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

@dawn-ai/cli@0.3.0

Minor Changes

• 8133553: Add opt-in conversation summarization (Phase 3 sub-project 6b). When a thread's history exceeds a token threshold, the agent is fed a condensed view — a running summary of older turns plus the most recent turns verbatim — while the full history stays intact in the checkpoint. This is non-destructive: summarization runs as a LangGraph preModelHook that returns llmInputMessages for the turn only and never rewrites saved messages, so GET /threads/:id/state, resume, and restart always see the complete history (and there is no tool-call/result pairing hazard).

  Enable it in dawn.config.ts:

  export default {
    summarization: {
      enabled: true, // default false
      maxTokens: 12_000, // threshold over which older turns are summarized
      keepRecentTurns: 6, // most-recent turns kept verbatim
      // model defaults to the route's model
      // tokenCounter defaults to a lazy gpt-tokenizer (o200k_base) counter
      // summarize defaults to a built-in single-LLM-call running-summary fold
    },
  };

  Both the token counter and the summarizer are pluggable (tokenCounter, summarize). The running summary is cached in agent state and refreshed incrementally — each turn folds only the newly-aged messages, so cost stays bounded. The turn-boundary split is pairing-safe (a tool-call message is never separated from its results). When summarization is disabled (the default), behavior is unchanged and gpt-tokenizer is never loaded. If the summarizer call fails on a given turn, the agent falls back to the full history for that turn rather than failing the run.

• 027b1cc: Add tool-output offloading. When a tool returns output larger than toolOutput.offloadThresholdChars (default 40,000), the full payload is written to workspace/tool-outputs/ and the in-context ToolMessage is replaced with a preview+pointer stub; the agent retrieves the full content with the existing readFile tool (which bypasses the size cap for tool-outputs/ paths). Active automatically when a workspace exists. The directory is bounded by a size + TTL cap (defaults 256MB / 3h) with throttled evict-on-write and LRU-by-access eviction (readFile bumps mtime for tool-outputs/ files). Large content never enters message state, so there is no tool-call/result pairing hazard. Configurable via dawn.config.ts toolOutput. The FilesystemBackend interface gains optional statFile/removeFile/touchFile/mkdir methods and an optional per-call maxBytes override on readFile.

Patch Changes

• 55b69f0: Fix tool-output offloading so retrieval tools are exempt. Previously the workspace readFile tool — the very tool the agent uses to read back an offloaded output — had its own (large) result offloaded again, replacing it with a second pointer stub. The agent could never see the retrieved content. Retrieval/inspection tools (readFile, listDir) are now never offloaded; the new dawn.config.ts toolOutput.noOffloadTools option adds further exemptions (merged with the always-exempt built-ins). Found by a live-API smoke test.
• Updated dependencies [30db6ed]
• Updated dependencies [55b69f0]
• Updated dependencies [2e3bc8d]
• Updated dependencies [8133553]
• Updated dependencies [027b1cc]
  • @dawn-ai/langchain@0.3.0
  • @dawn-ai/core@0.3.0
  • @dawn-ai/langgraph@0.3.0
  • @dawn-ai/permissions@0.1.8
  • @dawn-ai/sqlite-storage@0.2.0

@dawn-ai/core@0.3.0

Minor Changes

• 8133553: Add opt-in conversation summarization (Phase 3 sub-project 6b). When a thread's history exceeds a token threshold, the agent is fed a condensed view — a running summary of older turns plus the most recent turns verbatim — while the full history stays intact in the checkpoint. This is non-destructive: summarization runs as a LangGraph preModelHook that returns llmInputMessages for the turn only and never rewrites saved messages, so GET /threads/:id/state, resume, and restart always see the complete history (and there is no tool-call/result pairing hazard).

