chore: 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

@gemstack/ai-sdk@0.4.0

Minor Changes

• e784b5d: Decouple the cache-backed run stores from @rudderjs/cache (epic: framework-agnostic engine).

  CachedAgentRunStore and CachedSubAgentRunStore no longer lazy-import @rudderjs/cache or read a global CacheRegistry. They now take a required, caller-supplied cache via { cache }, typed against the new exported CacheAdapter contract (a 3-method interface: get / set / forget). Bring any cache (redis, Memcached, a Map, a framework's cache).

  Breaking (0.x):

  • new CachedAgentRunStore() / new CachedSubAgentRunStore() with no cache now throw; pass { cache }. (All in-repo and documented usage already passes it.)
  • Default key prefixes changed from rudderjs:ai:* to gemstack:ai:*. These guard 5-minute-TTL ephemeral run snapshots, so the only effect on upgrade is that in-flight parked runs fall back to "not found" once (they self-heal). Override keyPrefix to keep the old value if needed.

  Also made the Google prompt-cache registry fully neutral (it already accepted a BYO store): it now uses the shared CacheAdapter type and a gemstack:ai:google-cache: key prefix, with no @rudderjs/cache references.

• 97ed299: Remove the @gemstack/ai-sdk/doctor subpath (epic: framework-agnostic engine).

  The AI doctor check registered into @rudderjs/console's doctor registry, coupling the agnostic engine to the Rudder CLI. It has moved to the Rudder binding @rudderjs/ai/doctor (same import path on that package). The ./doctor export is removed here.

  Breaking (0.x): importing @gemstack/ai-sdk/doctor no longer resolves; use @rudderjs/ai/doctor. (The @rudderjs/console peer stays for now — make:agent and the /server provider still use it until they relocate too.)

• 4fa5820: Decouple from @rudderjs/orm (epic: framework-agnostic engine).

  The ORM-backed store subpaths @gemstack/ai-sdk/conversation-orm, /memory-orm, /budget-orm, and /memory-embedding are removed from this package. They imported @rudderjs/orm, coupling the agnostic engine to the Rudder ORM, so they have moved to the Rudder binding @rudderjs/ai under the same subpath names (@rudderjs/ai/conversation-orm, etc.). The @rudderjs/orm peer dependency is dropped.

  Breaking (0.x): update imports from @gemstack/ai-sdk/{conversation-orm,memory-orm,budget-orm,memory-embedding} to @rudderjs/ai/{...}. The relocated implementations are unchanged and still implement the neutral ConversationStore / UserMemory / BudgetStorage contracts, which remain exported from @gemstack/ai-sdk. Non-Rudder apps implement those contracts against their own persistence, or use the in-memory defaults.

• cf28664: Export GoogleCacheRegistry (+ GoogleCacheRegistryOptions / CacheStoreLike) from the main entry, and defaultFixturesDir / readFixture / writeFixture from the ./eval subpath.

  These were gemstack-internal symbols that framework bindings could not reach against a published build. Surfacing them lets a binding construct the Gemini context-cache registry with its own CacheAdapter (new GoogleCacheRegistry({ store })) and lets an ai:eval CLI binding read/write recorded fixtures. Purely additive; unblocks relocating the /server provider and the ai-eval command to the Rudder side.

• 3cb13db: Decouple ImageGenerator.store() / AudioGenerator.store() from @rudderjs/storage (epic: framework-agnostic engine).

  Both .store() helpers no longer lazy-import @rudderjs/storage. They now take a required, caller-supplied storage via a new exported StorageAdapter contract (a one-method interface: put(path, bytes)). Implement it against any blob store (S3, GCS, the filesystem, a framework's storage layer).

  Breaking (0.x): .store(path) is now .store(path, storage). Migrate await ImageGenerator.of(p).store('out.png') to await ImageGenerator.of(p).store('out.png', storage) where storage satisfies StorageAdapter. A Rudder app wraps @rudderjs/storage in a ~3-line adapter.

Patch Changes

• 035050e: Quality pass for ai-sdk: rebrand the error/log message prefix from the migration leftover [Rudder AI] to [ai-sdk] (108 messages across 38 modules), matching the sibling packages' package-name prefix convention, and fix the "file an issue" URL in the Bedrock provider to point at gemstack-land/gemstack. No API or behavior change beyond the message text.

@gemstack/ai-skills@0.2.0

Minor Changes

• 3eb72d4: Quality + docs pass for ai-skills:

  • SkillRegistry.discover() no longer aborts the whole scan when a single SKILL.md is unreadable or malformed; the bad bundle is skipped and the rest are still indexed. Pass discover(root, { onError }) to observe what was skipped. This restores the documented "index hundreds of skills safely" contract.
  • Skill name is now validated at parse time against [a-zA-Z0-9_-] instead of being silently mangled later at compose time, so invalid names fail fast with a clear message.
  • load() on an undiscovered name now lists the available skill names (or hints to call discover() first).
  • loadSkill() distinguishes a missing SKILL.md from an unreadable one.
  • Deduplicated the internal fileExists/isDirectory helpers into a shared module.
  • README: clarified that the tools module is loaded compiled (tools.js/.mjs/.cjs, not tools.ts), and added a direct loadSkill() usage example.

Patch Changes

• Updated dependencies [e784b5d]
• Updated dependencies [97ed299]
• Updated dependencies [4fa5820]
• Updated dependencies [cf28664]
• Updated dependencies [035050e]
• Updated dependencies [3cb13db]
  • @gemstack/ai-sdk@0.4.0

@gemstack/ai-autopilot@0.1.1

Patch Changes

• 81fe17b: Quality + docs pass for ai-autopilot:

  • Supervisor now validates its options at construction (plan must be a function, workers is required, concurrency/maxSubtasks must be positive integers) and run() rejects an empty task, so misconfiguration fails fast with a clear message instead of deep in a planner call.
  • An onEvent callback that throws is now isolated (logged and swallowed) so an observer bug can no longer abort a supervised run.
  • Corrected the SupervisorRun.usage docs: it aggregates dispatched-subtask usage only (the Planner/Synthesizer contracts return data, not usage, so planning/synthesis spend isn't observable).
  • Clarified that maxSubtasks and budget are optional, marked the internal runPool helper @internal, and added JSDoc examples.

• Updated dependencies [e784b5d]

• Updated dependencies [97ed299]

• Updated dependencies [4fa5820]

• Updated dependencies [cf28664]

• Updated dependencies [035050e]

• Updated dependencies [3cb13db]

  • @gemstack/ai-sdk@0.4.0

@gemstack/ai-mcp@0.1.1

Patch Changes

• Updated dependencies [e784b5d]
• Updated dependencies [97ed299]
• Updated dependencies [4fa5820]
• Updated dependencies [cf28664]
• Updated dependencies [035050e]
• Updated dependencies [3cb13db]
  • @gemstack/ai-sdk@0.4.0

@gemstack/mcp@0.2.1

Patch Changes

• 36f56e2: Quality + docs pass for mcp:

  • OAuth: reject an empty bearer token (Authorization: Bearer with no value) up front with a 401 invalid_token instead of forwarding an empty string to verifyToken.
  • Errors thrown when a @Handle dependency fails to resolve now chain the original via { cause }.
  • Documented McpResponse.text/json/error (and when to prefer error() over throwing); neutralized framework-specific wording in the OAuth core docs.
  • README: completed the OAuth 2.1 section (a real jose-based verifyToken, and that oauth2McpMiddleware + registerOAuth2Metadata must both be wired), softened the origin framing.