devops-run-burn

name	devops-run-burn
version	0.3.0
description	High-throughput autonomous task runner with aggressive parallelization. EXPLICIT INVOCATION ONLY — the user must type /devops-run-burn to activate. Do NOT trigger on any phrasing, keyword, or intent — not on "burn", "budget", "limit", "aufbrauchen", "maximize", "alles verbrauchen", "token", or ANY other wording. Only the literal /devops-run-burn command.
argument-hint	[task description or goal — additional tasks are discovered automatically]
allowed-tools	Bash(*), Read, Write, Edit, Glob, Grep, Agent, AskUserQuestion, CronCreate, CronDelete, CronList, EnterWorktree, ExitWorktree, TodoWrite, WebFetch, WebSearch, mcp__computer-use__*, mcp__Claude_Preview__*, mcp__plugin_playwright_playwright__*, mcp__plugin_devops_dotclaude-completion__*, mcp__plugin_devops_dotclaude-ship__*, mcp__plugin_devops_dotclaude-issues__*

Run Burn

Maximize the remaining weekly token budget. Collect tasks from all sources, then launch autonomous mode with aggressive parallelization to get maximum value out of the remaining time window.

This skill MUST only run when explicitly invoked via /devops-run-burn. Never trigger from hooks, prompt phrasing, or heuristic matching.

Step 0 — Load Extensions

Silently check (do not surface "not found"):

1. ~/.claude/skills/run-burn/SKILL.md + reference.md
2. {project}/.claude/skills/run-burn/SKILL.md + reference.md
3. Merge: project > global > plugin defaults

Step 1 — Burn Confirmation Gate

MANDATORY — never skip this step.

Before doing ANYTHING else, ask via AskUserQuestion:

BURN MODE

Dieser Modus verbraucht aggressiv dein verbleibendes Weekly-Budget durch maximale Parallelisierung. Bei aktivierter Zusatznutzung kann das über dein Standardlimit hinausgehen.

Bist du sicher, dass du den Burn-Modus starten willst?

Options: ["Ja, burn starten", "Nein, abbrechen"]

• "Ja, burn starten" → proceed to Step 2
• "Nein, abbrechen" → stop immediately, output: "Burn abgebrochen." — do nothing else

Step 2 — Budget Assessment

Fetch current usage data to understand how much budget remains:

node "${CLAUDE_PLUGIN_ROOT}/scripts/refresh-usage-headless.js" --quiet --summary

Read ~/.claude/usage-live.json and compute:

• Weekly remaining % — how much of the weekly budget is left
• Time remaining — hours until weekly reset
• Burn capacity — rough estimate: "remaining %" split across available time

Present a one-line summary:

Budget: {remaining}% | Reset in {hours}h | Modus: BURN

If weekly remaining is < 10%, warn:

"Nur noch {X}% übrig — Burn-Modus bringt hier wenig. Trotzdem starten?"

Step 3 — Primary Task Intake

Use $ARGUMENTS if provided. If empty, ask:

Was ist der Hauptauftrag für den Burn?

Parse into: Goal (one sentence), Scope (files/systems), Priority (high).

Step 4 — Task Discovery

Ask the user:

Soll ich zusaetzlich zu deinem Prompt automatisch Tasks aus folgenden Quellen ziehen?

1. GitHub Issues (offen, mir zugewiesen oder unassigned)
2. TODO/FIXME/HACK Kommentare im Code
3. Code-Qualitaet (Lint-Fehler, Type-Errors, veraltete Dependencies)
4. Test-Coverage-Luecken (ungetestete Dateien/Funktionen)
5. Offene PR-Reviews
6. Alle oben genannten
7. Nein, nur mein Prompt

(Mehrfachauswahl moeglich, z.B. "1,2,4")

Step 5 — Task Collection

Based on the user's selection, collect tasks in parallel:

5a. GitHub Issues (if selected)

gh issue list --state open --limit 30 --json number,title,labels,assignees,body

Filter: prioritize assigned-to-user > unassigned > others. Extract actionable items (skip discussions, questions, epics).

