/burn

name: burn
description: "Host-only quota-burning overlay for maximum useful parallel throughput across providers. Use when explicitly asked to /burn, burn usage, or spend available quota before a reset."
disable-model-invocation: true
metadata:
  khanelinix-invocation-mode: user-only

Burn

Turn available subscription usage into completed, verified work quickly. The task's normal workflow owns lifecycle, integration, commits, and final judgment; this overlay owns quota-aware fan-out. Requires `multi-provider-sdlc` and host delegation tools. Keep state in the task's ignored scratch directory.

Workflow

1. **Scope the queue.** Use the requested task or backlog and its completion criteria. If no work was specified, ask for the task before dispatching. Map dependencies and ready slices; keep dependent design and integration with the parent. Each slice needs a concrete artifact or check, paths, one write owner, and an exit condition. Quota is a resource, not a work target. 2. **Read capacity.** Discover live providers, models, and actual host limits. In T3 use `orchestrator_capabilities`; a thread-creation batch limit is not a delegated-worker concurrency limit. Disk is a host limit: check free space as `git-toolkit` cleanup describes before isolating any worker. Use `multi-provider-sdlc` quota advice with `--user-requested` for each relevant provider, fetching independent sources concurrently. Match each source to the account running the workers. Keep all reported windows, headroom, reset times, and available reset-credit inventory. Missing telemetry is unknown, not exhaustion or unlimited usage. When `reset-inventory.py` is installed, enrich the Codex read with `reset-inventory.py codexbar usage --provider codex --source cli --json --no-credits` and pipe its JSON to quota advice with `--provider codex --from-stdin --user-requested`. This optional read reports reset credits without redeeming them; skip it when the helper is unavailable. 3. **Allocate across subscriptions.** Among capable routed workers, favor capacity likely to expire unused before its reset, then replenishable capacity with confirmed available reset credits, then other headroom. Check longer windows and shared model pools too; raw percentages across providers are not comparable token budgets. Use provider preference to break ties, not to leave other subscriptions idle. Route every worker, including implementers, through `multi-provider-sdlc`, using its `subscription` constraint and the live catalog rather than choosing model IDs by hand. Keep role effort defaults; burning usage is not a reason to raise effort. 4. **Fill the ready queue.** Dispatch all useful independent slices that current provider and host capacity can sustain. Use concurrent tool calls and refill slots as workers finish, rather than waiting for a whole wave. There is no default worker count or forecast-based conservation ceiling in burn mode. Grow until ready work, real throttling, host pressure, free disk, or parent integration throughput becomes the constraint. For unknown capacity, let one required task probe each provider/pool, then expand after success. Give an isolated worktree only to a writer that cannot share the checkout; reviewers, playtesters, planners, and other readers never get one. Record each worktree before dispatch, and commit required inputs first because T3 worktrees start from the parent commit. A slot frees only after its worktree closes. 5. **Keep the parent productive.** Give workers bounded packets with task, paths, verified context, constraints, write policy, skill/tool lane, required evidence, and exit criteria. Integrate and verify returned slices while other workers run. Close out each slice's worktree through `git-toolkit` cleanup as soon as the slice resolves, before refilling its slot. Avoid overlapping writers, duplicate investigations, and extra review rounds without an unresolved question. Reserve enough capacity to finish integration and corrections, not an arbitrary quota percentage. 6. **React to actual limits.** Retain task-local routing state, claims, task IDs, worktree ownership records, and outcomes. Before each refill, compare `git worktree list` with those records. An unrecorded worktree, or a resolved one still open, blocks expansion until it is closed out. Low disk is a hard stop for isolated dispatch until resolved worktrees are closed out. On exhaustion, route remaining ready work to another capable subscription instead of trying another model in the same pool. Refresh quota before substantial expansion and after a confirmed reset. A reset credit being available or a reset timestamp passing does not prove replenishment. Redeem a credit only when the user authorized redemption and a supported mechanism exists; never purchase credits implicitly. Otherwise report the available reset while continuing work on other providers. After account-matched fresh telemetry or an explicit user report confirms replenishment, use `multi-provider-sdlc`'s `reset-quota` recovery procedure once for that reset. This rearms one pool for a required probe and preserves other failures. Treat rate throttling separately from exhausted allowance; honor retry timing and reduce pressure on the affected route. 7. **Finish the work.** Stop adding workers when the authorized queue is done, all capable routes are genuinely blocked, or the user stops the run. Collect and inspect every result, complete verification, and close out every recorded worktree; none may remain open after the run. Report completed work, provider distribution, observed usage where available, retained branches with their tips, and any remaining blocker. Never invent usage or claim a speedup without measurements.