Factory docs, home
Page navigation

See the input → responsibility → output diagram.

What it does for you: It asks one AI model, once, to propose small text-file edits for a Slice. It passes the model only the selected files and the brief, with every tool switched off. It checks the answer strictly and hands back proposed edits, or a plain reason why there are none: unavailable, failed or interrupted. It never changes a file, a branch or a record itself. Two transports share one contract: Claude Code on this Mac, and OpenAI's Responses API.

A playful violet bob with a pencil and blank sketchbook.

Meet the sketcher. Here is an idea to examine. Asks a model for proposed edits; a proposal is not an accepted change. Meet the whole Plumb family.

How it links: It takes files from a Repository snapshot, and it gives proposed edits back for Repository sealing, which decides again whether they apply. Before any call it asks a gate for permission and reports what was used. That gate is not built yet: the Product gate exists but admits no calls to an AI model, so nothing in the Factory uses producers yet.

Honest limits:

  • Claude is unavailable. One live check ran on 28 September 2026. Even with every tool, skill and server switched off, Claude Code still listed 5 built-in agents and 2 built-in plugins. Nobody has shown that these are harmless, so the producer stopped the run at once, and no answer was produced. What that stopped run cost is unknown. The rules have changed since that check (they are now revision 5), so Claude needs a fresh check in any case.
  • OpenAI is unavailable. No OpenAI Platform key was supplied here, so no live request was made. Only offline tests (104 of 104 passing now; 80 of 80 when its receipt was written) show how it would behave. That receipt describes an earlier version of the rules (revision 4); the current rules (revision 5) would need a fresh check once a Platform key exists.
  • The qualification tool withholds recognised credential-shaped process names while retaining the process rows and counts. Hidden names can leave comparison unknown; when new processes are definitely seen, the count excludes any ambiguous comparisons and states that limit. Both outcomes prevent qualification. This limited screen does not detect every secret or qualify the CLI. Source recovery evidence; independent source review and offline checks passed; live qualification remains separate.
  • Stopping Claude on this Mac does not prove the provider stopped working. Claude's spending limit is its own estimate, not a billing cap. Nothing contains Claude Code's own processes.
  • An independent code review accepted this module on 28 September 2026, in its ninth round; each of the first eight found something to fix. It checked the code and its offline tests, not a live run: its own sandbox could not run the file, process and network tests, so those ran on this Mac (397 of 397 passing).

Status: Accepted after independent review: the contract both providers share, and the adapters for the Claude Code command line and OpenAI's Responses API. Acceptance and evidence.

Contract and limits

Use the contract in another application

The private adoption draft lets another application parse bounded proposals and supply its own admission and accounting. It needs no Factory startup or Git checkout. Offline checks use a local pretend CLI and loopback server. They do not qualify either live provider. Current acceptance and review evidence: build review.

A lost accounting acknowledgement stays unknown. The caller checks its own stored record before any further action: calling the Producer again creates a fresh attempt and may spend again.

Neutral adoption contracts, 3 October 2026

The additive PinnedClaudeProducer accepts only an audited executable pin. Policy /2 binds full fixed capabilities, actual HOME and resolved managed paths; legacy qualification cannot admit it. Platform/model rules remain fixed, and no live qualification is supplied. The qualifier stages distinct content-addressed receipt names. See caller setup.

Current evidence and open gates.

Source: docs/guide/producer.md