chump-orchestrator

AUTO-013 — the Chump self-dispatching orchestrator. See docs/AUTO-013-ORCHESTRATOR-DESIGN.md for the full architecture.

Status: MVP COMPLETE (steps 1-5) + dispatch-backend pluggability (COG-025)

All five MVP steps shipped. The crate is the picker (step 1), subprocess dispatcher (step 2), monitor loop (step 3), per-outcome reflection writes (step 4), and the end-to-end synthetic smoke (step 5). COG-025 added the runtime backend selector so dispatched subagents can route through Chump's own multi-turn agent loop on any OpenAI-compatible endpoint instead of the Anthropic-only claude CLI — see "Dispatch backends" below.

With --no-dry-run --watch, chump-orchestrator drives a backlog from launch to all-PRs-merged with zero human input. Real-world dogfood (against docs/gaps.yaml) and Together-vs-claude A/B is COG-026; lesson-aware dispatch is AUTO-013-A.

Dispatch backends (COG-025)

The dispatched-subagent binary is selected at spawn time via env:

Both backends honour CHUMP_DISPATCH_DEPTH=1 (depth-1 enforcement) and the gap-claim/worktree contract. Reflection rows include backend=<label> in the notes field so PRODUCT-006 / COG-026 A/B can split by backend.

Operator smoke against the synthetic backlog (no real network):

CHUMP_DISPATCH_BACKEND=chump-local \
  OPENAI_API_BASE=https://api.together.xyz/v1 \
  OPENAI_MODEL=Qwen/Qwen3-235B-A22B-Instruct-2507-tput \
  ./target/release/chump-orchestrator --self-test

Self-test (synthetic 4-gap E2E)

chump-orchestrator --self-test

Runs the full pipeline in-process against docs/test-fixtures/synthetic-backlog.yaml. Uses an injected TestSpawner (touches a dummy file instead of forking claude) and a mock PrProvider that returns Shipped for every branch. No real network calls, no real worktrees. Completes in <10ms; CI-gated by crates/chump-orchestrator/tests/e2e_smoke.rs. Use this to verify the loop is healthy before spending real cloud calls.

Step 4 closes the self-improvement loop spec'd in docs/AUTO-013-ORCHESTRATOR-DESIGN.md §Q6: every dispatch outcome lands in chump_reflections + chump_improvement_targets with error_pattern = 'orchestrator_dispatch' and a structured directive of the shape dispatched gap=… effort=… outcome=… duration_s=… parallel_siblings=… pr_number=… notes=…. PRODUCT-006 (nightly synthesis) reads these rows and distils higher-priority lessons; MEM-006 surfaces those lessons inside the next dispatched subagent's prompt.

Stderr capture: while a subagent runs, a background tailer thread keeps the last 64 WARN/ERROR/FAIL/PANIC lines in a bounded ring. The snapshot is folded into the reflection's notes field on terminal outcome.

The monitor (crates/chump-orchestrator/src/monitor.rs) ticks every 30s per DispatchHandle, probing both the child PID (via try_wait) and the PR state (via gh pr list --head <branch> --state all). It applies the soft-deadline ladder from the design doc — S=20m / M=60m / L=180m — and SIGTERMs (then SIGKILLs after 30s grace) any subagent that has no PR by 2× the soft deadline. The PR probe is behind a PrProvider trait so the state machine is unit-tested without shelling out to gh.

What works today

Dry-run (default — safe, no side effects):

cargo run -p chump-orchestrator -- --backlog docs/gaps.yaml --max-parallel 2 --dry-run

Execute (actually spawns claude subprocesses, returns immediately):

cargo run -p chump-orchestrator -- --backlog docs/gaps.yaml --max-parallel 2 --no-dry-run

Execute and watch (spawns + blocks on the monitor loop until every dispatched subagent reaches a terminal outcome — Shipped, Stalled, Killed, or CiFailed):

cargo run -p chump-orchestrator -- --backlog docs/gaps.yaml --max-parallel 2 --no-dry-run --watch

The execute mode prints DISPATCHED: <GAP> in <worktree> as PID <pid> per spawn. With --watch it then prints a summary table when the monitor loop returns (one line per dispatched gap with its terminal outcome).

Filter rules (MVP)

A gap is pickable iff:

1. status: open
2. priority: P1 or P2 (P3+ skipped until the loop is trusted)
3. effort != xl (XL gaps require human breakdown per design doc §4)
4. all depends_on IDs have status: done

First N in YAML order win. Reflection-driven priority tuning is AUTO-013-A.

Roadmap — five-step MVP

Acceptance criteria status (vs design doc §4)

MVP complete as of step 5. The real-world smoke (against live docs/gaps.yaml with the real claude CLI) is deferred to AUTO-013-A (lesson-aware dispatch) where it composes with PRODUCT-006's synthesis loop. SIGINT teardown is AUTO-013-B.

Tests

cargo test -p chump-orchestrator

Unit tests cover the picker (P1/P2 ordering, P3/XL/done filtering, met & unmet dependency chains, n=0, empty input), the dispatcher (path derivation, step ordering via injected Spawner, prompt assembly, worktree-failure abort), and the monitor (soft-deadline table, every branch of the pure decide_tick state machine, plus MonitorLoop::watch_until_done driven by a ScriptedProvider mock). The CLI smoke tests also exercise --watch against an empty backlog to prove the wiring without forking a real claude subprocess. No claude binary or live gh invocation is needed to run cargo test.