fix(user-prompt): inject prompt-matched conclusions instead of stalest representation lines

[alvistar]

Problem

The UserPromptSubmit hook runs a semantic search over the user's prompt (peer.context({ searchQuery, ... })), but then formatCachedContext() injects the first 5 lines of the representation string. The representation merges search hits with frequent/recent conclusions into a single timestamp-ascending list, so the injected lines are always the oldest conclusions in the window — the search results never surface.

Measured on a live workspace (358 conclusions): with and without searchQuery, 4 of 5 injected lines were identical and unrelated to the prompt. In practice every turn re-injects the same stale facts regardless of what the user asks — closely related to the repetition reported in #39 (includeMostFrequent pinning is only half the story; head-of-representation selection is the other half).

A second, compounding issue: extractTopics() falls back to an English-stopword word filter, so non-English prompts produce low-signal queries ("come siamo messi" → word salad).

Fix

• Query matched conclusions explicitly (conclusionScope.query(searchQuery, 5)) in parallel with peer.context() and put them first in the injected block (cached alongside the context, so cache-served turns keep them too)
• Fill the remaining slots with the newest representation lines (timestamp-desc when lines carry a [timestamp] prefix) instead of the oldest, deduped by normalized text
• Drop the English-only stopword fallback; fall back to the raw prompt (truncated to 300 chars), which embeds better for semantic search and keeps non-English prompts working

No new API surface: conclusionScope.query already exists in the SDK; failures degrade gracefully to the previous behavior (.catch(() => [])).

Before / after (live workspace, same prompts)

Prompt (Italian): "come va il deriver di honcho su kubernetes? flux ha applicato la modifica ai worker?"

Before — 5 conclusions about an unrelated design discussion from a previous day, identical to what every other prompt received.

After — all 5 conclusions about the user's Kubernetes/k3s work from the same day.

Prompt (English): "help me create the new jira tickets for the REORG board" → after: all 5 conclusions about the user's recent Jira board work.

Test plan

• tsc --noEmit clean
• Deterministic unit tests: bun test — 16 cases covering search-matched-first ordering, newest-not-oldest fill (regression for the head-of-representation bug), normalized dedup, the 5-slot cap, and the empty-topics return that drives the raw-prompt fallback (src/hooks/user-prompt.test.ts)
• Fresh-fetch path: hook invoked with stdin payload, context cache cleared — injected conclusions match prompt topic (English and Italian prompts)
• Search-failure path: conclusionScope.query rejection falls back to representation-only selection
• Cache-served path: searchMatched survives the context cache round-trip (plain JSON property)

🤖 Generated with Claude Code

Summary by CodeRabbit

Release Notes

• Improvements
  • Enhanced search query generation for more consistent results
  • Optimized context retrieval through concurrent processing for improved performance
  • Refined prioritization of search-matched conclusions in context display
  • Improved chronological ordering of context information for better relevance

[coderabbitai]

Walkthrough

This PR refines the context search and formatting pipeline in the user-prompt hook. Topic extraction is simplified to return only high-signal terms; search operations are parallelized with matched conclusions stored for prioritized formatting; and context assembly now prioritizes prompt-matched conclusions with timestamp-aware deduplication and ordering.

Changes

Context Search and Formatting Improvements

Layer / File(s)	Summary
Topic extraction simplification plugins/honcho/src/hooks/user-prompt.ts	extractTopics() removes the fallback path that extracted non-stopword English words; now returns only high-signal terms (deduped).
Concurrent search execution and matched conclusion storage plugins/honcho/src/hooks/user-prompt.ts	fetchFreshContext() always generates a search query (topics-joined or truncated); context retrieval and conclusion matching run in parallel; matched conclusions are stored in contextResult.searchMatched.
Context formatting with deduplication and timestamp-aware ordering plugins/honcho/src/hooks/user-prompt.ts	formatCachedContext() prioritizes prompt-matched conclusions from context.searchMatched (deduplicated and stripped), then fills remaining slots from context.representation sorted newest-first by timestamp instead of taking the first 5 lines.

Estimated code review effort

🎯 3 (Moderate) | ⏱️ ~20 minutes

Possibly related issues

• UserPromptSubmit re-injects identical conclusions every turn — no dedup causes context bloat #39: The changes implement deduplication and prioritization of prompt-matched conclusions in formatCachedContext() and add concurrent query execution, which directly address the issue's concerns about repeated conclusion reinjection and inefficient context retrieval.

Poem

🐰 A rabbit hops through topics bright,
Now deduped with parallel light,
Fresh conclusions dance in time,
Sorted newest—a context sublime!
High-signal hops, no fallback blight. 🥕

🚥 Pre-merge checks | ✅ 5

✅ Passed checks (5 passed)

Check name	Status	Explanation
Title check	✅ Passed	The title clearly and specifically describes the main fix: prioritizing prompt-matched conclusions over stale representation lines in the user-prompt hook.
Docstring Coverage	✅ Passed	No functions found in the changed files to evaluate docstring coverage. Skipping docstring coverage check.
Linked Issues check	✅ Passed	Check skipped because no linked issues were found for this pull request.
Out of Scope Changes check	✅ Passed	Check skipped because no linked issues were found for this pull request.
Description Check	✅ Passed	Check skipped - CodeRabbit’s high-level summary is enabled.

_{✏️ Tip: You can configure your own custom pre-merge checks in the settings.}

✨ Finishing Touches

🧪 Generate unit tests (beta)

• Create PR with unit tests

---

Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out.

❤️ Share

• X
• Mastodon
• Reddit
• LinkedIn

_{Comment @coderabbitai help to get the list of available commands and usage tips.}