autonomous-loop

Autonomous Delivery Loop

Principle

The loop orchestrates roles as subagents and works without user input until the MVP gate, an escalation criterion, or context pressure. The core contract in project/BRIEF.md is the constitution; project/STATE.md the single state store.

Cycle (repeat)

1. Sync: read STATE.md; increment the cycle counter (basis for cooling-off) and check it against the cycle cap — cap reached → runaway stop (see stop conditions), no further cycle. Incorporate open merges, blockers, review results.
2. CEO tick (inline per role-ceo): activate next WORK items up to the WIP limit, evaluate gates. No new reason → no reprioritization.
3. PO step (per role-po): cut activated items into cards — with acceptance criteria, claim limits, and disjoint claim zones. Idea triage via project/IDEAS.md.
4. DEV fan-out (parallel): one DEV subagent per card, all in one invocation block. Embed context instead of having it read: the WORK card and profile extract (commands, quality rules) are verbatim in the prompt — the subagent only reads its role skill. Model choice by card complexity: S/M → sonnet (very large diff or wide context → sonnet[1m]); L or "sensitive" → opus. These override the subagent's own model field (the role skills carry inherit precisely so the spawn decides). The orchestrator verifies the rating before spawning: risk-regex hits in the claim zone, gate relevance, or security/concurrency aspects override a lower PO rating upward — never downward. Zones not safely disjoint or > 2 parallel DEVs → worktree isolation per agent. Parallel DEVs run only targeted tests of their zone plus a regression check of the foundation modules — the full suite run belongs to the orchestrator after the merge (otherwise everyone tests against half-finished neighbor zones).
5. Review pipeline (risk-based): as soon as a DEV finishes, start a reviewer subagent per role-reviewer — don't wait for the slowest card. Full review (gate-relevant → opus, otherwise sonnet) for: production code, gate packages, risk-regex hits in the diff, or included status claims. Light review (haiku) for trivial packages (tests/docs only, small diff): zone check, acceptance check, targeted tests, compact claim audit — without pattern spot-check.
6. Fix loop: findings/BLOCKED → targeted fix run (same card, same zone). Max. 2 attempts per blocker, then the autonomy contract applies (escalation or DEFER with ADR). Findings in files outside all zones (no-man's-land, e.g. legacy or skeleton tests): never have the DEV fix them along the way — the PO immediately cuts a dedicated fix card with its own zone, which runs before the blocked packages merge.
7. Merge sequentially: merge one package after another — before each, the mechanical check per package, executed on the package branch with a clean orchestrator tree (commit your own project/ changes such as STATE.md/IDEAS.md first, so only package changes land in the check): <skills-dir>/_shared/scripts/merge-check.sh <base> --zone <card-zone>… --allow project/backlog/<WORK-NNN>.md --allow project/PROFILE.md --test-cmd "<full test from PROFILE>" — <skills-dir> is the installation location of these skills (project-local .claude/skills/, global ~/.claude/skills/, or the plugin root when installed via /plugin install). No broad --allow project/: that would wave through DEV changes to STATE.md, BRIEF.md, or IDEAS.md. Allowed are only the package's own card (evidence) and PROFILE.md (command maintenance); for special cases the check knows --deny (overrides zone and allow). FAIL blocks the merge hard (zone violation → finding + fix card; red suite → fix loop) — no discretion. The full suite runs exactly once per package: here, via the check — DEV and reviewer only test targeted, duplicate runs are cut. Card to DONE incl. evidence, update STATE.md.
8. Collect ideas: transfer IDEAS/FOLLOW-UP from all DEV outputs to project/IDEAS.md — record only; evaluation happens in the PO step of the next cycle (cooling-off).
9. Stop check (see below), otherwise next cycle.