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

@context-chef/core@3.5.0

Minor Changes

• bb35e1f Thanks @MyPrototypeWhat! - feat: configurable memory placement (MemoryConfig.memoryPlacement) + Gemini parallel-tool-call fix

  memoryPlacement: 'after_system' | 'before_history_tail'

  New MemoryConfig.memoryPlacement controls where the volatile <memory> data block lands in the compiled payload. Default 'after_system' is byte-for-byte compatible with previous behavior. Opt into 'before_history_tail' when you use Anthropic prompt caching with cache breakpoints on history — under the default placement, every memory mutation invalidates the entire history cache because Anthropic / Gemini adapters extract every role: 'system' message into the top-level system parameter, and downstream cache breakpoints hash that block.

  const chef = new ContextChef({
    memory: {
      store: new VFSMemoryStore(dir),
      memoryPlacement: "before_history_tail",
    },
  });

  Placement	Top of sandwich	Last user message	Cache behavior
  'after_system' (default)	INSTRUCTION + <memory> data combined into one role: 'system'	untouched	Memory text rides into the system parameter — cache breakpoints downstream hash it and miss on every memory mutation
  'before_history_tail'	INSTRUCTION only (stable, cacheable)	<memory> data appended via the existing tail-injection mechanism	Memory text never enters the system parameter — earlier cache breakpoints survive memory mutations on every provider

  When dynamicStatePlacement: 'last_user' is also active, the tail order inside the last user message is: original content → <memory> → <dynamic_state> → <implicit_context> → anchor line.

  Caveat — when compile() is invoked mid-agent-loop (history tail is a tool turn awaiting interpretation), "most recent user" walks back past the tool result(s) to the user turn that kicked off the current tool sequence. The injection lands mid-conversation rather than at the absolute tail. Cache breakpoints placed AFTER that user turn (typical: end-of-history assistant) will see the modified user content and miss. To preserve cache in tool-ending tails, place breakpoints BEFORE the user turn that started the active tool sequence. Same placement convention as setDynamicState({ placement: 'last_user' }).

  Additions:

  • Memory.placement: MemoryPlacement — public getter exposing the configured placement for observability.
  • MemoryPlacement type re-exported from @context-chef/core.

  Gemini adapter: parallel-tool-call merge (latent bug fix)

  GeminiAdapter.compile() now collapses consecutive same-role Content entries before returning the payload. Previously, any IR sequence with parallel tool calls would emit non-canonical Gemini shape — three parallel tool IR messages map to three role: 'user' Contents (Gemini's tool-result convention), which is rejected by generateContent with:

  400 INVALID_ARGUMENT: Please ensure that multiturn requests alternate
  between user and model
  

  The canonical Gemini shape for parallel function responses is ONE user Content with multiple functionResponse parts (Vertex AI docs; same shape the Google Gen AI SDK constructs in its automatic function calling path). The new merge produces exactly that shape and additionally guards against any other source of adjacent same-role contents (e.g. stripped-thinking empty model turns, caller-supplied malformed IR).

  Implementation notes:

  • Runs AFTER prefill degradation so the trailing-model pop sees an unmerged view (a pre-merge would conflate the prefill candidate with the preceding model turn).
  • Lazy-allocation fast path: when the input is already strictly alternating (the common case for normal conversations), the original array is returned by reference — zero allocation. Allocates on the first detected merge point.
  • Pure function: never mutates the input. Result envelopes and parts arrays are fresh; inner Part objects are treated as immutable leaves.
  • Anthropic auto-merges server-side and OpenAI tolerates adjacent same-role, so this merge lives exclusively in the Gemini adapter — mirroring @tanstack/ai-gemini's mergeConsecutiveSameRoleMessages and LangChain's langchain-google-genai.

  Users on @context-chef/ai-sdk-middleware and @context-chef/tanstack-ai are unaffected by the latent bug — those paths already produce canonical Gemini shape via existing batching (toAISDK's consecutive-tool-message coalescing, TanStack's own merge).

  Internal API tightening — Assembler.compile() / AssembleOptions marked @internal

  AssembleOptions is no longer re-exported from the package barrel. Both AssembleOptions (the interface) and Assembler.compile() (the instance method) now carry @internal JSDoc tags so TypeScript / TypeDoc tool chains exclude them from generated .d.ts and documentation.

  This makes explicit what was already true in practice: Assembler.compile() is implementation glue for ContextChef.compile() and has no external use case. The Assembler class itself remains exported because its static helpers — orderKeysDeterministically and stringifyPayload — are genuine public utilities for callers that need to hash payloads consistently with the compile pipeline (cache-aware logging, custom adapters, etc.).

  Internal signature changes (under @internal):

  • AssembleOptions.dynamicStateXml → AssembleOptions.tailXml — the field is general-purpose tail injection, not dynamic-state-specific.
  • AssembleOptions.placement removed — the assembler no longer gates injection on a placement enum; ContextChef.compile() composes the tail stitch and supplies it only when injection should happen.
  • The hardcoded "Above is the current system state..." anchor line moved out of injectIntoLastUser and into ContextChef.compile() — appended only when dynamic state or implicit context is actually present in the tail.

  Migration — external callers of import { type AssembleOptions } from '@context-chef/core' (rare; never advertised) will see a TypeScript error after upgrade. Resolution: use ContextChef.compile() rather than invoking the assembler directly. If you have a genuine use case for direct assembler access, please open an issue.

  Other additions

  • Prompts.MEMORY_BLOCK_HEADER: 'You recall the following from previous conversations:' — extracted constant referenced by both Prompts.getMemoryBlock and ContextChef.compile()'s anchor-suppression logic. Makes the self-anchor contract explicit so future changes to the memory block prefix can't silently break anchor suppression for memory-only tail injections.

@context-chef/ai-sdk-middleware@1.3.5

Patch Changes

• Updated dependencies [bb35e1f]:
  • @context-chef/core@3.5.0

@context-chef/tanstack-ai@0.4.5

Patch Changes

• Updated dependencies [bb35e1f]:
  • @context-chef/core@3.5.0