_branch}:

   git -C {worktree_path} merge-base --is-ancestor origin/{base_branch} HEAD
   

   Exit code 0 means up to date. Log and continue to Step 4:

   python3 .plan/execute-script.py plan-marshall:manage-logging:manage-logging \
     work --plan-id {plan_id} --level INFO \
     --message "[STATUS] (plan-marshall:phase-5-execute) Baseline fast-path: worktree already up to date with origin/{base_branch}"
   

5. Drift detected — exit code non-zero means upstream has new commits the worktree does not contain. Do NOT merge, do NOT rebase, do NOT continue to Step 4 yet. First capture the divergent commits for both logging and the self-absorb decision below:

   git -C {worktree_path} log --oneline HEAD..origin/{base_branch}
   

   Record the output as {divergent_commits}.

6. Invoke baseline-reconcile to obtain a deterministic overlap predicate. The script runs git merge-tree against HEAD and origin/{base_branch} and returns conflict_count — the number of files where the three-way merge would conflict. This is the structural "overlap" signal: conflict_count == 0 means the upstream commits and the worktree's in-flight changes touch disjoint sets of files, so absorbing the upstream tip into the baseline metadata is safe without any working-tree mutation. --no-emit suppresses Q-Gate finding emission — phase-5-execute self-absorption is the wrong place to surface refine-time findings:

   python3 .plan/execute-script.py plan-marshall:workflow-integration-git:git-workflow \
     baseline-reconcile --plan-id {plan_id} --no-emit
   

   Parse conflict_count, upstream_commit_count, and upstream_commits from the returned TOON.

7. Self-absorption branch — conflict_count == 0 (zero-overlap case): the upstream tip can be absorbed into the baseline metadata without re-authoring the request, the outline, or any task. Persist the new worktree_sha (the current HEAD sha after the fetch — unchanged, but recorded for audit) and the new main_sha (the resolved origin/{base_branch} sha) into status.metadata via a single fused manage-status metadata --set call:

   git -C {worktree_path} rev-parse HEAD
   

   Capture as {worktree_sha}.

   git -C {worktree_path} rev-parse origin/{base_branch}
   

   Capture as {main_sha}. Then write both keys:

   python3 .plan/execute-script.py plan-marshall:manage-status:manage-status metadata \
     --plan-id {plan_id} --set --field worktree_sha --value {worktree_sha}
   python3 .plan/execute-script.py plan-marshall:manage-status:manage-status metadata \
     --plan-id {plan_id} --set --field main_sha --value {main_sha}
   

   Emit exactly ONE decision-log entry naming the absorbed commits — the entry is the audit trail that ties the new metadata to the specific upstream commits that were absorbed:

   python3 .plan/execute-script.py plan-marshall:manage-logging:manage-logging \
     decision --plan-id {plan_id} --level INFO \
     --message "(plan-marshall:phase-5-execute:self-absorb) Absorbed {upstream_commit_count} upstream commits with zero overlap: {divergent_commits}"
   

   Log the work-log [STATUS] line for grep-ability:

   python3 .plan/execute-script.py plan-marshall:manage-logging:manage-logging \
     work --plan-id {plan_id} --level INFO \
     --message "[STATUS] (plan-marshall:phase-5-execute) Self-absorbed zero-overlap drift: {upstream_commit_count} commits, new main_sha={main_sha}"
   

   Then continue the task loop — no return to orchestrator, no dispatch to phase-2-refine, no architecture reload, no source-premise verification, no Q-Gate. Self-absorption is metadata-only: the request narrative, solution outline, task list, and confidence score remain valid because the upstream commits touched no overlapping files. Proceed to Step 4.

8. Drift contract — conflict_count > 0 (non-zero-overlap case): the upstream commits touch files that overlap with the worktree's in-flight changes. ABORT the phase fail-loud — re-authoring is required and only refine's iterate-to-confidence loop can absorb the overlap correctly. Return the structured drift TOON for the orchestrator's drift-recovery branch to act on (see plan-marshall:plan-marshall/workflow/execution.md § "Baseline drift recovery (non-zero overlap)"):

   status: error
   error: baseline_drift
   divergent_commits: {divergent_commits}
   upstream_commit_count: {upstream_commit_count}
   conflict_count: {conflict_count}
   display_detail: "baseline drift: {upstream_commit_count} upstream commits"
   

   Log the failure to work-log:

   python3 .plan/execute-script.py plan-marshall:manage-logging:manage-logging \
     work --plan-id {plan_id} --level ERROR \
     --message "[ERROR] (plan-marshall:phase-5-execute) Baseline drift at {worktree_path} with non-zero overlap ({conflict_count} conflicting files) — origin/{base_branch} contains commits not in HEAD: {divergent_commits}. Returning structured drift TOON; orchestrator will re-dispatch phase-2-refine."
   

   Phase-5-execute does NOT perform substantive reconciliation for non-zero overlap. The orchestrator's drift-recovery branch dispatches phase-2-refine, which surfaces the upstream commits as Q-Gate findings and runs the iterate-to-confidence loop to absorb the overlap.

Proceed to Step 4.

Step 4: Log Phase Start and Surface Active Worktree (Once per phase)

At the start of execute or finalize phase, resolve the pending-task count and emit the canonical [STATUS] entry:

python3 .plan/execute-script.py plan-marshall:manage-tasks:manage-tasks list \
  --plan-id {plan_id} --status pending

Parse the row count from the returned tasks_table and substitute it as {N}.

Differentiate first entry from re-entry: Read the persisted phase status from manage-status read to determine whether this is the first time phase-5-execute is being entered or a re-dispatch of an already-in-progress phase. The 5-execute phase row's status is pending on the very first entry and in_progress on every subsequent re-dispatch (the manage-status transition --completed 4-plan call sets it to in_progress).

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

Locate the phases[name=5-execute] row in the returned TOON and read its status field. Then:

• If status == pending (first entry) → emit:

  python3 .plan/execute-script.py plan-marshall:manage-logging:manage-logging \
    work --plan-id {plan_id} --level INFO --message "[STATUS] (plan-marshall:phase-5-execute) Starting execute phase — {N} tasks pending"
  

• If status == in_progress (re-entry; e.g., orchestrator re-dispatched a execution-context after a previous turn ended without completing the queue) → emit:

  python3 .plan/execute-script.py plan-marshall:manage-logging:manage-logging \
    work --plan-id {plan_id} --level INFO --message "[STATUS] (plan-marshall:phase-5-execute) Re-entering execute phase — {N} tasks pending"
  

Both forms emit exactly one [STATUS] line; the wording difference makes it possible to grep for re-entries during retrospective gap analysis.

Surface the active worktree absolute path so it remains visible in model context for every subsequent Edit/Write/Read call. Read the worktree path from status metadata:

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

Extract worktree_path from the output. If present (plan runs in an isolated worktree), emit:

python3 .plan/execute-script.py plan-marshall:manage-logging:manage-logging \
  work --plan-id {plan_id} --level INFO --message "[STATUS] (plan-marshall:phase-5-execute) Active worktree: {worktree_path} — all Edit/Write/Read tool calls MUST target this path. See workflow-integration-git/standards/worktree-handling.md for the full worktree contract."

If worktree_path is absent (plan runs against the main checkout), skip emission. See workflow-integration-git/standards/worktree-handling.md for the worktree-specific application of this rule (path binding, tool cwd flags, Write/Edit-only file authoring, never-edit-main-checkout invariant).

For each task in current phase:

Step 5: Locate Task with Context

python3 .plan/execute-script.py plan-marshall:manage-tasks:manage-tasks next \
  --plan-id {plan_id} \
  --include-context

Returns next task with status pending or in_progress, including embedded goal context (title, body) for immediate use without additional script calls.

Step 6: Execute Steps

For each step in task's steps[] array:

1. Parse the step text
2. Execute the action (delegate if specified) — when delegating to a subagent via Task:, Skill: (prompt-accepting), or execution-context, the subagent inherits the pinned cwd per the Dispatch Protocol section above; pass plan_id as the structured input and optionally embed the reminder header.
3. Mark step complete via manage-tasks:finalize-step

Infeasible deliverable — report, never silently substitute: when a step's declared deliverable turns out to be infeasible during execution — the target cannot be cleanly built as the task specifies (the required surface does not exist, a precondition the deliverable assumed is false, or building the named artifact is structurally impossible as scoped) — the agent MUST report the infeasibility back through the gate and MUST NOT silently substitute a different, weaker deliverable under the same name. Mark the task infeasible — a first-class terminal status — via manage-tasks update --status infeasible, and have the leaf return a structured status: infeasible TOON carrying an infeasibility_reason field naming why the deliverable cannot be built as scoped. The infeasible task flows into the Step 11 (Triage Verification Failure) path, where the dedicated "For infeasible blocks" sub-section routes it to a planning-level gate decision (drop / re-scope / abort) rather than the verification-feedback code-fix loop. 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 prohibited; the infeasibility is a real failure and belongs on the triage surface, not hidden behind a substituted deliverable. infeasible is terminal: it is resolved by the gate decision, never by resuming the same task.

