fix(honcho): pass user_message as search_query for topic-relevant context injection

[qxxaa]

Problem

get_prefetch_context() accepts a user_message parameter but intentionally discards it, with a comment explaining that passing it as search_query "would expose conversation content in server access logs."

This rationale is inconsistent: Honcho already receives and persists every message in full via saveMessages. The content is already in the database. A search query string appearing in an access log adds negligible additional exposure over what is already persisted — and is entirely moot for self-hosted deployments where the operator owns the logs.

Current behaviour

Without search_query, Honcho's peer.context() returns the full peer representation: all observations, derived conclusions, and peer card, in insertion order. When contextTokens is set to limit injection size, the budget is consumed by raw observations first. The higher-signal parts — peer card and synthesised conclusions — appear last and are truncated.

The result is topic-agnostic context injection on every cold start, regardless of what the session is actually about.

Fix

Pass user_message as search_query to _fetch_peer_context() when available. This enables Honcho's semantic retrieval to return conclusions relevant to the current session topic rather than the full observation set.

_fetch_peer_context() already accepts and passes search_query to the Honcho API — this change simply connects the two. The total diff is 2 lines.

Behaviour when user_message is None

user_message or None preserves the existing broad-fetch behaviour when no message is available (e.g. session prewarm before the first turn). No regression for existing deployments.

Files changed

• plugins/memory/honcho/session.py — 2 lines

Testing

1. Configure recallMode: hybrid and injectionFrequency: first-turn in honcho.json
2. Start a fresh session with a topic-specific first message
3. Inspect the injected context — should contain topic-relevant conclusions rather than the full observation dump
4. Verify sessions without a first message (prewarm) still function correctly