  Enable it in dawn.config.ts:

  export default {
    summarization: {
      enabled: true, // default false
      maxTokens: 12_000, // threshold over which older turns are summarized
      keepRecentTurns: 6, // most-recent turns kept verbatim
      // model defaults to the route's model
      // tokenCounter defaults to a lazy gpt-tokenizer (o200k_base) counter
      // summarize defaults to a built-in single-LLM-call running-summary fold
    },
  };

  Both the token counter and the summarizer are pluggable (tokenCounter, summarize). The running summary is cached in agent state and refreshed incrementally — each turn folds only the newly-aged messages, so cost stays bounded. The turn-boundary split is pairing-safe (a tool-call message is never separated from its results). When summarization is disabled (the default), behavior is unchanged and gpt-tokenizer is never loaded. If the summarizer call fails on a given turn, the agent falls back to the full history for that turn rather than failing the run.

• 027b1cc: Add tool-output offloading. When a tool returns output larger than toolOutput.offloadThresholdChars (default 40,000), the full payload is written to workspace/tool-outputs/ and the in-context ToolMessage is replaced with a preview+pointer stub; the agent retrieves the full content with the existing readFile tool (which bypasses the size cap for tool-outputs/ paths). Active automatically when a workspace exists. The directory is bounded by a size + TTL cap (defaults 256MB / 3h) with throttled evict-on-write and LRU-by-access eviction (readFile bumps mtime for tool-outputs/ files). Large content never enters message state, so there is no tool-call/result pairing hazard. Configurable via dawn.config.ts toolOutput. The FilesystemBackend interface gains optional statFile/removeFile/touchFile/mkdir methods and an optional per-call maxBytes override on readFile.

Patch Changes

• 55b69f0: Fix tool-output offloading so retrieval tools are exempt. Previously the workspace readFile tool — the very tool the agent uses to read back an offloaded output — had its own (large) result offloaded again, replacing it with a second pointer stub. The agent could never see the retrieved content. Retrieval/inspection tools (readFile, listDir) are now never offloaded; the new dawn.config.ts toolOutput.noOffloadTools option adds further exemptions (merged with the always-exempt built-ins). Found by a live-API smoke test.
• 2e3bc8d: Fix tool-input schema extraction for standalone literal types. A single string-literal type (e.g. a discriminated-union discriminant like by: "date") was not recognized as an enum (only multi-member literal unions were), so it fell through to object extraction and was misread as an object carrying String.prototype methods (charAt, toString, …). This produced a bogus schema that rejected the correct argument, breaking every discriminated/object-union tool parameter end-to-end. Standalone string/number/boolean literals now extract correctly, and object extraction is guarded to genuine object types. Found by a live-API smoke test.
• Updated dependencies [027b1cc]
  • @dawn-ai/workspace@0.2.0
  • @dawn-ai/sdk@0.3.0
  • @dawn-ai/permissions@0.1.8
  • @dawn-ai/sqlite-storage@0.2.0

@dawn-ai/langchain@0.3.0

Minor Changes

• 8133553: Add opt-in conversation summarization (Phase 3 sub-project 6b). When a thread's history exceeds a token threshold, the agent is fed a condensed view — a running summary of older turns plus the most recent turns verbatim — while the full history stays intact in the checkpoint. This is non-destructive: summarization runs as a LangGraph preModelHook that returns llmInputMessages for the turn only and never rewrites saved messages, so GET /threads/:id/state, resume, and restart always see the complete history (and there is no tool-call/result pairing hazard).

  Enable it in dawn.config.ts:

  export default {
    summarization: {
      enabled: true, // default false
      maxTokens: 12_000, // threshold over which older turns are summarized
      keepRecentTurns: 6, // most-recent turns kept verbatim
      // model defaults to the route's model
      // tokenCounter defaults to a lazy gpt-tokenizer (o200k_base) counter
      // summarize defaults to a built-in single-LLM-call running-summary fold
    },
  };

  Both the token counter and the summarizer are pluggable (tokenCounter, summarize). The running summary is cached in agent state and refreshed incrementally — each turn folds only the newly-aged messages, so cost stays bounded. The turn-boundary split is pairing-safe (a tool-call message is never separated from its results). When summarization is disabled (the default), behavior is unchanged and gpt-tokenizer is never loaded. If the summarizer call fails on a given turn, the agent falls back to the full history for that turn rather than failing the run.