5b. TODO/FIXME/HACK Comments (if selected)

Use Grep to find all actionable code comments:

• Pattern: TODO|FIXME|HACK|XXX|OPTIMIZE|REFACTOR
• Group by file, deduplicate near-identical entries

5c. Code Quality (if selected)

Run in parallel:

npm run lint 2>&1 || true
npx tsc --noEmit 2>&1 || true

Parse errors into fixable tasks. Skip warnings unless trivial to fix.

5d. Test Coverage Gaps (if selected)

npm test -- --coverage --reporter=json 2>&1 || true

Identify files/functions below coverage threshold. Create tasks for missing tests.

5e. Open PR Reviews (if selected)

gh pr list --state open --json number,title,reviewDecision,reviewRequests --limit 20

Find PRs awaiting review or with requested changes.

Step 6 — Task Consolidation & Prioritization

Merge all discovered tasks into a single prioritized backlog:

Priority Tiers

Tier	Source	Rationale
P0	User's primary prompt	Explicitly requested — always first
P1	Blocking issues / failing tests / lint errors	Unblocks other work
P2	Assigned GitHub Issues	User's committed work
P3	TODO/FIXME in changed files	Close to current context
P4	Coverage gaps, PR reviews	Valuable but deferrable
P5	Remaining TODOs, unassigned issues	Fill remaining budget

Deduplication

• Merge tasks that touch the same files/modules into a single work unit
• Identify conflicts (two tasks modifying the same function) — sequence them

Plan Presentation

Present the consolidated plan:

## Burn Plan

**Budget**: {remaining}% | **Tasks**: {count} | **Geschaetzte Agents**: {N}

### P0 — Hauptauftrag
- {user's primary task}

### P1 — Blocking
- {lint errors, failing tests, ...}

### P2 — Zugewiesene Issues
- #{number}: {title}

### P3–P5 — Zusaetzliche Tasks
- {list, grouped}

### Execution Strategy
- Waves: {N waves planned}
- Agents per wave (+ model): {Wave 1: core (sonnet) · Wave 2: frontend, windows (sonnet) · Wave 3: qa (sonnet)}
- Estimated token usage: {aggressive / moderate}

Model per agent comes from deep-knowledge/agent-orchestration.md § Model & Effort Defaults (opus for po/research/redteam, sonnet for the rest). Show a default → override when you downgrade a model for cost (§ Model override rules) so the budget impact is visible.

Wait for user confirmation:

• "go" / "ja" / "burn" → proceed as planned
• Modifications → adjust
• "weniger" → reduce scope to P0–P2 only

Step 7 — Launch Autonomous with Burn Guidance

Invoke /devops-run-autonomous with the following composite prompt. The autonomous skill handles everything from here (desktop questions, permission priming, execution, reporting, shutdown).

Composite Prompt Construction

Build the autonomous task prompt from the template in deep-knowledge/composite-prompt.md — it specifies the Hauptauftrag + prioritized task list + burn-guidance (browser Edge Credo, parallelization rules, throughput optimization, task ordering, result consolidation).

Important: Pass the filled-in prompt as $ARGUMENTS to the autonomous skill. The autonomous skill then handles Steps 2–8 of its own flow (desktop questions, permission priming, execution, reporting, optional shutdown).

Rules

• NEVER auto-trigger — this skill runs ONLY on explicit /devops-run-burn invocation
• NEVER skip the plan step — always present the consolidated plan and wait for confirmation
• NEVER skip budget assessment — the user needs to know remaining capacity
• Respect autonomous guardrails — burn mode amplifies throughput, not permissions. All safety rules from devops-run-autonomous still apply (no push, no ship, no external comms)
• Deduplication is mandatory — never let two agents modify the same file simultaneously
• Primary task always wins — if budget is tight, drop P3+ tasks, never the user's prompt
• NEVER skip the confirmation gate — Step 1 is mandatory, even if the user says "just do it"
• Task discovery is optional — if the user says "nur mein Prompt", skip Steps 4–5 entirely and jump straight to Step 7 with only the primary task