Step 6.5: Scope-Creep Guard (per-task)

After Step 6 completes its file-system changes but BEFORE running task verification (Step 7's finalize-step records "done" only after this guard clears), invoke the deterministic scope-creep helper. The helper computes the residual file-set drift — files modified since the plan was created that are NOT declared in the union of all deliverables' affected_files — and emits a scope_creep_warning finding when the residual cardinality exceeds the configured threshold.

python3 .plan/execute-script.py plan-marshall:phase-5-execute:scope_creep_check \
  check --plan-id {plan_id}

The helper reads plan_creation_sha from references.json, computes git diff --name-only {plan_creation_sha}..HEAD against the worktree, subtracts the union of affected_files from every deliverable, and returns:

status: success
residual_count: N
threshold: T
finding_emitted: true|false
residual_files[N]: [paths]

When finding_emitted: true, the helper has already persisted a scope_creep_warning finding to the Q-Gate findings store via manage-findings qgate add --type scope_creep_warning. The finding flows into the Step 11 triage loop alongside other verify findings (same resolution path: FIX / SUPPRESS / ACCEPT). No additional surface action required here — the standard triage loop handles it.

Threshold configuration: default is 5; override via phase_5.scope_creep_threshold in marshal.json's plan-scoped config. Set to 0 to disable the guard entirely.

Step 7: Mark Step Complete

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

Step 8: Log Task Completion

After each task completes, the canonical [OUTCOME] work-log line is emitted inside manage-tasks finalize-step — see manage-tasks/SKILL.md § "Script-Level [OUTCOME] Emission" for the contract. The script fires exactly one [OUTCOME] (plan-marshall:phase-5-execute) Completed TASK-NNN: {title} ({M} steps) line on the task-closing finalize call. Skills MUST NOT emit a manual [OUTCOME] line here — duplicating the script-level guard creates double entries; the line is lost whenever an execution-context is re-dispatched and the original agent's working context is discarded before its caller-side [OUTCOME] can fire, which is exactly why the emission was moved into the script.

Immediately after the script-emitted [OUTCOME] line, emit one [ARTIFACT] work-log entry per file the task changed by diffing the task-start SHA (recorded at in_progress transition as task_start_sha) against the current HEAD. See standards/workflow.md § Artifact Emission at Task Completion for the authoritative procedure, status-code mapping, and rename-handling rule. The artifact entries use a deliberate three-segment caller prefix (plan-marshall:phase-5-execute:{task_number}) — a documented exception to the usual two-segment (bundle:skill) convention in manage-logging/standards/log-format.md. Emit nothing when the diff is empty. This step precedes manage-tasks next so the audit trail for each task is flushed before the orchestrator advances.

Step 8b: Persist Per-Task Subagent Usage to Accumulator

Applies when: the task was executed by dispatching to a Task agent / execute-task Skill that returned a <usage> tag. Inline tasks (or task agents that produced no <usage> tag) skip this step.

Persist the agent's <usage> totals to the on-disk per-phase accumulator so manage-metrics phase-boundary can read them at end-of-phase, regardless of whether the model context survives until the next orchestrator turn:

python3 .plan/execute-script.py plan-marshall:manage-metrics:manage-metrics accumulate-agent-usage \
  --plan-id {plan_id} --phase 5-execute \
  --total-tokens {total_tokens} --tool-uses {tool_uses} --duration-ms {duration_ms}

Replace the placeholders with the integers parsed from the dispatched agent's <usage>...</usage> block. The canonical token key inside the <usage> block is total_tokens — the same integer forwarded to --total-tokens — so emitters use the canonical key the manage-metrics enrich parser prefers. The script reads .plan/plans/{plan_id}/work/metrics-accumulator-5-execute.toon (initialising it on first call), sums in the supplied values, increments samples, and writes the file back. The on-disk file is the only source of truth — do NOT also keep a parallel tally in model context. See manage-metrics/standards/data-format.md § "Per-Phase Subagent Accumulator" for the file schema.

The orchestrator's phase-boundary call in workflow/execution.md (recorded at end of execute) reads this accumulator as a fallback when its --total-tokens / --tool-uses / --duration-ms flags are omitted. Inline tasks contribute nothing — manage-metrics enrich (run by phase-6-finalize:default:record-metrics) sweeps the transcript for any subagent <usage> tags whose timestamp falls inside the 5-execute window and adds them to the per-phase subagent_* columns of the metrics report as a post-hoc safety net.

Step 8c: Record Per-Step Execution Outcome to the Manifest

Applies to: every Phase 5 verification step the envelope dispatches — the per-task default:verify:{canonical} built-in steps, the Step 11b final quality sweep, and each external (project: / bundle:skill) verification step. This is the consuming side of the record-step contract published by manage-execution-manifest (see that skill's Producers table — phase-5-execute is named as a record-step producer).

After a verification step settles (its build/check completes with a known outcome), append one execution-log row to the manifest so per-step execution metadata is loggable per-plan deterministically, rather than relying on the fragile orchestrator <usage>-forwarding boundary call alone:

python3 .plan/execute-script.py plan-marshall:manage-execution-manifest:manage-execution-manifest record-step \
  --plan-id {plan_id} --step-id {step_id} --phase 5-execute --outcome {executed|skipped|error} \
  --total-tokens {total_tokens} --tool-uses {tool_uses} --duration-ms {duration_ms}

See manage-execution-manifest Canonical invocations → record-step for the authoritative argument surface. Contract:

• --phase is always 5-execute in this phase; --step-id is the verification step ID (e.g. verify:quality-gate, verify:module-tests, verify:coverage, or an external step's notation).
• --outcome is executed when the step ran, skipped when a skip rule fired (e.g. the Step 11b skip when verification_steps is empty, or the Step 10b documentation-only / empty-list skip), and error when the step's build/check exited non-zero (recorded BEFORE the Step 11/11b triage_required return so the failed attempt is on the execution log).
• The token-attribution triple (--total-tokens / --tool-uses / --duration-ms) is the per-step cost; supply the integers parsed from the dispatched agent's <usage> block when one is available, and 0 for inline build invocations that carry no <usage> tag (a skipped step legitimately reports zeros). These are the SAME integers forwarded to the Step 8b accumulate-agent-usage call — Step 8b sums them into the per-phase accumulator that fills the total_tokens column, while Step 8c records the per-step breakdown; the two are complementary, not redundant.
• The manifest MUST already exist (composed by phase-4-plan Step 8b); record-step returns file_not_found otherwise. The append is atomic and one decision-log line is emitted per record.

Exec-blind contract: Phase 5 exec token counts MUST be recorded on EVERY plan. The combination of Step 8b (accumulate-agent-usage → fills the per-phase total_tokens column) and the orchestrator's end-of-execute phase-boundary call (which always fires at the 5-execute → 6-finalize transition, reading the accumulator as its fallback) guarantees the phase-5 total_tokens in metrics.toon is non-zero — closing the historical exec-blind path where phase-5 total_tokens==0. Step 8c's per-step execution_log[] rows are the auditable per-step breakdown behind that aggregate. Because the phase-boundary record is reached unconditionally at the transition (Step 12 → Phase Transition), there is no plan path that skips the closing-phase token record.

Step 9: Independent Change Verification

Applies to: implementation and module_testing profile tasks only. Skip this step for verification profile tasks.

After task completion but before committing, independently verify that the task agent produced genuine results rather than trusting self-reports. Any subagent dispatch made during this step (e.g., a follow-up Task invocation) inherits the pinned cwd per the Dispatch Protocol section above and resolves .plan/ cwd-relatively.

