The Software Factory — Proof Before Commit
Who Decides What
The Owner Decides
Agents do the work; the owner makes the calls that leave the machine. One table in the toolkit's startup reference records who approves each action, in four tiers:
| Tier | Meaning | Examples | Approver |
|---|---|---|---|
| T0 free | No outward effect, git can recover it | Edit files, run tests or the gate | Agent |
| T1 cheap undo | Local and reversible | Commit through commit.js with the ship check |
Agent, after verification evidence |
| T2 outward | Leaves the working tree or changes who can do what | Push, open a PR, publish, deploy, fleet sync, change CI, permissions or hooks | Owner's explicit go, each time |
| T3 irreversible | Cannot be taken back | Force or mirror push, deleting a remote ref, exposing a credential | Owner; a deny rule cannot be waived by an in-session go |
Agents never push, deploy or publish on their own. A green build is not a go, and an earlier go does not carry over to the next action.
The Scheduled Detect-Only Job
An optional recipe (copied and installed by you, not a shipped file) notices a red gate while nobody is at the keyboard:
- a systemd user timer every four hours, or a path unit on a new gate failure, starts a wrapper script
- the wrapper reads the last local gate result with
gate-signal.js --only-if-red; it never runs a gate. A quiet result ends the run with no model call - for an unseen red, the model gets file tools only (Read, Glob, Grep, Edit, Write), in a throwaway copy of HEAD, and writes one intake note for a person to triage
- it never fixes, commits or pushes; anything changed outside the intake folder fails the run
The Improvement Loop
The toolkit measures and improves its own agents, but never lets them edit their own judge:
- agent mistakes are logged as structured rows (
agent-issue-log.js add) - intake attributes them to an exact skill
- a proposed skill change is measured, and
skill-eval.js verdictaccepts or rejects it - a protected set — the gate, hooks, guardrail rules, settings, eval tasks and judge code — is off limits to any improvement run or subagent
Everything is propose-only: the weekly cycle recipe leaves a local auto/evolve-<skill>-<date> branch that never merges and never pushes, autoApply is off unless the committed config says exactly "on", and the managed pre-push hook refuses pushes to auto/* branches. The owner reads the branch and merges or deletes it.
"The tests are green, so I deployed" is the failure this table exists to prevent. Commit is the last thing an agent does on its own; everything outward waits for the owner.
A run's tests and seal pass. Which next step needs the owner's explicit go?