Review and policy

Intake triage

GitHub Edit page

Intake triage

  • Status: active by default with native OpenAI authentication
  • Owner: ClawSweeper maintainers
  • Source of truth: src/intake-triage-contract.ts, src/intake-triage-api.ts, src/intake-triage-decisions.ts, src/intake-triage-workflow.ts, .github/workflows/sweep.yml
  • Last verified: 2026-10-16; configuration, admission, and workflow ordering covered by the focused intake and sweep-workflow tests
  • Update when: credentials, model defaults, questions, input limits, backends, provisional labels, or workflow ordering change

Intake classifies the author's submission before source review. It answers three bounded questions: request category, whether trigger steps are supplied, and whether version and environment are supplied. It does not establish that a bug exists or that a patch is correct.

The ordinary Responses adapter works without the Decisions API. Both adapters use the same packet and rubric. A caller must record request intent before a billable request and record returned usage separately from full-review usage. Missing credentials, timeouts, refusals and malformed output produce no labels. An input scanner refusal propagates and must not trigger another backend.

The packet includes only repository, item number, kind, title and body. It is limited to 12 KB; oversized submissions abstain rather than silently truncate. No source checkout, comments, tools or full-review prompt is sent. Each request has a 20-second deadline, a 32 KB response limit, and no application retry. Both adapters default to gpt-6-luna. Responses uses strict structured output and disables response storage.

Labels in the finite clawsweeper:intake:* family are preliminary. They never set priority, impact, reproduction proof, good-first-issue eligibility, maturity, merge risk, or readiness. They must be removed by the normal completed-review publication. Intake must not satisfy review requests or populate review caches.

Adding a cheap call before a full review adds spend. Cost savings require moving an existing metadata task to intake, not counting both calls as a saving.

The default Decisions adapter calls POST /v1/decisions with the selected model and three choice questions. It validates names, choices, complete distributions, finite probabilities, and refusals. Native choice distributions are retained separately; Responses does not invent matching confidence values. No probability threshold is presented as calibrated without labeled production examples.

See the official Decisions contract. Decisions failures abstain. They do not automatically spend on Responses or resubmit scanner-refused content. The ordinary Responses adapter remains an independent path for deployments without Decisions access.

#Hosted enablement and lifecycle

Intake uses the existing repository secret OPENAI_API_KEY and is enabled by default in native OpenAI modes. The separate OpenClaw runner can use the same key. Private ClawRouter Codex routes keep upstream keys withheld; their workload credential is never forwarded to the public OpenAI API. An absent native key disables intake without model, GitHub, queue or ledger work.

The default backend is decisions and the default model is gpt-6-luna. Set the repository variable CLAWSWEEPER_DECISION_MODEL to override the model, or to off to disable intake. Unset and blank values use the default. This model is independent of the full-review model; the selected value is sent to the provider and recorded in the paid inference intent.

Set CLAWSWEEPER_INTAKE_BACKEND=responses to use ordinary Responses without Decisions access. The model must be supported by the selected API. Unsupported models abstain without an automatic cross-backend retry.

For admitted exact reviews, intake runs after the review-in-progress reservation and trusted scanner setup, before target-source checkout and review hydration. Held or superseded reservations skip runner and target-source preparation. Its preliminary labels remain visible during target-source preparation and the full review. This is a bounded serial intake phase, not a second background writer. Oversized metadata-only reviews skip it. A failed optional phase does not block full review. The automatic acknowledgement changes to reviewing only after source preparation succeeds.

The publisher checks the queue lease/generation and unchanged item snapshot before each label-definition or item write. Only the six known intake labels can change; all other labels stay owned by their existing publishers. Request intent, usage and mutation receipts use a distinct intake operation in the existing action ledger. Interrupted writes are finalized as unknown outcomes, even when the later full review succeeds. They never complete a review request or its cache.

Enabled intake invalidates the earlier PR-admission handoff even after an uncertain write or process failure, so source review hydrates current labels and timestamps. Only a current complete review removes preliminary labels, through the existing guarded label batch. Failed, stale and incomplete reviews leave them explicitly preliminary. There is no automatic cross-backend retry.