9a. File-change invariant: Verify that at least one file was modified in the worktree. Run in the worktree directory (or main checkout if no worktree):

git -C {worktree_path} diff --name-only HEAD

If the diff output is empty (no files changed) for an implementation or module_testing task:

• Mark task blocked with reason no_changes_detected
• Log:

  python3 .plan/execute-script.py plan-marshall:manage-logging:manage-logging \
    work --plan-id {plan_id} --level WARNING --message "[VERIFY] (plan-marshall:phase-5-execute) No file-system changes detected for {task_id} — marking blocked"
  

• Skip Steps 9b and 9c, proceed to Step 11 (Triage)

9b. Obfuscation spot-check (conditional): When the task's verification criteria include checking for absence of a specific token (e.g., "zero grep hits for --body"), grep the modified files for common obfuscation patterns around that token:

• String concatenation splitting the token (e.g., '--' + 'body', "--" + "body")
• Variable assignment that reconstructs the token from parts

If any obfuscation pattern is found:

• Log each hit:

  python3 .plan/execute-script.py plan-marshall:manage-logging:manage-logging \
    work --plan-id {plan_id} --level WARNING --message "[VERIFY] (plan-marshall:phase-5-execute) Obfuscation pattern detected in {file}: {pattern} — manual review recommended"
  

• Do NOT auto-block (false positives are possible) — flag for human review only

9c. Verification cross-check: Re-execute the task's verification.commands independently and compare the exit code against what the agent reported:

# Run the same verification command the agent claims to have passed
{verification_command}

If the agent reported verification.passed: true but the independent run returns a non-zero exit code:

• Mark task blocked with reason verification_mismatch
• Log:

  python3 .plan/execute-script.py plan-marshall:manage-logging:manage-logging \
    work --plan-id {plan_id} --level WARNING --message "[VERIFY] (plan-marshall:phase-5-execute) Verification mismatch for {task_id}: agent reported pass but independent run failed — marking blocked"
  

• Proceed to Step 11 (Triage)

If independent verification also passes, continue to Step 10.

Step 10: Per-Deliverable Chain-Tail (Commit + Focused Build)

Step 10 fires at the per-deliverable chain-tail point — the moment all tasks for the just-completed deliverable are done. Two independent concerns hang off this point: the (unconditional) per-deliverable commit and the focused per-deliverable build (gated on per_deliverable_build). Both evaluate the same chain-tail predicate; the commit runs first, then the build.

Chain-tail predicate: Does any other pending/in-progress task have depends_on pointing to the just-completed task?

• YES → a downstream task still needs to run → this is NOT the chain tail → skip both the commit and the focused build, proceed to Step 11.
• NO → all tasks for this deliverable are done → this IS the chain tail → run the commit (Step 10a) then the focused build (Step 10b).

Step 10a: Per-Deliverable Commit

The per-deliverable commit fires UNCONDITIONALLY at every chain tail — regardless of which tier ran the preceding build. Whether the deliverable's verification ran as an inline per_task build (reaped synchronously by the leaf) or as an orchestrator-tier build detached through the await-long-running seam, the commit obligation is identical: it is owner-independent. The build's tier decides who runs and reaps the build, never whether the commit fires. The commit consumes only the tier-agnostic kind=build freshness stamp written at the executor dispatch boundary (see ../manage-change-ledger/SKILL.md), which is present for the detached orchestrator build exactly as for an inline per-task build.

Commit-ownership contract: the phase-5-execute envelope is the SOLE owner of every per-deliverable commit — Step 10a is the one place a deliverable's changes are committed. A leaf (execute-task load) returning done while its edits sit uncommitted past the deliverable's chain tail is a contract violation, which the clean-tree post-condition at the 5-execute → 6-finalize transition converts into a worktree_dirty_at_boundary refusal (see manage-status cmd_transition). The boundary settlement bookkeeping documented in ../plan-marshall/workflow/execution.md § "Execute Phase Completion" is the RECOVERY for that refusal, not the norm — it attributes each dirty path back to its owning deliverable, cuts one settlement commit per implicated deliverable (each with its own Step 10a-shaped kind=change ledger append), and halts loudly on any path that maps to no deliverable; commits belong here, at the chain tail. The initial-envelope backgrounding concern is closed by the leaf's live tier re-resolve, which runs before EVERY whole-tree verification step it would execute — so the leaf's per_task-only runnable slice and the orchestrator's await-long-running ownership of orchestrator-tier steps apply uniformly from the very first envelope dispatch onward, with no step for a leaf to background-and-lose. The compose-time stamp (manage-execution-manifest._stamp_phase_5_step_execution_tier) is total over verification_steps but advisory — it records the expected tier for planning, and the live resolve is what routes.

1. Commit:

   Skill: plan-marshall:workflow-integration-git
   Parameters:
     - message: conventional commit derived from task title
     - push: false
     - create-pr: false
   

2. Log commit outcome:

   python3 .plan/execute-script.py plan-marshall:manage-logging:manage-logging \
     work --plan-id {plan_id} --level INFO --message "[OUTCOME] (plan-marshall:phase-5-execute) Per-deliverable commit: {task_id} ({commit_hash})"
   

3. Resolve the git-sourced commit_sha — capture the SHA the commit just produced:

   git -C {worktree_path} rev-parse HEAD
   

   Record the output as {commit_sha}.

4. Source the changed paths from git — the kind=change entry's changed_paths MUST be git-sourced (NOT self-computed from the deliverable's declared affected_files). Enumerate the paths the commit touched:

   git -C {worktree_path} diff-tree --no-commit-id --name-only -r {commit_sha}
   

   Record the newline-separated output as {changed_paths} (join into a comma-separated list for the verb's --changed-paths argument).

5. Append the kind=change ledger entry — record the per-deliverable commit transition to the unified change-ledger. The verb stores the supplied changed_paths list verbatim; do NOT inline-copy the ledger API details — see ../manage-change-ledger/SKILL.md § "Canonical invocations" → append (kind=change) for the authoritative argument surface:

   python3 .plan/execute-script.py plan-marshall:manage-change-ledger:manage-change-ledger append \
     --kind change --deliverable-id {deliverable} --commit-sha {commit_sha} --changed-paths {changed_paths} \
     --task-id {task_id}
   

Step 10b: Focused Per-Deliverable Build (module-scoped canonical-verify steps)

The mid-execute per-deliverable build is focused by design: it runs the per_deliverable_build list of default:verify:{canonical} step IDs module-scoped over the changed module(s), never a whole-tree sweep. The whole-tree quality sweep stays once at end-of-phase (Step 11b/11c) and is never repeated mid-execute — whole-tree gates (e.g. integration-tests, e2e) are NOT permitted in per_deliverable_build; they live in verification_steps. This step fires at the chain tail, after the Step 10a commit decision.

1. Read the per_deliverable_build list from the plan-scoped config:

   python3 .plan/execute-script.py plan-marshall:manage-config:manage-config \
     plan phase-5-execute get --field per_deliverable_build --audit-plan-id {plan_id}
   

   The value is a LIST of default:verify:{canonical} step IDs (the same vocabulary as verification_steps); the default is [default:verify:compile, default:verify:module-tests], which reproduces today's compile + scoped-test behaviour. The empty list [] means "no per-deliverable build" (the end-of-phase sweep is the only build). The vocabulary and its list-membership validator are owned by plan-marshall:manage-config (per_deliverable_build field) — do NOT restate the validation here; this step consumes the resolved list.

2. Empty list [] → skip the per-deliverable build entirely. The end-of-phase Step 11b/11c sweep is the only build. Log the decision and proceed to Step 11:

   python3 .plan/execute-script.py plan-marshall:manage-logging:manage-logging \
     decision --plan-id {plan_id} --level INFO \
     --message "(plan-marshall:phase-5-execute) per_deliverable_build=[] — skipping focused build for deliverable {deliverable}; end-of-phase sweep is the only build"
   

3. For each default:verify:{canonical} entry in the list, invoke the canonical-verify step module-scoped over the changed module(s): resolve architecture resolve --command {canonical} --module {changed_module} and run the resolved executable, honouring the returned execution_tier / bash_timeout_seconds. Do NOT restate the resolution/execution-tier logic here — see standards/canonical_verify.md § "Module-scoped vs whole-tree invocation" and § "Workflow" for the authoritative step body (module-scoped invocation supplies --module {changed_module}; the unresolved-canonical skip and the tier hand-off apply identically). After each build call, inspect the result TOON — read status and the errors[] rows, not the harness exit code (the build wrapper exits 0 even on failure).

   Documentation-only short-circuit: a changed-path set with no buildable module yields no module-scoped run (the canonical does not resolve / there is no changed module). Log:

   python3 .plan/execute-script.py plan-marshall:manage-logging:manage-logging \
     decision --plan-id {plan_id} --level INFO \
     --message "(plan-marshall:phase-5-execute) No buildable module for deliverable {deliverable} — no module-scoped build runs for this deliverable"
   

4. On non-zero exit — route the failure through the existing Step 11 per-task triage path: persist each failing finding to the Q-Gate store (manage-findings qgate add) and return the triage_required signal to the orchestrator. Do NOT invent a new triage surface — reuse the Step 11 contract verbatim (producer=build-runner, finding_type=test-failure).

