On PR #86 the reviewer approved with "cleanly delivers both payoff items … following the repo's existing idiom … and the plan's stated approach exactly" and filed zero findings, while the diff carried a redundant double-404, a flag standing in for an unnamed error path, and a test writing straight into DBOS's status table. Its own WHAT TO LOOK FOR list already names dead branches, foot-guns, reuse, and test shape. It never got that far: every clause of its verdict grades plan conformance, which is the evaluator's job and which this prompt explicitly forbids it from redoing.

It does that because of what it is handed. All seven roles include one _header.md, which renders the plan markdown, every acceptance criterion, the ruled-out list, prior evaluations, human feedback, and a MANDATORY FIRST ACTION to fetch the ticket as binding truth. The diff appears only as two SHAs.

The split

_contract.md now carries the ticket-fetch order, the plan, the acceptance criteria, the ruled-out approaches, prior evaluations, and human feedback. The six contract-bearing roles include it immediately after _header.md, so their prompts keep every section in the same order.

review_code.md includes _header.md alone — workflow context, workspace paths, verification, base_sha/head_sha — and gains one mandatory first action: git diff base..head, then read every changed file end to end, before any finding. It fetches the ticket when it needs scope to judge a finding and to file the follow-up sub-issue, rather than opening with it as a rubric.

Review prompt changes

• Added to WHAT TO LOOK FOR: edits outside the ticket scope, named by file path — the class only a reader of the diff can catch. PR ENG-736 - Delete model_prefixes: route models off the harness's own fetched-or-shipped list #82's unrelated startswith → text[:3] rewrite in subprocess.py is the live example; it is in main.
• Added out-of-scope edits at high severity.
• Removed from WHAT NOT TO FLAG: "Style nits that don't change behavior or maintainability (whitespace, alphabetical import order, single vs double quote unless the repo enforces one)" — it read as licence to ignore house style.

Lint

ignore = ["SIM108"]. It rewrites if/else into a ternary; formatter-wrapped ternaries are banned here, so the lint gate was arguing against the house form.

Tests

test_contract_context_is_omitted_only_from_code_review renders all seven roles against a fully-populated journal and asserts the six carry every contract marker in order while the reviewer carries none of them and does carry the diff instruction.

test_build_prompt_orders_the_ticket_fetch now runs over _CONTRACT_TEMPLATES — the reviewer is deliberately outside that set.

Note

The six roles are not byte-identical to main: the ticket-fetch paragraph moved from the Workflow context bullets down into _contract.md, since the ticket is the contract. Section order and content are otherwise unchanged.

uv run ruff check backend, uv run ruff format --check backend, and backend/tests/test_build_prompts.py (23 passed) are green locally; the rest gates in CI.