feat(recall): recallMulti() for multi-query recall + pluggable taskQueryBuilder

[evilh2019]

Summary

Adds two new capabilities to improve recall quality:

1. recallMulti(queries, options) — multi-query recall

The Recaller class now supports running multiple independent queries in parallel and merging the results. This is a natural generalization of the existing single-query recall() — each query runs its own precise + generalized path internally, and results are merged with dedup via the new mergeMany() static method.

API added:

• recaller.recallMulti(queries: string[], options?) — parallel multi-query recall
• Recaller.merge(a, b) — merge two results (static, was previously private mergeResults)
• Recaller.mergeMany(results[]) — merge any number of results

2. Pluggable taskQueryBuilder — configurable query construction

The DSH integration now captures skill-catalog events (user/message with source.kind === "skill-catalog") and passes them to a configurable taskQueryBuilder that can produce additional recall queries from skill descriptions relevant to the current user message.

Default builder: bigramRelevanceFilter

• Only includes skills whose name or description has a 2-char overlap with the user message
• Excludes lark-* infrastructure skills
• Caps at 500 characters total
• Returns undefined when no skills match (no additional queries)

Config:

taskQueryBuilder?: (
  userQuery: string,
  skills: Array<{ name: string; description: string }>,
) => string[] | undefined;

Custom builders can use embedding similarity, keyword expansion, or LLM-based skill selection — the interface is intentionally minimal and host-agnostic.

Motivation

Single-query recall only searches the user message ("what is the user talking about"), which misses procedural knowledge ("how to do the task"). For example, when a user says "send a task to mama1", the user query finds workflow nodes like dispatch-task-to-mama1-correctly but misses mama1-executor-json-command-send (JSON format requirement that previously caused rejected tasks). The task-oriented query from relevant skill descriptions fills this gap.

Testing

• 148 existing tests pass (1 pre-existing failure in pro-dsh.test.ts due to missing @deepseek-ai/dsh-typert-protocol dependency, unrelated)
• Verified in production with 5 scenarios: KB quality review, A2A dispatch, web restart, plugin development, and code review — dual-recall finds complementary nodes in all cases

Files changed

File	Change
src/recaller/recall.ts	+recallMulti(), +static merge(), +static mergeMany(), extract mergeResultsImpl
dsh.ts	+skillEntries capture, +taskQueryBuilder config, +bigramRelevanceFilter default, dual-recall in system-prompt/assemble
dist/	Compiled JS output