Guideline — file-relocation lint gap. When a deliverable RELOCATES a file (moves it from one path/module to another), the focused module-scoped per-deliverable build above can MISS lint/structural findings that only surface in a whole-tree sweep: a moved file implicates BOTH its old and new module, but a module-scoped gate sees only one of them. Sequence a relocation deliverable so the whole-tree verify runs AFTER the move completes, and NEVER skip the orchestrator whole-tree gate (Step 11c, Execute-Exit Verify Gate) on a plan that relocates files. The module-scoped per-deliverable build is NOT a substitute for the whole-tree gate on relocations.

Step 11: Triage Verification Failure

Applies when:

• A profile=verification task completes with verification.passed: false / next_action: requires_triage, OR
• Step 9 marked a task blocked with reason no_changes_detected or verification_mismatch, OR
• A task was marked infeasible per Step 6 ("Infeasible deliverable — report, never silently substitute") — the leaf returned status: infeasible with an infeasibility_reason; this routes to the dedicated "For infeasible blocks" sub-section below, NOT the verification-feedback code-fix loop, OR
• The in-context execute-task load returned escalate_ask: true with prompt_options[] — a scope-deviation and/or smart_and_ask gate fired and the task is left not-done; this routes to the dedicated "Scope-deviation / smart_and_ask escalation yield" sub-section below, NOT the verification-feedback code-fix loop

The per-finding LLM core (FIX / SUPPRESS / ACCEPT / AskUserQuestion decisions over the failing findings) is owned by ../plan-marshall/workflow/verification-feedback.md. This per-task body is a leaf and does NOT dispatch it — the leaf persists the findings and returns a triage_required signal; the main-context orchestrator dispatches verification-feedback under --phase phase-5-execute --role verification-feedback with producer=build-runner (see ../plan-marshall/workflow/execution.md and the canonical contract in ref-workflow-architecture/standards/agents.md).

Scope-deviation / smart_and_ask escalation yield (escalate_ask)

Applies when the in-context execute-task load returned escalate_ask: true with prompt_options[] — a scope-deviation escalation (Handle Verification Results) and/or a smart_and_ask compatibility gate fired, and the leaf left the task not-done. This is a planning-level operator decision, NOT a fixable code failure — it is explicitly NOT routed through verification-feedback (there is no test/lint/build finding to FIX, SUPPRESS, or ACCEPT). The canonical deviation taxonomy, the three-option shape, and the prohibited "log-and-continue" anti-pattern live in ref-workflow-architecture/standards/scope-deviation-escalation.md.

Because this per-task body is a leaf, it CANNOT fire AskUserQuestion and CANNOT dispatch. It collects the prompt_options[] from the returned escalate_ask — batching across every deviation / smart_and_ask gate that fired anywhere in this envelope into ONE prompt_options[] list — and yields an escalate_ask envelope to the main-context orchestrator as a new TASK-boundary yield reason alongside budget_yield / blocked / infeasible. The leaf returns a wrapped terminal payload and stops; the orchestrator fires ONE batched AskUserQuestion and applies each option's side effect post-return (see ../plan-marshall/workflow/execution.md § "Post-return escalate_ask batched deviation dispatch"). The leaf's return payload:

status: escalate_ask
display_detail: "{task_number} escalate_ask: {N} deviation prompt(s)"
escalate_ask: true
plan_id: {plan_id}
prompt_options[N]{id,question,header,options,recommended}:
  ...

The orchestrator resolves the prompt and re-dispatches phase-5-execute with each resolution baked in: Hold the line resumes the fix loop with the requirement intact; Accept with rationale persists the rationale to decision.log and the PR body; Split into follow-up plan seeds a successor lesson. The leaf performs none of the operator-facing interaction.

Pre-triage scope cross-reference

Before composing the triage dispatch, classify the failing file paths against the plan's live footprint (derived on demand from the worktree — {base}...HEAD ∪ porcelain). The cross-reference is deterministic — a small Python helper that subtracts the footprint from the union of error paths and returns a exclusively_out_of_scope flag:

python3 .plan/execute-script.py plan-marshall:phase-5-execute:verify_failure_scope \
  classify --plan-id {plan_id} --error-paths {comma_separated_paths}

The script derives the live footprint from the worktree (reading references.json only for the base ref), classifies each error path, and returns:

status: success
total: N
in_scope_count: I
out_of_scope_count: O
exclusively_out_of_scope: true|false
out_of_scope_paths[O]: [paths]

When exclusively_out_of_scope: true: the failing tests originate ENTIRELY outside the plan's declared scope (a sibling refactor on the same branch surfaced unrelated breakage). The [BLOCKED] triage message MUST include the distinction (e.g., "All N failures originate outside plan scope: {paths}") and the AskUserQuestion offered to the user MUST present "Stash foreign files and re-verify" as the default recommended action, alongside the standard FIX / SUPPRESS / ACCEPT options.

When exclusively_out_of_scope: false (the common case): proceed to the standard triage dispatch below without the foreign-failure annotation. The classification is informational only.

Planned-failure exception (breaking-refactor task split)

Applies before the standard triage branches below. When a task with profile: implementation produces a verification failure and a downstream task with profile: module_testing and explicit depends_on: [TASK-{current_task_number}] exists, the dispatcher MAY proceed to the dependent task without flagging the failure as an error — this is the only case where "tests fail" is the planned outcome of the implementation step.

Boundary conditions (ALL must hold; if any fails, fall through to the standard triage branches below):

1. The downstream task's profile is module_testing AND its deliverable matches the current task's deliverable AND its description enumerates the pre-existing tests being rewritten.
2. The downstream task has explicit depends_on: [TASK-{current_task_number}] linkage declared at planning time. A downstream task that happens to run later without a depends_on edge does NOT qualify.
3. The set of failing tests reported by the implementation task's verification command is a subset of the tests enumerated in the downstream task's description. New failures (tests not on the list) are real regressions and MUST fall through to standard triage.

When all three boundary conditions hold, log the planned-failure decision, mark the implementation task as done (not blocked), and proceed to the next task in the queue (which will be the test-contract task by depends_on ordering):

python3 .plan/execute-script.py plan-marshall:manage-logging:manage-logging \
  decision --plan-id {plan_id} --level INFO \
  --message "(plan-marshall:phase-5-execute) Planned-failure exception applied for {task_id}: verification failed as expected; downstream test-contract task TASK-{downstream_number} will rewrite the affected tests"

After the test-contract task completes, the standard verification path resumes — the test-contract task itself MUST produce a green test run; if it does not, that is a real failure and goes through standard triage.

Rationale and boundary documentation: see ../phase-4-plan/standards/breaking-refactor-task-split.md for the full contract spanning phase-4-plan task allocation and this phase-5-execute exception.

For infeasible blocks: The leaf returned status: infeasible with an infeasibility_reason — the declared deliverable cannot be built as scoped. This is a planning decision, NOT a fixable code failure: it is explicitly NOT routed through verification-feedback (there is no test/lint/build finding to FIX, SUPPRESS, or ACCEPT). The orchestrator's handling:

1. Mark the task infeasible in the ledger (the leaf has already done this via manage-tasks update --status infeasible; the orchestrator confirms the terminal status is recorded). Because infeasible is a terminal state the loop-exit guard does NOT count, it never re-enters the task loop.
2. Raise AskUserQuestion surfacing the infeasibility_reason and offering exactly three gate-level options:
   • (a) Drop the task — accept that this deliverable will not be built; remove it from the active scope and continue with the remaining queue.
   • (b) Re-scope via a new task — create a replacement task with a buildable, value-preserving deliverable that supersedes the infeasible one; the new task enters the queue.
   • (c) Abort the plan — the infeasible deliverable is load-bearing for the whole plan; stop and return control for re-planning.
3. Record the chosen option to decision.log and act on it. Do NOT dispatch verification-feedback on this path — the AskUserQuestion gate is the resolution mechanism.

For no_changes_detected blocks: The implementation task produced no file changes. Triage options:

• RETRY → reset task to pending for re-execution
• FAIL → mark task failed with outcome no_changes_detected, log, continue

For verification_mismatch blocks: The agent claimed verification passed but independent re-run failed. Triage options:

• FIX → create fix task to address the actual verification failure
• RETRY → reset task to pending for re-execution
• FAIL → mark task failed with outcome verification_mismatch, log, continue

For verification task failures (original behavior):

11a: Read verify_iteration counter from task metadata (default: 0).

