Unwrap — Inbox Task Recall + Resume

Parses targets, spawns briefing subagents, collects briefings, spawns task workers. All autonomous — no user interaction between briefing and execution.

Step 1 — Parse Target

Accept an argument: all (resume all ongoing tasks) or a single task slug.

• all → resume every task in ~/.claude/tasks/inbox/ongoing/
• [slug] → resume only that task — spawn exactly one briefing + one worker

If slug is given but no matching folder exists in ongoing/, output:

TASK NOT FOUND: [slug]
Hint: Check ~/.claude/tasks/inbox/ongoing/ for the current task list.

Then stop.

Step 2 — Spawn Briefing Subagent(s) in Parallel

Use the Agent tool to spawn general-purpose sonnet subagents. Spawn one per target simultaneously.

Subagent prompt per task:

"Read-only briefing for task '{slug}'. Do NOT write any files.

PERMISSIONS: read-only. No write/edit/create.

Scope

• Task dir: ~/.claude/tasks/inbox/ongoing/{slug}/
• Session logs: ~/.claude/tasks/inbox/ongoing/{slug}/sessions/

FILES TO READ:

• ~/.claude/tasks/inbox/ongoing/{slug}/TASK.md
• Most recent session log in ~/.claude/tasks/inbox/ongoing/{slug}/sessions/ — list sessions/ first to find the most recent file

OUTPUT FORMAT (write exactly this to /tmp/unwrap-{slug}.briefing):

UNWRAP — {slug}

Title: [from TASK.md] Status: [from TASK.md] Priority: [from TASK.md — high/medium/low] Next: [specific next action — one sentence] Blockers: [list one per line, or "none"] Context: [what was happening — 1-2 sentences from session log or TASK.md]

--- Session History --- [session date] — [1-line summary] [...most recent 3 sessions...]

RULES:

• Read-only. Do NOT write or edit any files except /tmp/unwrap-{slug}.briefing.
• If a file doesn't exist, skip that field and note "N/A".
• Keep every field to 1-2 lines max.
• Be specific — "write unit tests for auth module" not "write tests".
• Write ONLY the briefing above to /tmp/unwrap-{slug}.briefing. Then stop."

subagent_type: general-purpose model: sonnet

Wait for all briefing subagents to complete.

Step 3 — Read Briefings

Read each /tmp/unwrap-{slug}.briefing file. Collect all.

Step 4 — Spawn Task Worker(s) in Parallel

After all briefings are collected, read each /tmp/unwrap-{slug}.briefing file. Then spawn one task worker per target simultaneously — briefing embedded directly in the spawn prompt. Workers start immediately.

Task worker spawn prompt:

"You are resuming work on inbox task '{slug}'. Your briefing from the last session is below. Read it carefully — this is your context. Start the stated 'Next' action immediately.

============================================ [READ /tmp/unwrap-{slug}.briefing — paste its full content here as the briefing]

Task Scope (do not read — always embedded here)

• Task dir: ~/.claude/tasks/inbox/ongoing/{slug}/
• TASK.md: ~/.claude/tasks/inbox/ongoing/{slug}/TASK.md
• Sessions: ~/.claude/tasks/inbox/ongoing/{slug}/sessions/

Task Protocol — NON-NEGOTIABLE

1. DECOMPOSE: Break every task into the smallest possible independent sub-tasks. Do not do any work yourself until sub-tasks are decomposed.

2. PARALLELIZE: In a SINGLE message, spawn one Agent per sub-task. All subagents must be launched simultaneously. Do not wait between spawns.

3. REPORT: After EACH subagent completes, send a SendMessage to "team-lead":

   • DONE: "{subtask-label}: DONE — [brief description]"
   • BLOCKED: "{subtask-label}: BLOCKED — [reason] — [workaround]" Report immediately on completion, not at the end of all work.

Agent template: general-purpose (task workers use Sonnet)

Start immediately on 'Next'. Do not re-read task docs unless the briefing says to. When a task block is complete or you are blocked, run /wrap then stop.

subagent_type: general-purpose model: sonnet

Wait for all task workers to report back.

Step 5 — Output Summary

Read /tmp/unwrap-{slug}.briefing to extract Status, Priority, Next, and Blockers for each task. Output:

UNWRAP — {n} task(s)

{slug} — [Status] [Priority] — spawned
  next: [Next from briefing]
  blockers: [Blockers from briefing]

[...one per task...]

All task workers running. Results arrive as each completes.

Then stop. No further narration.