• 027b1cc: Add tool-output offloading. When a tool returns output larger than toolOutput.offloadThresholdChars (default 40,000), the full payload is written to workspace/tool-outputs/ and the in-context ToolMessage is replaced with a preview+pointer stub; the agent retrieves the full content with the existing readFile tool (which bypasses the size cap for tool-outputs/ paths). Active automatically when a workspace exists. The directory is bounded by a size + TTL cap (defaults 256MB / 3h) with throttled evict-on-write and LRU-by-access eviction (readFile bumps mtime for tool-outputs/ files). Large content never enters message state, so there is no tool-call/result pairing hazard. Configurable via dawn.config.ts toolOutput. The FilesystemBackend interface gains optional statFile/removeFile/touchFile/mkdir methods and an optional per-call maxBytes override on readFile.

Patch Changes

• 30db6ed: Offloaded tool-output filenames are now deterministic — keyed on the originating tool_call_id (with a content-hash fallback when absent) instead of timestamp+random. This makes offloaded paths stable and traceable and enables deterministic agent e2e tests. The openai chat model now also honors OPENAI_BASE_URL, allowing a local mock provider (used by the new CI-safe aimock-based agent e2e regression tests for the discriminated-union tool-input and tool-output-offload-retrieval paths).
• Updated dependencies [55b69f0]
• Updated dependencies [2e3bc8d]
• Updated dependencies [8133553]
• Updated dependencies [027b1cc]
  • @dawn-ai/core@0.3.0
  • @dawn-ai/workspace@0.2.0
  • @dawn-ai/sdk@0.3.0

@dawn-ai/workspace@0.2.0

Minor Changes

• 027b1cc: Add tool-output offloading. When a tool returns output larger than toolOutput.offloadThresholdChars (default 40,000), the full payload is written to workspace/tool-outputs/ and the in-context ToolMessage is replaced with a preview+pointer stub; the agent retrieves the full content with the existing readFile tool (which bypasses the size cap for tool-outputs/ paths). Active automatically when a workspace exists. The directory is bounded by a size + TTL cap (defaults 256MB / 3h) with throttled evict-on-write and LRU-by-access eviction (readFile bumps mtime for tool-outputs/ files). Large content never enters message state, so there is no tool-call/result pairing hazard. Configurable via dawn.config.ts toolOutput. The FilesystemBackend interface gains optional statFile/removeFile/touchFile/mkdir methods and an optional per-call maxBytes override on readFile.

create-dawn-ai-app@0.3.0

Patch Changes

• @dawn-ai/devkit@0.3.0

@dawn-ai/langgraph@0.3.0

Patch Changes

• @dawn-ai/sdk@0.3.0

@dawn-ai/vite-plugin@0.3.0

Patch Changes

• Updated dependencies [55b69f0]
• Updated dependencies [2e3bc8d]
• Updated dependencies [8133553]
• Updated dependencies [027b1cc]
  • @dawn-ai/core@0.3.0

@dawn-ai/config-biome@0.3.0

@dawn-ai/config-typescript@0.3.0

@dawn-ai/devkit@0.3.0

@dawn-ai/sdk@0.3.0

@dawn-example/chat-server@0.0.2

Patch Changes

• Updated dependencies [30db6ed]
• Updated dependencies [55b69f0]
• Updated dependencies [2e3bc8d]
• Updated dependencies [8133553]
• Updated dependencies [027b1cc]
  • @dawn-ai/langchain@0.3.0
  • @dawn-ai/core@0.3.0
  • @dawn-ai/cli@0.3.0
  • @dawn-ai/sdk@0.3.0

[vercel]

The latest updates on your projects. Learn more about Vercel for GitHub.

Project	Deployment	Actions	Updated (UTC)
dawnai	Ready	Preview, Comment	Jun 4, 2026 9:26pm