The Loop

The Loop

One pass of the method, from an idea to a merged change. Each step links to the page that covers it and names the README section that holds the reasoning.

Step	What happens	Who or what	README
1	Shape the ticket: one agent pass, agent-ready, risks answered. See Shaping a Ticket.	You, with /lean-agent:shape-ticket	0
2	Triage inline. Sort into tiers, do the cheap ones by hand, close duplicates.	You, no agents	1
3	Draft the change, with builds only. Open a brain agent item when the drafter starts and close it when it reports back.	The cheap model	2, 14
4	Run the mechanical checks: the repository's full gate script, the holdout check, needs-review.sh.	Scripts	4, 10
4a	Run the tests once, as one phase. Each test project goes in with brain queue add, one test-phase drains the queue, and failures land in brain. The fix round works from brain list --open and closes each entry with its fix commit. See Test Queue.	brain, test-phase	11, 14c
5	If needs-review.sh prints anything, the critic runs. See Review Rules.	The diff decides	3
6	Critic: hunt, then refute. See Adversarial Review.	Two fresh agents on the cheap model	2, 12
7	Findings go back to the agent that wrote the change.	The same drafter, resumed	5
8	A finding not fixed in one round goes to the strong model.	The strong model	2
9	Merge. At most four changes in flight.	A person, or a standing authorisation	6, 14h
10	Measure the batch. See Measuring a Batch.	workflow-cost.py, log-review.sh	7, 15
11	Retro. See Retro.	/lean-agent:lean-retro, you approve	16

Tiers

Tier	What
0	Docs, tests only, config, CI files, README sweeps
1	A feature inside one module, additive
2	Core, auth, tenancy, data shape, anything reachable without a login

The tier comes from the ticket and decides how you triage. It does not decide whether a critic runs. The diff does, because the tickets that read as small are the ones that touch auth.

Rules that hold through every step

• Scripts, not instructions. Run the repository's full gate script, not the one gate that seems relevant.
• The second time something is reasoned through, it becomes a script with a test. Fix the script, not the run.
• No follow-up issues from review findings. A finding is fixed in the change or dropped. One that is real and cannot be fixed there goes on the checklist of the open issue that owns that area.
• A task is not done until the pull request URL exists. An agent that hands back without one says so on the first line of its report.
• Every agent writes under its own scratchpad subdirectory. Never a bare pr-body.md at the root.
• Every wait loop has a deadline. Subagents run commands in the foreground, and every gate piece finishes under the 600 second tool timeout.
• Nobody waits in a loop. Enqueue the test job and end the turn. The drainer's exit is the one notice. Never a sleep or pgrep loop.
• State that changes lives in brain. Failures, deferred minors, questions for the person, blockers and running agents go in Brain, so they survive a compaction. Ledgers keep rulings only.
• Heavy commands queue. Builds, installs, containers and browsers go through heavy.sh, one command per lane.
• Parallel changes never append to one shared file. One fragment file per change.
• Plan the steps that need a person up front. A run ends with one block of exact commands for the person.
• Verify content and identity, not status codes. A page baked empty still answers 200. Use check-deployed-sha.sh to prove a deploy.

In one session

With LEAN_AGENTS=one, steps 3 to 8 happen in the one session, one ticket at a time. The critic still runs as two passes: write the candidates down, then refute each one from the code and a check. See Settings.