Last Human Commit

Last Human Commit is a business-first, least-cost agent workflow. It keeps the next real user result ahead of process, chooses the cheapest sufficient agent and proof, and adds governance only when a concrete risk justifies its cost.

Core workflow

1. Define what the user needs now, including an accepted MVP/80/20 Definition of Done.
2. Trace the actual production consumer path before choosing an implementation surface.
3. Name the shortest real business canary and cheapest sufficient proof.
4. Let Lead work directly when delegation costs more; otherwise use the lowest sufficient Worker and preserve its useful context.
5. Treat every 20 active minutes as a reporting/control checkpoint, not an agent lifetime limit. Continue, redirect/resume, or consult Overseer; cancel only exceptionally.
6. Use the harness wait/join mechanism for required children. A timeout is an observation, never proof of terminal state.
7. Consult the supreme Overseer at every crossed hour, overrun, and repeated failure; finish user-facing results only with a real-surface test.
8. Match evidence to the claim and stop when that claim is proven.
9. Keep one unified history: review every path, repair unsafe or unreviewable work, commit the complete result, and end every Full cycle with every repository clean, pushed, deployed, and real-surface tested.

Workers ask Lead at decision boundaries because Lead retains the broader user context and owns business decisions. With a proven non-blocking parent transport, the Worker sends evidence, recommendation/default, parallel-safe work, and the exact blocked action, then continues safe work while Lead decides.

Every declared work cycle records an immutable minimum/maximum estimate. The dependency-free src/common/tools/lhc_time_guard.py emits idempotent hourly business reports and original-maximum overrun diagnostics. Native hooks call it when attested; otherwise Lead calls it at observable checkpoints and reports any delayed hourly delivery honestly.

Overseer, Tester, and Reviewer are the only gates. Gates are tools, not milestones. When a route choice matters, LHC compares two genuinely different approaches after compressing each from ideal to normal to YAGNI/Pareto MVP. It does not manufacture a third option, double testing, per-wave reviews, or hardening loops merely because a task is important.

State and workspace

Use one compact .agents/tasks/ record when recovery, coordination, or audit value justifies it. Update it in place. Legacy todo/work/done lineages remain valid, but new work does not require snapshot copies or snapshot commits.

Routine work stays in the primary checkout. Preserve foreign edits and stage only task-owned paths. Never silently create/switch/merge/delete a branch or worktree. A user-requested worktree lives only at <primary-project-root>/.worktrees/<task-slug>.

Human and secret boundaries

The active harness owns approval policy. Ordinary missing decisions use one compact direct question. Secrets are not work: read them from an environment variable, .env, or a secret file in one step; secret-handoff infrastructure is forbidden.

Install the portable router

Copy src/common/ and templates/ into a project, then use the explicit marker-only helper. It preserves project text outside one canonical block:

scripts/lhc-block init AGENTS.md /path/to/project/AGENTS.md
scripts/lhc-block init CLAUDE.md /path/to/project/CLAUDE.md
scripts/lhc-block apply AGENTS.md /path/to/project/AGENTS.md
scripts/lhc-block apply CLAUDE.md /path/to/project/CLAUDE.md

Optional harness adapters live under adapters/; canonical skills live under skills/ and are mirrored into plugins/last-human-commit/skills/.

Validation:

python3 -m pytest -q tests/test_business_first_contract.py
python3 -m pytest -q tests/test_time_guard.py
python3 tests/validate.py
sh tests/test_task_states.sh
python3 -m pytest -q adapters/hermes/plugin/tests/test_plugin.py