deliver

name	deliver
description	Persistent completion mode — tasks must finish fully with verify/fix loops until done

When this skill is invoked, you MUST follow this exact output sequence.

Step 1: Header (print IMMEDIATELY)

[BW] deliver mode — persistent completion, no task left behind...

Step 2: Task registration

Break the user's request into discrete completion targets. Each target has a clear done-condition.

[BW] deliver targets:
  [ ] 1. {target} — done when: {concrete condition}
  [ ] 2. {target} — done when: {concrete condition}
  [ ] 3. {target} — done when: {concrete condition}

Step 3: Execute with verify/fix loops

For EACH target, call TaskCreate for live progress: TaskCreate: subject="target N: {target}", activeForm="delivering target N..." Update with TaskUpdate as targets complete.

Execute in a loop until the done-condition is met:

[BW] deliver: working on target 1...
[BW] ▶ bestwork:{agent} implementing target 1

After implementation, VERIFY the done-condition:

[BW] deliver: verifying target 1...

Verification must use concrete checks:

• If code: run type check (npx tsc --noEmit), run tests, check imports
• If config: validate JSON/YAML, check references resolve
• If docs: verify links, check accuracy against code
• If refactor: ensure no regressions (tests still pass, no broken imports)

If verification passes:

[BW] ✓ target 1 complete (1 attempt)

If verification fails:

[BW] ✗ target 1 failed verification: {reason}
[BW] ↻ deliver: fixing target 1 (attempt 2)...

Fix the specific issue, then re-verify. Repeat until done.

Step 4: Progress tracker

After each target completes or fails, show progress:

[BW] deliver progress:
  [✓] 1. {target}
  [→] 2. {target} (attempt 3)
  [ ] 3. {target}

Step 5: Escalation (after 5 failed attempts on one target)

If a single target fails verification 5 times:

[BW] deliver: target 2 stuck after 5 attempts
[BW]   last failure: {description}
[BW]   attempted fixes: {list}
[BW]   options:
    a) skip and continue (mark as incomplete)
    b) simplify the approach
    c) halt deliver mode

Wait for user input. This is the ONLY time deliver asks for input.

Step 6: Completion

[BW] ═══════════════════════════════════
[BW] deliver complete: {N}/{M} targets done, {K} total attempts
[BW]   {list completed targets}
[BW] ═══════════════════════════════════

Step 6.5: Summary Report

After printing the completion block, call report-writer (src/observe/report-writer.ts) to save detailed output:

writeReport("deliver", {
  total: {M},
  done: {N},
  agents: [list of all agents used],
  tasks: [list of target descriptions with attempt counts],
  durationMs: {elapsed ms},
  decisions: [escalation decisions, fix strategies tried]
}, projectRoot)

Then call summarize (src/observe/result-summarizer.ts) and print the 1-liner:

[BW] ✓ {N}/{M} done — details: .bestwork/reports/deliver-{timestamp}.md

After writing the meeting log footer to .bestwork/state/meeting.jsonl, also append a decisions entry to .bestwork/context/decisions.md. Create the file and directory if they do not exist.

Format:

## {date}: {task}
- **Mode**: deliver ({N} targets, {K} attempts)
- **Verdict**: {COMPLETE/INCOMPLETE}
- **Key agents**: {comma-separated agent list}
- **Summary**: {1-sentence summary of what was delivered}

State persistence

Save deliver state to .bestwork/state/deliver.json:

{
  "created": "<ISO timestamp>",
  "task": "<user's original request>",
  "targets": [
    { "description": "...", "doneCondition": "...", "status": "complete", "attempts": 1 },
    { "description": "...", "doneCondition": "...", "status": "in_progress", "attempts": 3 }
  ]
}

Resume with ./deliver — picks up incomplete targets.

Rules

• NEVER mark a target as done without concrete verification
• Verification must run actual commands (not just visual inspection)
• No max retry limit per target (except the 5-attempt escalation prompt)
• Targets are executed sequentially (later targets may depend on earlier ones)
• Each verify/fix cycle should attempt a DIFFERENT fix strategy (not the same thing repeatedly)
• Agent selection: pick the best-fit tech specialist per target, switch agents on escalation
• Write meeting log to .bestwork/state/meeting.jsonl with attempt tracking
• deliver does NOT use PM or critic agents — the verify loop replaces them