Version Packages

[github-actions]

This PR was opened by the Changesets release GitHub action. When you're ready to do a release, you can merge this and the packages will be published to npm automatically. If you're not ready to do a release yet, that's fine, whenever you add more changesets to main, this PR will be updated.

Releases

@harpua/agent-tools@0.8.1

Patch Changes

• 693f878: prepareChunks(markdown, options?) — export the pure chunk-prep half of
  ingest() (field report ci(release): pin npm to 11.5.1 (npm@latest broke provenance/sigstore) #18): chunk → sanitize → junk-filter → embed-text
  formatting, with no embedding or storage. A consumer running its own
  embed/upsert path into a separate collection no longer needs a hand-rolled
  duplicate of embeddingInputFor and the alnum junk floor just to keep chunk
  geometry in sync with ingest() — they call prepareChunks directly and
  get the same { text, embedText, chunkIndex, startLine, endLine, headingTrail } per chunk that ingest() embeds and stores.

  Options are the same four chunking knobs ingest() exposes
  (maxChunkChars, minAlnumChars, embedHeadingTrail, sanitize), same
  defaults, same strict zod validation (unknown keys throw). ingest() itself
  now composes over prepareChunks — the embed/upsert half is all that's left
  inline — so there is exactly one code path for chunk geometry and
  embedding-input formatting; direct prepareChunks callers and ingest()
  can never drift apart.

• e3728ce: SkillRegistry skip visibility (field report Version Packages #17): a consumer's skill silently
  tripped the 16KB SKILL.md body cap and cost an hour of misdiagnosis because
  the only trace was a generic onWarn string — skips were a log, not data.
  rescan() now also returns skippedSkills: { name, reason }[] (one entry per
  skip, reason the same detail sent to onWarn minus the skills: prefix),
  so a caller can actually act on a skip instead of grepping console output.
  Frontmatter failures — a bad name vs. an empty description — previously
  folded into one generic "no valid frontmatter" reason; the structured
  reason (never the onWarn text, which stays byte-identical for existing
  parsers) now folds in the first zod issue so the two are distinguishable.
  Separately, a directory with no SKILL.md at all (a misnamed skill.md?)
  now gets a single onWarn line instead of skipping silently — it still
  isn't counted in skipped/skippedSkills, since it was never a skill
  candidate. renderSkillMenu also grows an optional opts.header so a
  caller can swap the leading TOC line; omitting it reproduces the exact
  current bytes, keeping the prompt cache warm for everyone who doesn't need it.

@harpua/langgraph@0.6.2

Patch Changes

• 1e57a51: askUserTool(options?) — a new model-callable tool that lets the model ask
  the user one or more typed questions and get answers back as the tool result
  mid-turn, instead of emitting a question as prose for a host to reconstruct
  after the fact. It joins requireApproval as the second member of the
  interrupt vocabulary: it pauses with { type: "ask_user_request", intro?, questions }, and the host resumes with { answers: [...] } (index-aligned
  to questions) or { dismissed: true, reason? } for headless hosts.

  Ships with a flat default question preset (prompt, inputType:
  select/multi_select/boolean/free_text, options?) that normalizes
  select/multi_select choices (trim, drop empties, dedupe, coerce to
  free_text below 2 survivors) — or bring your own questionSchema (any zod
  type extending { prompt: string }) and serializeAnswers.