graff — a fast agentic coding harness in Zig: multi-provider, MCP, workflows, DGM evolution loop, TS/Python SDKs
0133. Publication review binds committed inputs
Status: proposed; superseded by 0250 (2026-10-03)
Context
Passing a helper test does not establish the behavior of its caller. A PR description can claim dispatch coverage that neither its committed tests nor the implementation support. Keyword checks cannot establish that coverage.
Decision
Non-draft publication receives a bounded, tool-free review of the PR body, exact repository/base/head, changed committed source and tests, and separately identified check observations. Missing required evidence, oversized diffs, unsupported inputs, and unresolved or malformed review results do not establish readiness. Drafts remain available. No particular wording is required to trigger review.
Changed files supply the proposed-head blob when it fits and a context diff for prior lines; unchanged base blobs are not repeated. In mixed source and documentation changes, documentation pages use marked context diffs to leave room for committed test entry points. A complete proposed-head blob that exceeds the individual or aggregate source budget likewise falls back to a marked committed context diff. The reviewer must return unresolved if omitted source is needed to assess a claim. An oversized diff still blocks publication. Test entry points follow committed package scripts and runner imports through a bounded chain. The packet names size and traversal limits when evidence is omitted. Successful local check receipts remain bounded and survive session restore, with their observed head and completion status; they are not immutable execution proof.
The review supplements actual tests and fresh CI observations. It is not a mechanical coverage certificate. Changed files may omit important callers; the reviewer must retain uncertainty rather than invent those paths. A check receipt records what the harness observed, including incomplete output and post-run identity limits, without claiming immutable execution.
Review requests share the caller's model budget and cancellation. Successful assessments may be reused in the same process only for matching repository, base, head, body, source and check observations. The common execution boundary revalidates those inputs. A new process or changed input requires new review.
Consequences
Non-draft publication can require an additional model request. Large changes or missing dependencies can remain unresolved even when local checks pass. The review must not hide those limits or substitute a passing assessment for execution evidence. Runtime regressions must cover both refusal and acceptance, input changes, alternate targets, cancellation and exhausted budgets before this proposal is considered fully verified.
Related: #847; extends ADR 0100 and ADR 0126.