| name | factory-implement |
| description | Claim one live GitHub queue item, implement it end to end, run fail-closed gates and independent verification, then open a draft PR. Use for one ready-to-implement issue or the implementation routine. |
Factory implementation
You implement exactly one queue item per run. Not two. If you finish early, stop.
Batching items is how a single wrong assumption becomes a wide diff nobody can review.
Before writing any code
- Read
docs/factory/CONTRACT.md, then docs/factory/CHARTER.md.
- Query GitHub for open issues labeled
factory:ready-to-implement. The live label is
authoritative; read the latest factory-handoff:v1 comment for done_when, expected
files, gate level, and confidence. QUEUE.md is only a snapshot. If the handoff is
missing, duplicated, malformed, or inconsistent with the charter, move the issue to
factory:needs-info and stop. If running locally without GitHub access, stop unless a
human explicitly selects an item for an interactive run.
- Select one item and win the deterministic remote-branch claim described below. Only
after that push succeeds, replace
factory:ready-to-implement with
factory:in-progress. Re-read the issue after the write. If either step failed, stop.
- Re-read
done_when. This is your stopping condition. You are done when it is true and
gates are green, not when the code looks finished.
- Check the item against
LOAD_BEARING yourself. Triage can be wrong. If the work turns
out to touch a load-bearing path, stop, move the item to ready-to-spec, and record
why. Do not proceed carefully; proceed not at all.
If the review queue is already at the charter limit, do not claim an item. Stop and record
the back-pressure condition.
Branch
From the current default branch, create the deterministic branch
claude/fq-<issue-number>. Add an empty commit whose message includes the unique run ID,
then push it without force:
git switch -c claude/fq-<issue-number>
git commit --allow-empty -m "factory: claim FQ-<issue-number> (<run-id>)"
git push origin HEAD:refs/heads/claude/fq-<issue-number>
Two sessions may read the ready label at the same time. Their claim commits differ, so only
the first push can create the remote ref; a later push is rejected as non-fast-forward.
Treat that rejection as "already claimed" and stop. Never force the branch.
In a cloud session the claude/ prefix is accepted for pushes. Do not add a slug or use a
different branch: the deterministic name is the lock.
Implementing
Work in the smallest diff that satisfies done_when.
Write the failing test first. Not as ceremony: the test is what converts "I believe
this works" into a machine-checkable fact, and it is the artifact the verifier uses to
decide whether you actually fixed anything. If you cannot write a test that fails before
your change and passes after, say so explicitly in the PR body and flag the item for a
human read.
Rules while implementing:
- Do not modify existing test files in an unattended run. In an interactive session,
stop and obtain explicit human approval before doing so. The PR stays draft and requires
a human read.
- Do not add abstractions the item does not need. One caller means no interface.
- Do not clean up unrelated code. Note it for the queue instead.
- Do not add dependencies. If one is genuinely required, stop; that is a
ready-to-spec
decision.
- Stay inside
files_expected where possible. Every file beyond it is a signal that the
triage estimate was wrong, and if the count exceeds the charter's limit, stop.
Gates
Run the gate level the queue item specifies:
./.claude/scripts/gates.sh full
Iterate until the final line reads status=GREEN. status=MISCONFIGURED blocks the run;
do not edit the gate configuration yourself to make it pass.
You must quote the FACTORY_GATES: line verbatim in your PR body. You may not
describe the result in your own words instead, and you may not report success if that line
says status=RED. If gates go red twice in a row on the same item, stop and hand it back
per the charter's STOP_IF.
Report optional skipped gates in the PR body. A required skip produces
MISCONFIGURED, so it cannot be mistaken for green.
Commit the scoped implementation and test after gates are green, then confirm both the
index and working tree are clean. The verifier's reversible negative-test proof refuses a
dirty checkout so it cannot hide or overwrite unrelated work.
Independent verification
When gates are green, delegate verification to the factory-verifier subagent using the
Agent tool. Give it the queue item, branch name, and verified base SHA, not your account
of what you did. The verifier reads the diff cold and reaches its own verdict.
You may not skip this step because you are confident. Confidence is what it is checking.
If the verifier returns verdict: rejected, fix what it names and re-run gates and
commit the corrected diff before verification. After two rejections on the same item, stop and hand it to a human. Do not
argue with the verifier in a third pass; two failed attempts means the item was misclassified.
The verifier uses ./.factory/scripts/prove-test.sh from a clean committed branch for the
negative test. Do not substitute git stash.
Pull request
Open a PR only after gates are green and the verifier returns verdict: accepted.
PR body template. Fill every field. Empty fields are how unreviewed work gets merged.
## What
<one sentence>
## Queue item
FQ-<n> - <link to issue>
done_when: <copied verbatim from the queue>
## Why this is safe
<the thing a reviewer would otherwise have to work out for themselves>
## Gates
<paste the FACTORY_GATES line verbatim>
Skipped gates: <list, or "none">
## Verification
Verifier verdict: accepted
<the verifier's one-line reasoning>
## Human read required
<yes + reason, or no>
## Not done
<anything in scope you deliberately left , "">
Mark the PR as draft if any of these hold:
- the change touches a load-bearing path
- an existing test file was modified
- a gate was skipped
- the verifier accepted with reservations
Then replace the source issue's factory:in-progress label with
factory:awaiting-review, link the PR on the issue, and write one unique implement run
record under docs/factory/runs/.
If the run stops after claiming the issue, move it to the correct live state before ending:
- ambiguity or missing human decision ->
factory:needs-info
- load-bearing or scope decision ->
factory:ready-to-spec
- transient infrastructure failure with no code PR ->
factory:ready-to-implement
Never leave an issue factory:in-progress without a run record explaining who owns it.
What you never do
- Merge. Ever. On any tier. The merge decision is the human's, and it is the one place
accountability actually lives.
- Modify
docs/factory/CHARTER.md.
- Modify anything under
.claude/.
- Modify
.factory/gates.conf, AGENTS.md, .agents/, or .codex/.
- Pick up a second item.