Plan 0019
Search results you can trust
The search surfaces (fff.grep and friends) stop failing silently: a query that cannot be answered completely says so, index staleness is visible, and a recall corpus in CI keeps false negatives from shipping.
- Status
- Sketch
- Ambition
- 4 a correctness bar for existing search surfaces, not a new engine
- Impact
- 6 silent false negatives poison audits that claim covered everything
- Effort
- 4 index-state visibility, honest degradation, and a recall test corpus
- Risk
- 2 worst case is slower answers that admit their gaps
- Maturity
- 3 the surfaces work in the common case; the failure modes are silent
- Leverage
- 6 every audit and every agent that greps through it inherits the bar
- Taste
- 3 the bar is obvious; the work is tests and plumbing
- Authors
andrewgazelka- Created
- Updated
- Tracking issue
- -
- Supersedes
- -
- Superseded by
- -
Summary
A search tool that silently returns partial results is worse than none: an audit that greps for callers and misses some claims a coverage it does not have. The fff.grep issue cluster is exactly that shape: plain-mode misses, regex classes that quietly do not match, path filters that exclude more than asked, stale indexes serving old trees. The bar this plan sets: every query either answers completely or says why it cannot, staleness is part of the result, and a recall corpus runs in CI so regressions are caught before an audit trusts them.
Non-goals
Ranking or performance work; correctness first.
Absorbed issues
#1136, #1134, #1131, #1129, #1123, #1115, #1103, #1049, #1033
Drafted by Claude Code (Fable) from the open-issue corpus; scores are seed estimates.