Deepwork Execute Amend Skill

Deepwork Execute Amend

Invoke a lightweight single-gate amendment cycle for the active execute-mode session.

What this skill does

Spawns a MICRO-TEAM of CRITIC + 1 specialist to re-verdict a single gate in the execute session without re-running the full deepwork team. Use when:

SCOPE-GUARD identifies a necessary-but-out-of-scope change for a specific gate
A scope-delta discovery with proposed_outcome: escalate has been appended to discoveries.jsonl
EXECUTOR reports a plan gap that cannot be resolved within existing plan language

When NOT to use this skill — require fresh `/deepwork --mode default` instead

Per plan §6 (full re-run threshold), redirect to /deepwork --mode default if ANY of:

The amendment touches ≥3 gates (>30% of the plan scope)
The amendment contradicts a categorical ban (G7 secret-scan, G8 CI-bypass)
The amendment changes the plan's test acceptance criteria (affects test_manifest[] entries)
The amendment changes the plan's declared environments (affects env_attestations entries)

In these cases, output a HALT recommendation and ask the user via AskUserQuestion to authorize a full re-run or cancel.

Steps

Read the active execute session state:
```
Glob: ${CLAUDE_PROJECT_DIR:-$(pwd -P)}/.claude/deepwork/*/state.json
```
Filter for the instance with execute.phase not null and not "halt". Read state.json.execute.plan_ref, plan_hash, change_log, scope_amendments.
Identify the gate being amended from the <gate-id> argument (e.g., G-exec-3).
Read the gate's task from TaskList. Read the relevant change_log[] entry with metadata.bar_id == <gate-id>. Read the discovery entry from discoveries.jsonl that triggered this amendment.
Check full re-run threshold (step 2 above). If threshold is met, output HALT recommendation and stop.
Determine which specialist to spawn alongside CRITIC:
- PA dimension issue (scope, plan citation) → scope-guard specialist
- EG dimension issue (test failures, adversarial gaps) → adversary specialist
- RA dimension issue (regression) → executor specialist
- Env attestation issue → auditor specialist
Spawn the MICRO-TEAM via two Agent tool calls in one message:
- critic — include references/critic-stance.md verbatim; instruct to re-verdict ONLY the PA dimension of <gate-id> against the proposed amendment
- <specialist> — include the appropriate stance from profiles/execute/stances/; instruct to read the discovery context and produce an amendment.v<N>.md in proposals/amendments/
Wait for the micro-team to complete. The specialist produces proposals/amendments/amendment.v<N>.md. CRITIC verdicts it.

If CRITIC PASS on the amendment:

Append the scope amendment record via state-transition.sh append_array:

bash "${CLAUDE_PLUGIN_ROOT}/scripts/state-transition.sh" \
  --state-file "${CLAUDE_PROJECT_DIR:-$(pwd -P)}/.claude/deepwork/<instance-id>/state.json" \
  append_array .execute.scope_amendments \
  '{"id":"SA-<N>","gate_id":"<gate-id>","amendment_file":"proposals/amendments/amendment.v<N>.md","reason":"<description>","approved_at":"<ISO>","triggered_by":"<discovery-id or null>"}'

If the amendment modifies the plan file, recompute and store plan_hash and clear drift via set_field:

NEW_HASH=$(sha256sum "$plan_ref" | cut -d' ' -f1)
bash "${CLAUDE_PLUGIN_ROOT}/scripts/state-transition.sh" \
  --state-file "${CLAUDE_PROJECT_DIR:-$(pwd -P)}/.claude/deepwork/<instance-id>/state.json" \
  set_field .execute.plan_hash "$NEW_HASH"
bash "${CLAUDE_PLUGIN_ROOT}/scripts/state-transition.sh" \
  --state-file "${CLAUDE_PROJECT_DIR:-$(pwd -P)}/.claude/deepwork/<instance-id>/state.json" \
  set_field .execute.plan_drift_detected false

Mark the discovery entry's resolution field: "resolved via SA-<N>"
Report to user: amendment approved, session may resume.

If CRITIC FAIL on the amendment:
- Output the FAIL reason to the user.
- Ask user for direction via AskUserQuestion: "Amendment for gate <gate-id> was rejected: <reason>. Options: (1) revise the amendment, (2) cancel this gate, (3) halt session for manual review."

CC hook capability note

permissionDecision:"ask" silently degrades to "deny" in non-interactive mode. This skill does not use "ask".
PreToolUse blocking uses permissionDecision:"deny" (via hookSpecificOutput) or exit 2, not the deprecated decision:"block".

Companion commands

/deepwork-execute-status — view current execute session state before invoking amendment
/deepwork --mode default — full re-run when amendment threshold is exceeded

Agent Skills: Deepwork Execute Amend

Install this agent skill to your local

Skill Files