11b: If verify_iteration >= max_iterations (from phase-5-execute config, default 5) → mark task blocked, log, continue to Step 12.

11c: Persist each failing finding to the Q-Gate findings store (producer-side; the triage dispatch reads from the store by reference):

python3 .plan/execute-script.py plan-marshall:manage-findings:manage-findings \
  qgate add --plan-id {plan_id} --phase 5-execute \
  --source qgate --type test-failure --severity {severity} \
  --title "{finding_message}" --detail "{file}:{line}"

(One qgate add call per finding; the verification task's structured findings[] output drives this loop.)

11d: This per-task body is a leaf — it MUST NOT dispatch verification-feedback itself. After §11c has persisted each failing finding to the Q-Gate store, return a structured terminal payload to the main-context orchestrator and stop; the orchestrator owns the triage dispatch. See ref-workflow-architecture/standards/agents.md for the canonical leaf/dispatch-topology contract.

The leaf's return payload carries the discriminators the orchestrator needs to compose the dispatch:

status: blocked
display_detail: "{task_number} triage_required: {N} verification finding(s)"
triage_required: true
producer: build-runner
finding_type: test-failure
plan_id: {plan_id}

(finding_type: lint-issue for the Step 11b sweep path.) The findings are already in the per-plan store — the orchestrator's verification-feedback dispatch queries them by reference; the leaf does not embed the findings in its return.

The orchestrator-side handling — resolving the verification-feedback target via manage-config effort resolve-target --phase phase-5-execute --role verification-feedback, emitting the [DISPATCH] log line, dispatching verification-feedback (producer=build-runner, caller_phase: phase-5-execute) as a top-level Task: in the main context, and consuming the triage return to drive the §11e branch — lives in ../plan-marshall/workflow/execution.md § "Verification-feedback triage (leaf returned triage_required)". The per-finding triage core (FIX / SUPPRESS / ACCEPT / AskUserQuestion, smart grouping, overflow, and the Scope-Deviation Escalation guard) is owned by ../plan-marshall/workflow/triage.md; the dispatch is by-reference (the subagent queries the store as its first workflow step).

11e (orchestrator-owned): The orchestrator inspects the verification-feedback return per ../plan-marshall/workflow/execution.md:

• If fix_tasks_created > 0 → increment verify_iteration in task metadata, reset the verification task to pending, continue the execution loop (fix tasks will execute before the re-queued verification task via depends_on).
• If fix_tasks_created == 0 AND overflow_deferred == 0 → mark the verification task complete (all findings suppressed / accepted / taken_into_account).
• If overflow_deferred > 0 → leave the verification task pending; the orchestrator re-fires the triage dispatch on the next phase-5-execute entry (the iteration cap is unchanged).

Step 11b: Final Quality Sweep (After All Tasks)

After every task in the phase has completed (and Step 11 has resolved any per-task verification failures), but before Step 12 transitions the phase, run one canonical quality-gate invocation as a final sweep — but ONLY when BOTH (a) phase_5.verification_steps (cached from Step 2) is non-empty AND (b) the live footprint requires a build. The whole-tree quality sweep is the ONE footprint-blind phase-5 build surface (Steps 10b and 11c already derive the live footprint on demand and skip when no buildable module is present — see the Footprint gating note under the Built-in Step Dispatch Table); condition (b) closes that gap by consulting the existing manage-config build-decision verb (a thin wrapper over extension_base.should_execute_build, the same build.map ∩ footprint authority phase-6 uses — NOT a new verb and NOT a blanket phase-5 short-circuit).

Skip rule (empty list): If phase_5.verification_steps is empty (e.g., docs-only plans where the manifest composer dropped all verification steps), skip this step entirely — no final sweep, no log, proceed directly to Step 12.

Footprint-gate consult (fire condition (b)): When phase_5.verification_steps is non-empty, consult the build-necessity verdict for quality-gate BEFORE resolving or running the sweep:

python3 .plan/execute-script.py plan-marshall:manage-config:manage-config build-decision \
  --command quality-gate --plan-id {plan_id}

Parse decision and reason from the returned TOON. When decision == not_necessary (the live footprint is empty or touches no registered build.map glob — the docs-only nifi scenario), SKIP the sweep: do NOT resolve, do NOT run any build. Emit a decision-log line naming the footprint reason, record the step as skipped, and proceed directly to Step 11c:

python3 .plan/execute-script.py plan-marshall:manage-logging:manage-logging \
  decision --plan-id {plan_id} --level INFO \
  --message "(plan-marshall:phase-5-execute) Final quality sweep skipped — build-decision quality-gate returned not_necessary: {reason}"

python3 .plan/execute-script.py plan-marshall:manage-execution-manifest:manage-execution-manifest record-step \
  --plan-id {plan_id} --step-id verify:quality-gate --phase 5-execute --outcome skipped \
  --total-tokens 0 --tool-uses 0 --duration-ms 0

When decision != not_necessary (the live footprint touches a buildable glob), the sweep fires — proceed to run exactly one quality sweep, regardless of whether quality-gate already appears in the list:

1. Resolve the canonical quality-gate build command via the architecture API:

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

2. Branch on the tier the resolve in step 1 just returned. That live execution_tier is the routing authority (see Per-step execution_tier above); the manifest's verify:quality-gate stamp is the advisory expectation, not the decision. When the live tier == orchestrator, the sweep is NOT in the leaf's runnable slice — do NOT run it; return the orchestrator-tier yield signal (status: blocked, voluntary_checkpoint) naming the sweep so the orchestrator runs it via await-long-running. Only when the live tier == per_task does the leaf execute the returned executable inline. On non-zero exit, persist the failures to the Q-Gate findings store (manage-findings qgate add --type lint-issue …) and return the triage_required signal to the orchestrator with producer=build-runner and finding_type=lint-issue — same leaf-returns-signal shape as Step 11d above, only the finding type changes. The leaf does NOT dispatch verification-feedback itself; the orchestrator owns the dispatch (see ../plan-marshall/workflow/execution.md § "Verification-feedback triage (leaf returned triage_required)") and drives the same fix-task / suppress / accept branch (Step 11e). After the orchestrator's triage resolves, the sweep is NOT re-run — Step 11b runs at most once per phase entry.

3. Log the outcome:

   python3 .plan/execute-script.py plan-marshall:manage-logging:manage-logging \
     work --plan-id {plan_id} --level INFO \
     --message "[STATUS] (plan-marshall:phase-5-execute) Final quality sweep: {pass|fail}"
   

This step is the single source of "did the phase end clean?" — it appends the canonical quality-gate once after all task-level verification has settled, providing a stable end-of-phase quality signal. Only the manifest's verification_steps list controls whether it fires; per-doc skip logic has been removed in favor of this manifest-driven gate (the single parameterized canonical_verify.md step carries no embedded skip logic).

Step 11c: Execute-Exit Verify Gate (One verify per Affected Bundle)

Per-task verification runs each task's pre-stamped verification.commands — the derived ladder the deriver wrote at compose time (per-class compile / test-compile / module-tests, scoped to the deliverable's changed module). That ladder is the per-deliverable gate. The execute-exit verify gate is the single end-of-phase whole-bundle verify that fires exactly once per affected bundle at the FINAL state of execute, after every deliverable has settled — replacing both the per-task full suites and the retired holistic verification tasks. Running it once at the queue tail, rather than once per task, is the asymmetric-cost collapse: a whole-bundle verify re-runs the same suite no matter how many deliverables touched the bundle, so one run at the final state covers them all.

This gate runs after Step 11b's quality sweep and before Step 12.

1. Resolve the affected bundle set from the live plan footprint. Derive the footprint ({base}...HEAD ∪ porcelain) and map each path to its owning module:

   python3 .plan/execute-script.py plan-marshall:manage-architecture:architecture \
     which-module --path {changed_path}
   

   Collect the distinct module values into the affected-bundle set. The build-class classification and the derived-command mapping are owned centrally by the deriver — see ../manage-architecture/SKILL.md (derive-verification subcommand) and ../manage-architecture/standards/resolve-command.md for the build_map consumer API. Do NOT inline-copy the path heuristics or the build_class → command table here; the central standards are the single source of truth.

2. Run exactly one verify per affected bundle — for each distinct {bundle} in the affected set, resolve and execute the whole-bundle verify:

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

   Branch on the execution_tier this resolve just returned — the live tier is the routing authority (the manifest's verify:module-tests stamp, which backs the verify canonical, is the advisory expectation only). For tier=per_task run the build inline with timeout: bash_timeout_seconds * 1000; for tier=orchestrator the step is NOT in the leaf's runnable slice — return control to the orchestrator to run the long build via await-long-running (do NOT run it inline, do NOT background it). After each build call the leaf runs, inspect the result TOON — read status and the errors[] rows, not the harness exit code.

