feat(ENG-4364): opt-in tier3.dropShapeOnlyFields for whole-result reviews
Merged
HI
1d ago+474+474+012345678901234567890123456789-3-3-0123456789
What
A new opt-in Tier 3 option, tier3.dropShapeOnlyFields (default false), for whole-result reviews (tier3_only and cascade with cascadeReview: "result").
With it on, the field: value serializer leaves a line out of the reviewer input when both hold:
- the key is a plain field path of short field names, e.g.
data.nodes[0].status.id: each key segment is at most 32 characters and at most 3 words (split on_,-,$,@and camelCase); - the whole value, trimmed, is one of:
null,undefined,trueorfalse;- a number;
- a UUID or a hex ID (8ā64 hex characters, at least one digit);
- an ISO date or date-time;
- a hex colour.
These values can't carry an instruction.
Why
List-heavy tool results serialize mostly to lines like items[0].id: <uuid>, items[0].canceledAt: null or items[0].progress: 0. They fill the review window (maxChunks Ć maxTextLength), so free text past it goes unreviewed. They also make each judge call longer and slower. Leaving them out puts the window on text.
Design choices
- Where: the filter runs inside the serializer, before the cap and chunking, so the freed budget goes to text lines in the same pass.
- Never left out:
- lines of a multi-line value, so a bare number or date inside text stays;
- keys with spaces or other characters (keys are attacker-controlled text);
- empty values (
key:); - top-level values.
- Paths: matched with ASCII word characters only, so non-ASCII keys always stay.
- URLs and short words stay. Dropping them also removed pieces of injected multi-line values in testing.
- Coverage accounting: a left-out string counts as covered, not as dropped for budget. Filtering alone never sets
oversizeorcoverageDegraded. A result with nothing else left skips the provider, as an empty result does today. - Scope: only the reviewer's input changes. Tier 1, Tier 2, SFE and the returned result are untouched.
Tests
specs/tier3-shape-filter.spec.ts:- shape-only values left out; text, URLs, emails, short words and empty values kept;
- non-plain keys, text-like keys and multi-line values kept;
- hex-looking words without a digit kept;
- off by default;
- coverage: left-out values don't flag oversize, an overflowing result is fully covered with the filter on, and real overflow is still reported;
- the provider is skipped when nothing is left.
- Five new
tier3-serializeconformance cases for the Python port, and the rule inconformance/README.md.
pnpm test, pnpm lint and pnpm test:typecheck pass.
š¤ Generated with Claude Code