Skip to content

fix(search): cli-core 0.665.0 and a skill that starts a question with meaning search - #893

Merged
leemour merged 1 commit into
mainfrom
fix/search-skill
Nov 15, 2026
Merged

leemour merged 1 commit into
mainfrom
fix/search-skill

Conversation

@leemour

@leemour leemour commented Nov 15, 2026

Copy link
Copy Markdown
Member

Task

Experiment B (an agent answers the owner's questions from one real Telegram account): the agent with tg used it as keyword search, and the strong search kinds failed on the 2-second limit. #884, #887.

Why

  • On a store of 950,000 messages, search conversations failed on 12 of 29 sentence questions with query.time-limit; --timeout did not reach it.
  • The skill led to search messages first; the agent ran 71 word searches, one meaning search and no --discover, filtered the source chat out with grep -v, and missed from:me.

What

  • @wirecat/cli-core 0.665.0 (was 0.662.0), @wirecat/cli-base 0.102.1: a slow word half no longer fails search conversations, stop words left out, --timeout reaches it (cli-core#2959, #2972); --discover joins a hyphenated name.
  • Skill, trap 5: for a question in the owner's words, search conversations, then --discover, then key words; the question as written; no filtering by hand; from:me for "I advised"; --timeout after query.time-limit.
  • Regenerated docs/commands.md, contracts/wirecat.json, the test matrix (three shared options from cli-core 0.663–0.665 listed with the cli-core test that drives them). Fragments changes/chore-cli-core-0.665.0.md, changes/fix-search-skill.md.

Testing

  • Locally: typecheck, the full suite (88 files, 2,314 passed), lint, docs:check, parity:check, test:matrix:check.
  • On the real store (tg 0.128.0 on cli-core 0.662.0 + #2959): search conversations found the source in the top 10 in 8 of 29 one-query lookups (was 4), with 0 time-limit failures (was 12); some tg search kind found it in 14 of 29 (was 11). Numbers in search conversations stops at its 2 s time limit on a full-history store #887.

Closes #884

🤖 Generated with Claude Code

…ch and --discover

In Experiment B round 3 an agent with the default skill ran 71 word
searches, one meaning search and no --discover, filtered the source chat
out with grep -v, and missed from:me. The skill now gives the order for a
question in the owner's own words and the traps seen in that run.

Refs #884

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
@leemour
leemour merged commit cf83a23 into main Nov 15, 2026
14 checks passed
@leemour
leemour deleted the fix/search-skill branch November 15, 2026 08:37
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

Agent with the default skill never uses meaning search, which finds the answers it misses

1 participant