Outputs and decisions¶

ReviewOutput¶

One thing the reviewer produced, with the evidence it was based on.

from attest.code_review import FileRead, ReviewOutput

ReviewOutput(
    output_type="issue",
    text="cart_total reads past the end of items.",
    diffs={"billing/cart.py": "--- a/billing/cart.py\n+++ b/billing/cart.py\n@@ ..."},
    changed_files=["billing/cart.py"],
    reads=[FileRead(arguments={"path": "billing/models.py"}, output="class Item: ...")],
)

Field	Meaning
`output_type`	`issue`, `diff_summary`, `group_summary` or `canvas_answer`.
`text`	What the reviewer said.
`diffs`	Unified diff for each file the output talks about, keyed by path.
`changed_files`	The whole pull request's list of changed files, not just this output's. Lets attest catch a wrong "this PR also changes X". Usually left out: `review()` copies it from the pull request.
`reads`	Other code the reviewer looked at. Each `FileRead` has a `tool` (`read_file`, the default, or `grep`), its `arguments` and its `output`.
`question`	The question, for a `canvas_answer`.
`group_name`	The group's name, for a `group_summary`.

If diffs is empty or changed_files is None, review() fills them from the pull request.

paths() lists every file the output's evidence covers, so injection findings in unrelated files don't affect it.

PullRequest¶

The whole change. review() scans it for injection, and attest's own reviewer writes from it.

Field	Meaning
`diffs`	Diff for every changed file.
`changed_files`	Every changed path.
`title`, `description`, `commit_messages`	Author-written text, which is also scanned for planted instructions.

ReviewResult¶

review() returns one ReviewResult for the whole pull request.

Field	Meaning
`decisions`	One `OutputDecision` per output, in order.
`injection`	The `code_injection` check result for the pull request.
`guidelines`	The `ReviewGuidelines` used: your rules, or the general standard (`builtin=True`).
`written`	`True` when attest wrote the outputs itself.
`warnings`	Things to check by hand, e.g. the pull request changes the file its review rules came from.
`kept`	Indexes of the outputs to keep.
`dropped`	Indexes of the outputs to drop.

filter(outputs) returns the kept outputs in their original order. With no argument it returns the kept outputs from the ones review() checked, which is how you get attest's own review when written is True. Passing a list of a different length raises ValueError.

OutputDecision¶

Field	Meaning
`index`	Position of the output in the list.
`output_type`	`issue`, `diff_summary`, `group_summary` or `canvas_answer`.
`text`	What the output said.
`question`	The question, for a `canvas_answer`.
`paths`	The files the output was checked against.
`keep`	Show the output (`True`) or hide it (`False`).
`mode`	The gate mode used.
`score`	The faithfulness score, or `None` when no claim could be checked.
`summary`	The faithfulness summary, e.g. "2/2 checkable claims grounded."
`reasons`	The findings that dropped it. Empty when kept.
`notes`	Findings shown alongside it that did not drop it.
`omissions`	Files whose content was left out of the evidence, with the reason.

GateDecision¶

You only see this type if you call gate() yourself. OutputDecision above is the same decision plus the output's identity and scores.

Field	Meaning
`keep`	Show the output (`True`) or hide it (`False`).
`mode`	The gate mode used: the default for the output type, or the `mode=` you passed.
`reasons`	What dropped the output: each failing claim, each broken rule of yours, and each injection finding that blocks it. Empty when kept.
`notes`	Findings shown alongside the output that did not drop it: a guideline it missed (`misses_guideline`), a `suspicious` hit in its files, or a `compromised` finding in its files that affected a different output.

What drops an output¶

Source	Lenient	Strict
Faithfulness `unsupported`	drop	drop
Faithfulness `unverifiable`	keep	drop
A claim no model judged (a failed call, `metadata["checked"]` is `false`)	drop	drop
More claims than `max_claims` (40 by default): each claim past the cap is unchecked	drop	drop
One of your rules broken (with `follow_guidelines=True`)	drop	drop
Guideline missed (with `follow_guidelines=True`)	keep, listed as a note	keep, listed as a note
Injection `payload` in this output's files	drop	drop
Injection `compromised`, this output affected	drop	drop
Injection `compromised`, another output affected	keep, listed as a note	keep, listed as a note
Injection `suspicious`	keep, listed as a note	keep, listed as a note

An output past the claim cap lists each unchecked claim in reasons, and its summary says how many were not checked. Raise the cap with review(..., max_claims=60) or attest review --max-claims 60.

Injection findings in files outside the output's paths() are ignored. Override the mode with review(..., modes={"issue": "strict"}) or gate(report, "issue", mode="strict").