continue-work

Continue Work

Resume automated implementation for a task whose remediation loop budget was exhausted. This skill resets the counters, optionally accepts new user guidance, and relaunches a fresh code session with the full original context.

---

Sentinel — acquire before writing TASK cards

A PreToolUse hook (tasks-folder-guard.py) rejects every Write/Edit/MultiEdit/ NotebookEdit call targeting tasks/<CAP>/TASK-*.md unless the shared task-pipeline sentinel /tmp/.claude-task-pipeline.active is present and ≤30 min old. This skill is on the allowlist (together with /task, /task-refinement, /launch-task, /code, /fix, and /pr-merge-watcher). The code session this skill relaunches acquires its own sentinel inside its session — /continue-work only needs the sentinel for its own counter-reset edits.

Before the first TASK-card write (resetting loop_count, stalled_reason, status):

touch /tmp/.claude-task-pipeline.active

At the very end (success or graceful abort):

rm -f /tmp/.claude-task-pipeline.active

A stale sentinel grants write access to the next agent — explicit rm -f on exit is preferred. The hook treats sentinels older than 30 minutes as expired.

---

Usage

/continue-work TASK-NNN
/continue-work TASK-NNN --max-loops 15
/continue-work TASK-NNN --max-loops 20 "Look for #consent-gate not #gate-consentement"

Argument	Required	Default	Description
TASK-NNN	yes	—	Task identifier to resume
--max-loops N	no	same as previous max_loops	New loop budget for the session
"guidance text"	no	—	Optional hint injected into the remediation context

---

Step 1 — Locate and Validate the Task

1. Find the task file at /tasks/{capability-id}/TASK-NNN-*.md. If not found, scan all tasks/*/ directories and report a list of stalled tasks.

2. Verify the task status is stalled:

   • If status: done → tell the user: "TASK-NNN is already done — nothing to resume."
   • If status: in_progress → tell the user: "TASK-NNN is already running."
   • If status: in_review → tell the user: "TASK-NNN is awaiting PR merge, not stalled."
   • If status: todo → tell the user: "TASK-NNN has not been started yet. Use /code TASK-NNN."

3. Read from the task frontmatter:

   • loop_count: number of loops used before stalling
   • max_loops: the budget that was exhausted
   • stalled_reason: multi-line failure summary written by the code skill

---

Step 2 — Show the Stall Context

Present the situation clearly before asking the user to confirm:

⚫ TASK-[NNN] — [Title]
Capability: [ID] — [Zone]

Stalled after [loop_count] loop(s) / budget was [max_loops].

Last failing criteria:
  ❌ [Criterion 1]: [failure description from stalled_reason]
  ❌ [Criterion 2]: [failure description from stalled_reason]

Loop budget for new session: [new_max_loops]
  (pass --max-loops N to change — current default: [max_loops])

Optional guidance to inject into the first remediation prompt:
  [guidance if provided, or "(none — will retry with the same context)"]

Shall I reset and relaunch? [yes / no / --max-loops N "guidance"]

Wait for the user's confirmation (or adjustment) before proceeding.

---

Step 3 — Reset the Loop Counters

Determine the new max_loops value:

• If --max-loops N was provided: use N
• Otherwise: reuse the previous max_loops value (not 10 — respect the setting already in the file)

Update the task file frontmatter:

status: todo
loop_count: 0
max_loops: [new_max_loops]
# stalled_reason field: keep it for history but prefix with "# (resolved attempt):"

Write the updated frontmatter. Do not touch any other field (preserves depends_on, priority, pr_url history, etc.).

---

Step 4 — Notify the Board

Invoke /sort-task to refresh /tasks/BOARD.md and reflect the todo status (→ will transition to in_progress once the code agent starts). /launch-task will pick it up reactively when its Step 6 fires.

Report:

Board updated: TASK-[NNN] moved from ⚫ stalled → 🟢 ready.

---

Step 5 — Relaunch via Code Skill

Invoke the code skill for TASK-NNN with the following additional context prepended:

── CONTINUE-WORK CONTEXT ──
This task was previously stalled after [loop_count] remediation loop(s).

Stall reason (last failing criteria):
[stalled_reason content]

[IF guidance provided]:
User guidance for this session:
"[guidance text]"

New loop budget: [new_max_loops] iterations.
Loop counter has been reset to 0.
── END CONTINUE-WORK CONTEXT ──

The code skill will:

• Read the updated loop_count: 0 / max_loops: [new_max_loops] from the frontmatter
• Detect the zone and route to the correct implementation path
• Skip the initial implementation step if artifacts already exist and only re-run tests + remediation (the code skill handles this via the Stall Procedure check in Step 3)
• Run the matching test skill — /test-business-capability for non-CHANNEL tasks (test-business-capability agent) or /test-app for CHANNEL tasks (test-app agent) — and loop up to new_max_loops times

Say:

"Relaunching code skill for TASK-[NNN] with [new_max_loops] loop(s) available..."

---

Step 6 — Report

After the code skill completes (pass or new stall):

If all tests pass:

✅ TASK-[NNN] resumed and completed.
All DoD criteria validated after [loop_count] additional loop(s).
PR: [PR_URL]

If the budget is exhausted again:

⚫ TASK-[NNN] stalled again after [new_loop_count] additional loop(s).

Failing criteria (new run):
  ❌ [Criterion 1]
  ❌ [Criterion 2]

Suggestions:
  - Increase the budget: /continue-work TASK-NNN --max-loops [N+10]
  - Add targeted guidance: /continue-work TASK-NNN "specific hint about the failure"
  - Review the FUNC ADR or task definition — the criterion may be ambiguous
  - Run /task-refinement TASK-NNN to clarify the acceptance criteria before retrying

---

Skill Boundaries

• This skill only resumes stalled tasks — it does not modify design decisions or task definitions. To change what a task must produce, use /task-refinement TASK-NNN first.
• The stalled_reason written by the code skill is preserved in the file for traceability. Each new /continue-work run annotates it with # (resolved attempt N): prefix rather than deleting it.
• This skill does not create new branches — it resumes work on the existing feat/TASK-NNN-{slug} branch where the previous implementation left off.