3. On non-zero exit — route the failure through the same leaf-returns-signal path as Step 11b: persist each failing finding to the Q-Gate store (manage-findings qgate add --type lint-issue …) and return the triage_required signal to the orchestrator with producer=build-runner and finding_type=lint-issue. The leaf does NOT dispatch verification-feedback itself. The gate runs at most once per phase entry.

4. Log the outcome per affected bundle:

   python3 .plan/execute-script.py plan-marshall:manage-logging:manage-logging \
     work --plan-id {plan_id} --level INFO \
     --message "[STATUS] (plan-marshall:phase-5-execute) Execute-exit verify for {bundle}: {pass|fail}"
   

Skip rule: when the affected-bundle set is empty (no buildable footprint — e.g., a documentation-only plan whose changed paths resolve to no module), skip the gate entirely. The deriver already stamped zero Python commands on each such deliverable's per-task ladder, so there is no whole-bundle verify to run.

Step 12: Next Task or Phase

• If more tasks in phase → Continue to next task
• If phase complete → run Step 12a (Pending-tasks transition guard) below, then log phase outcome and auto-transition to next phase
• If all phases complete → Mark plan complete

Step 12a: Pending-tasks transition guard

Before invoking manage-status transition --completed 5-execute (see Phase Transition section below), refuse to transition when any pending tasks remain AND when the on-disk worktree has not been observed by a fresh verify run. Pending-queue emptiness is necessary but not sufficient: a task that was marked done against a prior code state still leaves the queue empty, yet the codebase the orchestrator is about to ship has never been verified end-to-end. The canonical failure mode for this gap: loop-exit-guard returns pending_count: 0 while the most recent verify predated the last source-file mutation, and CI fails on the pushed commit. Step 12a therefore enforces two co-equal gates: (a) manage-tasks next only surfaces the head of the queue, so a null next does NOT prove the queue is empty when downstream tasks are still in pending — fix tasks created by Step 11 triage commonly land here, and a premature transition silently abandons them; (b) the worktree state itself must be fresh with respect to the change-ledger — a successful kind=build entry must exist whose worktree_sha matches the current working-tree currency hash.

Script-level enforcement: the authoritative pending-count check is python3 .plan/execute-script.py plan-marshall:manage-tasks:manage-tasks loop-exit-guard --plan-id {plan_id} — see manage-tasks/SKILL.md § "Loop-Exit Guard". status: continue (with pending_count > 0 and pending_ids) forces the orchestrator to re-dispatch the execution-context; status: success (with pending_count: 0) is the precondition for recording the clean_exit_queue_empty termination cause via the manage-metrics record-dispatch-boundary verb. The list-based check below remains documented for backwards compatibility with existing callers — both forms read the same on-disk state, but loop-exit-guard is the canonical surface and the verb the orchestrator MUST consult.

Worktree-state freshness enforcement: the authoritative freshness check is python3 .plan/execute-script.py plan-marshall:manage-tasks:manage-tasks pre-commit-verify-freshness --plan-id {plan_id} — see manage-tasks/SKILL.md § "Pre-Commit Verify Freshness". The script recomputes the current working-tree currency hash (worktree_sha) and scans the unified change-ledger for a kind=build entry with exit_code == 0 whose worktree_sha matches. The query is tier-agnostic and build-tool-agnostic — it filters on kind, exit_code, and worktree_sha only, never notation or plan_id, so a Maven/Gradle/npm build or an orchestrator-driven global-tier build satisfies the gate exactly as a plan-scoped pyproject build does. See marketplace/bundles/plan-marshall/skills/manage-change-ledger/SKILL.md for the ledger and worktree_sha primitive. The script returns one of three statuses. status: fresh permits transition; status: stale or status: undecidable blocks transition with the same [BLOCKED] log line shape used for the pending-tasks branch. The gate fails closed by design — there is no LLM judgement and no "probably fine" fallback. Pending-queue emptiness and worktree freshness are co-equal gates: both MUST succeed before the phase may transition.

1. Query the pending-task list:

   python3 .plan/execute-script.py plan-marshall:manage-tasks:manage-tasks list \
     --plan-id {plan_id} --status pending
   

2. Parse the row count from the returned tasks_table. If the count is zero, proceed to step 2.5 (freshness check). If non-zero, jump to step 3.

2.5. Run the freshness check:

python3 .plan/execute-script.py plan-marshall:manage-tasks:manage-tasks \
  pre-commit-verify-freshness --plan-id {plan_id}

Parse status from the returned TOON. On status: fresh, proceed to Phase Transition. On status: stale or status: undecidable, log a [BLOCKED] line and abort the transition:

python3 .plan/execute-script.py plan-marshall:manage-logging:manage-logging \
  work --plan-id {plan_id} --level ERROR \
  --message "[BLOCKED] (plan-marshall:phase-5-execute) Worktree state not verified: {reason} (worktree_sha={worktree_sha}, ledger_path={ledger_path}) — refusing to transition 5-execute → 6-finalize. Re-dispatch a verify run, or invoke with --force to override."

Substitute the placeholders with the corresponding fields from the script's TOON output. Each branch omits a different field set: stale omits reason; undecidable carries reason set to one of no_registry (the ledger file is absent or empty) or head_unresolvable (the working-tree sha cannot be computed), and the head_unresolvable sub-case omits worktree_sha and ledger_path. Substitute - for any field absent in the returned TOON. Do NOT call manage-status transition and do NOT auto-continue to finalize. The orchestrator's recovery path is to dispatch a fresh verify run, after which Step 12a is re-entered.

3. If the pending count is non-zero, the phase is NOT complete. Log a [BLOCKED] line and abort the transition:

   python3 .plan/execute-script.py plan-marshall:manage-logging:manage-logging \
     work --plan-id {plan_id} --level ERROR \
     --message "[BLOCKED] (plan-marshall:phase-5-execute) Pending tasks: {ids} — refusing to transition 5-execute → 6-finalize. Re-enter the execute loop to complete pending tasks, or invoke with --force to override."
   

   {ids} is a comma-separated list of TASK-{number} identifiers parsed from the tasks_table. Do NOT call manage-status transition and do NOT auto-continue to finalize.

4. --force escape (mirrors the verification-cap escape in Step 11b): when the orchestrator is invoked with --force, log the override decision, then proceed to Phase Transition. The escape covers both gates — pending tasks left intact AND a non-fresh freshness status. Emit one decision line per gate that the override bypasses:

   python3 .plan/execute-script.py plan-marshall:manage-logging:manage-logging \
     decision --plan-id {plan_id} --level WARNING \
     --message "(plan-marshall:phase-5-execute) Pending-tasks guard overridden via --force — transitioning with {count} pending task(s): {ids}"
   

   python3 .plan/execute-script.py plan-marshall:manage-logging:manage-logging \
     decision --plan-id {plan_id} --level WARNING \
     --message "(plan-marshall:phase-5-execute) Worktree-freshness guard overridden via --force — transitioning with status={status}"
   

   Append reason={reason} to the message body only when status is undecidable; the stale branch does not emit a reason field, so the appended fragment is omitted for that branch. This mirrors the --force escape format in phase-6-finalize/standards/push.md § Freshness precondition.

   The --force escape is a deliberate safety valve for triage-driven aborts (the user has already decided the pending tasks are out-of-scope, or that the stale-freshness signal is being addressed elsewhere) — never invoke it programmatically from inside the loop.

Step 4.5: Capture Session Token (Once per phase)

At phase start, capture the runtime session token so downstream metrics operations can reference it:

python3 .plan/execute-script.py plan-marshall:platform-runtime:platform_runtime \
  session capture --plan-id {plan_id}

On Claude Code the runtime reads the stored session_id from the SessionStart hook. On OpenCode it returns no-op (no platform session id available) — the phase proceeds normally and uses manual --total-tokens for metrics.

Step 13: Log Phase Completion (When phase completes)

Substitute {N} with the count of tasks marked done during this phase entry and {M} with the total task count from the plan, then emit the canonical phase-exit [STATUS] line:

python3 .plan/execute-script.py plan-marshall:manage-logging:manage-logging \
  work --plan-id {plan_id} --level INFO --message "[STATUS] (plan-marshall:phase-5-execute) Execute phase complete — {N}/{M} tasks done"

Add visual separator after END log:

python3 .plan/execute-script.py plan-marshall:manage-logging:manage-logging \
  separator --plan-id {plan_id} --type work

---

Delegation

When checklist items specify delegation, invoke the appropriate agent/skill:

| Checklist Pattern | Delegation | |-------------------|------------| | "Run build" / "maven" / "npm" | See standards/operations.md | | "Delegate to {agent}" | Task: {agent} | | "Load skill: {skill}" | Skill: {skill} | | "Run /command" | SlashCommand: /command |

---

Auto-Continue Behavior

Execute continuously without user prompts except:

• Error blocks progress
• Decision genuinely required
• User explicitly requested confirmation

Do NOT prompt for:

• Phase transitions
• Task transitions
• Routine confirmations

B7 — voluntary_checkpoint no-progress reclassification

After a phase-5-execute dispatch returns and the orchestrator's classification rules in plan-marshall/workflow/execution.md provisionally label the return as termination_cause: voluntary_checkpoint, the orchestrator evaluates the deterministic no-progress predicate:

in_progress_count > 0 AND completed_tasks_delta == 0 AND consumed_tokens > 50000

When all three sub-conditions hold, the orchestrator reclassifies termination_cause from voluntary_checkpoint to error BEFORE invoking record-dispatch-boundary. The reclassification routes the dispatch into the shorter retry-budget / escalation path already coded for error, instead of letting another round of voluntary-checkpoint re-dispatches burn budget on a loop that is not making progress. Plans that DO make progress — even a single task completed (completed_tasks_delta >= 1), or cheap no-op iterations under the 50K-token threshold — keep the voluntary_checkpoint classification and continue along the standard recovery path.

The reclassification is a forensic + control-flow decision: the dispatch is still recorded via record-dispatch-boundary, only the --termination-cause value changes. The full predicate definition, sub-condition resolution rules, and decision-log shape (carrying all three predicate values for forensic reconstruction) live in plan-marshall/workflow/execution.md § "B7 — voluntary_checkpoint no-progress reclassification".

Forbidden: agent-initiated checkpoints

Phase-5-execute MUST drive the task loop to one of three terminal outcomes inside a single dispatch:

1. All pending tasks complete and the phase transitions to 6-finalize.
2. A fatal error captured via the Error Handling section (including the pending-task drift error below).
3. A triage-driven blocked outcome that the skill itself acknowledges via manage-tasks status updates.

Improvising a "progress checkpoint" return is a workflow violation. Specifically, the dispatched agent MUST NOT:

• Emit a "Returning control to orchestrator" / "checkpoint reached" / "partial-completion handoff" line and stop with pending tasks still in the queue.
• Return a TOON payload that summarises "N of M tasks done, please re-dispatch" without one of the three terminal outcomes above.
• Ask the user whether to continue when no genuine decision is required (loop fatigue is not a decision point).
• Return the bare task_complete payload from the in-context execute-task load verbatim while pending tasks in the SAME envelope_id group remain. This is the task_complete_returned_verbatim defect and the direct violation of the one-context-per-phase invariant (§ "Dispatched workflows vs inline steps"): the phase completes in exactly ONE dispatch per envelope group — envelope_count dispatches, NEVER one per task — so a bare task_complete echo collapses the envelope loop into a wasteful per-task re-dispatch. execute-task's next_action: task_complete is a PER-TASK signal THIS envelope loop consumes to advance to the next same-envelope_id task; it is never the envelope's terminal return. The envelope MUST keep looping over its own group until the group is exhausted, then yield via the wrapped budget_yield payload (§ "Deterministic exit clause") — never a bare task_complete.

Agent-initiated re-dispatch is a control-flow drift that can cause [OUTCOME] log coverage gaps — the script-level [OUTCOME] guard in manage-tasks finalize-step closes the audit-trail gap, but the underlying drift also needs to be ruled out at the skill level. The orchestrator (plan-marshall workflows) is the single component allowed to start, re-dispatch, or terminate phase-5-execute; the dispatched agent does not get to vote.

The one legitimate yield — budget_yield: the envelope-group read (see § "Deterministic exit clause (envelope-group read)") yields when the next pending task's envelope_id differs from the assigned group. This is NOT a forbidden checkpoint — it is the plan-time-packed dispatch boundary the bin-packer pre-computed. It is distinguished from a forbidden checkpoint by two MANDATORY observable signals: (1) a budget_yield decision-log entry naming the assigned and next envelope_id, and (2) a wrapped terminal TOON carrying budget_yield: true with tasks_remaining > 0 — never the bare task_complete echo. A yield that omits either signal IS a forbidden checkpoint. The orchestrator classifies the wrapped, logged yield as termination_cause: budget_yield (a legitimate yield) and re-dispatches the next envelope group; a bare task_complete return with pending tasks is still classified as the task_complete_returned_verbatim drift.

Deterministic exit clause (envelope-group read)

The loop's continue-vs-yield decision is pre-computed at plan time by the bin-packer (phase-4-plan's manage-tasks pack-envelopes, deliverable 3) — the executor makes NO runtime cost decision at all. Harness reality (the WHY): a running subagent CANNOT measure its own context-window usage mid-turn — no tool, env var, signal, or API returns "tokens used/remaining" to the model during execution; the <usage> block reaches the caller only AFTER the subagent returns. A runtime remaining_budget therefore has no source and cannot be the comparand. The packing decision is consequently moved entirely to plan time, where every task's predicted_cost_tokens is a known constant: the bin-packer groups tasks (in depends_on order, accumulating cost under per_envelope_budget_tokens) and stamps each with an envelope_id. At runtime the executor only READS that grouping.

The continue-vs-yield clause is therefore a trivial countable check, evaluated in canonical order:

Small-plan short-circuit: If tasks_total <= 2 (read from phase_5.tasks_total in the execution manifest cached at Step 2), the queue is trivially one envelope — continue to the next task until the queue is empty or a terminal outcome fires. (Informational only under plan-time packing; a small plan already packs to a single envelope.)

Final-task long-running-verify short-circuit: If BOTH (a) the current task is the final task in the queue (task_index + 1 == tasks_total) AND (b) its resolved verification command is in the known long-running build set (verify, coverage, quality-gate fully scoped), continue and finish in-dispatch. Re-dispatching at the queue tail to run a long-running build pays the full dispatch overhead for zero scheduling benefit (no subsequent task ever runs). Log the suppression decision via manage-logging decision. (Informational only under plan-time packing; the final task is already inside the last envelope.)

Envelope-group read (applies when neither short-circuit fires): After the just-completed task, peek the next pending task via manage-tasks next and read its envelope_id (surfaced on the next result dict). If the next pending task's envelope_id equals the assigned group: continue to it. Else (the next task belongs to a different envelope, or the queue is empty): yield. This is a pure equality check — NO cost summing, NO threshold evaluation, NO remaining_budget, NO self-measurement.

The packing already encodes the budget decision, so the two short-circuits above are now informational rather than the budget mechanism. The clause runs once after each task completes — between manage-tasks finalize-step of the closing step (which fires the canonical [OUTCOME]) and the next manage-tasks next call. There is no intermediate decision point.

Group-exhausted yield path — when the envelope-group read decides to yield because the next pending task's envelope_id differs from the assigned group (and the queue is NOT empty), the executor MUST distinguish this legitimate plan-time-packed yield from a forbidden checkpoint by doing BOTH:

1. Log a budget_yield decision naming the assigned envelope_id, the just-completed task, and the next task's differing envelope_id:

   python3 .plan/execute-script.py plan-marshall:manage-logging:manage-logging \
     decision --plan-id {plan_id} --level INFO \
     --message "(plan-marshall:phase-5-execute) budget_yield: envelope {my_envelope_id} exhausted after {task_id}; next pending task TASK-{next_number} belongs to envelope {next_envelope_id} — yielding to orchestrator for the next envelope-group dispatch"
   

2. Return a wrapped, discriminated terminal TOON carrying an explicit yield discriminator — status: success, a display_detail naming the envelope-group yield, and tasks_remaining > 0 — NEVER the bare task_complete echo from execute-task:

   status: success
   display_detail: "envelope {my_envelope_id} exhausted — {tasks_remaining} task(s) remain in later envelopes"
   plan_id: {plan_id}
   tasks_completed: {N}
   tasks_remaining: {M}
   budget_yield: true
   envelope_id: {my_envelope_id}
   next_envelope_id: {next_envelope_id}
   

   The orchestrator classifies this return as termination_cause: budget_yield (see plan-marshall/workflow/execution.md § "Termination-cause classification") — distinct from the task_complete_returned_verbatim failure mode — and re-dispatches the next envelope group.

Cross-reference to the three terminal outcomes — the envelope-group read is the continue-vs-yield decision, not a fourth terminal outcome. When it says "yield", the agent still exits via the group-exhausted yield path above (a wrapped terminal TOON with budget_yield: true), which is a legitimate use of the queue-not-empty success return — NOT a partial-completion checkpoint (that path is explicitly forbidden by the section above). The in-flight task's state is already persisted by manage-tasks finalize-step so resumption by the next envelope dispatch is lossless.

Audit diagnostic ledger — when investigating throughput or calibration questions, compare each envelope's per-dispatch ACTUAL tokens (the orchestrator's post-return <usage>) against the SUM of predicted_cost_tokens for the tasks in that envelope (prediction-vs-actual calibration of the size→token table). A systematic over- or under-prediction across envelopes is the signal that the cost_size_token_table magnitudes (per_envelope_budget_tokens consumers) need recalibration; the orchestrator owns that feedback loop because only it sees the post-return token counts.

---

Phase Transition

When transitioning from execute phase to finalize, first record phase metrics through the platform runtime:

python3 .plan/execute-script.py plan-marshall:platform-runtime:platform_runtime \
  metrics capture --plan-id {plan_id} --phase 5-execute

On Claude Code the runtime reads the stored session_id and captures token usage from the transcript. On OpenCode it returns no-op — the phase still transitions and accepts manual --total-tokens when available.

Then update the phase status:

python3 .plan/execute-script.py plan-marshall:manage-status:manage-status transition \
  --plan-id {plan_id} \
  --completed 5-execute

This automatically updates status.json and moves to the next phase.

After transition, check finalize_without_asking config:

python3 .plan/execute-script.py plan-marshall:manage-config:manage-config \
  plan phase-6-finalize get --field finalize_without_asking

• IF finalize_without_asking == true: Log and auto-continue to finalize phase
• ELSE (default): Stop and display "Run '/plan-marshall action=finalize plan={plan_id}' when ready."

---

Output

phase-5-execute returns on five terminal paths (queue empty → transition; fatal error; triage blocked; deliverable infeasible; scope-deviation escalate_ask). The minimum contract every workflow doc that implements ext-point-execution-context-workflow MUST return is:

status: success | error | blocked | infeasible | escalate_ask
display_detail: "<{tasks_completed} tasks complete, {tasks_remaining} remaining>"
plan_id: {plan_id}
tasks_completed: {N}
tasks_remaining: {N}
infeasibility_reason: {required when status=infeasible — why the declared deliverable cannot be built as scoped}
prompt_options[N]{id,question,header,options,recommended}: {required when status=escalate_ask — the batched scope-deviation / smart_and_ask questions for the orchestrator to fire}

display_detail shape on success: "{tasks_completed} tasks complete, {tasks_remaining} remaining" (e.g. "7 tasks complete, 0 remaining"). On blocked: "{task_number} blocked: {short reason}". On infeasible: "{task_number} infeasible: {short reason}". On escalate_ask: "{task_number} escalate_ask: {N} deviation prompt(s)". On error: short error label from § Error Handling. All values are ≤80 chars, ASCII, no trailing period. The infeasible return carries infeasibility_reason; the orchestrator routes it to the Step 11 "For infeasible blocks" planning gate (drop / re-scope / abort via AskUserQuestion), NOT the verification-feedback code-fix loop. The escalate_ask return carries prompt_options[]; the orchestrator fires ONE batched AskUserQuestion (§ "Post-return escalate_ask batched deviation dispatch" in execution.md) and re-dispatches with the resolutions baked in, NOT the verification-feedback code-fix loop.

---

Error Handling

On any error, first log the error to work-log:

python3 .plan/execute-script.py plan-marshall:manage-logging:manage-logging \
  work --plan-id {plan_id} --level ERROR --message "[ERROR] (plan-marshall:phase-5-execute) {task_id} failed - {error_type}: {error_context}"

Script Failure (Lessons-Learned Capture)

ON SCRIPT FAILURE: When any python3 .plan/execute-script.py invocation exits non-zero, emit the canonical [ERROR] script-failure line to work-log BEFORE any retry or abort. This is distinct from the [ERROR] task-failure line above — that one captures end-of-task failure context; this one captures every individual non-zero script exit so caller-name drift, argparse rejections, and "Unknown notation" failures stay visible in work.log instead of hiding in script-execution.log.

python3 .plan/execute-script.py plan-marshall:manage-logging:manage-logging \
  work --plan-id {plan_id} --level ERROR --message "[ERROR] (plan-marshall:phase-5-execute) Script {notation} {sub} failed: exit_code={N}, args={...}"

Substitute {notation} with the failing script's bundle:skill:script notation, {sub} with the subcommand (or - when none), {N} with the observed exit code, and {...} with a compact rendering of the call's arguments (mask any obviously sensitive values).

After the emit:

1. Capture error context (script path, exit code, stderr)
2. Continue with normal error recovery (retry, fail task, etc.)

Pending-task drift (fatal)

ON manage-tasks next returning a null next while pending tasks remain: this is a fatal control-flow drift, not a routine "no work to do" signal. The two known triggers are (a) a malformed depends_on graph that leaves every pending task waiting on a non-existent predecessor and (b) a misclassified in_progress task that the loop cannot advance. Either way, transitioning to finalize would silently abandon the pending tasks.

When the loop receives next: null from manage-tasks next, immediately query manage-tasks list --status pending. If the pending count is non-zero, treat it as a fatal error:

1. Emit the canonical [ERROR] line to work-log:

   python3 .plan/execute-script.py plan-marshall:manage-logging:manage-logging \
     work --plan-id {plan_id} --level ERROR --message "[ERROR] (plan-marshall:phase-5-execute) Pending-task drift: manage-tasks next returned null while {N} task(s) still pending — {ids}. This is a fatal control-flow error; do NOT transition to finalize."
   

2. Do NOT call manage-status transition --completed 5-execute. Do NOT auto-continue. Return a structured error payload (see Error Handling above) so the orchestrator can either re-enter execute (if the cause is a recoverable dependency-graph repair) or surface the failure for human review.

The Step 12a "Pending-tasks transition guard" (in the Execution Loop) is the structural check that prevents the transition; this section names the failure mode at the error-taxonomy level so the orchestrator can route the recovery.

Other Errors

| Error | Options | |-------|---------| | Build failure | Fix and retry / View log / Skip task | | Test failure | Fix tests / View details / Skip task | | Dependency not met | Complete dependency / Skip check |

---

Integration

Command Integration

• /plan-marshall action=execute - Primary entry point invoking this skill

Related Skills

• phase-4-plan - Creates tasks from deliverables (previous phase)
• phase-6-finalize - Shipping workflow (commit, PR) (next phase)

Phase-boundary metric bookkeeping

The 5-execute → 6-finalize phase boundary itself is recorded by the orchestrator (plan-marshall:plan-marshall workflows) via the fused manage-metrics phase-boundary call — see marketplace/bundles/plan-marshall/skills/manage-metrics/SKILL.md § phase-boundary for the API. Per-task manage-tasks finalize-step calls during the execution loop are unchanged.

Per-task subagent token aggregation is handled by Step 8b (accumulate-agent-usage) which persists each dispatched agent's <usage> totals to .plan/plans/{plan_id}/work/metrics-accumulator-5-execute.toon. The orchestrator's phase-boundary call reads this accumulator file as a fallback when its explicit token flags are omitted — so the orchestrator does not need to maintain a parallel running sum in model context.

Loop-back re-entry and boundary monotonicity. A finalize loop-back re-enters 5-execute and re-records its work under phase=5-execute, so a later phase's start_time can end up preceding this phase's already-closed end_time. manage-metrics generate carries a render-time monotonicity detector that surfaces the resulting non-monotonic boundary (a top-level boundary_monotonicity warning plus a per-phase annotation) and guards the idle residual for the affected phase, rather than silently emitting a corrupt residual derived from the overlapping window. The detector is read-only — it never rewrites the recorded start_time / end_time fields and never touches the #812 end_time-keyed partial verdict. See manage-metrics/standards/data-format.md § "Boundary Monotonicity (Loop-Back Re-entry)".

Canonical invocations

The canonical argparse surface for the two entry-point scripts this skill registers: scope_creep_check.py and verify_failure_scope.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".

scope_creep_check — check

python3 .plan/execute-script.py plan-marshall:phase-5-execute:scope_creep_check check \
  --plan-id PLAN_ID [--threshold THRESHOLD]

verify_failure_scope — classify

python3 .plan/execute-script.py plan-marshall:phase-5-execute:verify_failure_scope classify \
  --plan-id PLAN_ID [--error-paths ERROR_PATHS]