ci: release 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

@moonshot-ai/kimi-code@0.29.0

Minor Changes

• #1992 a8f1ca3 Thanks @RealKai42! - Support selecting a thinking effort level from ACP clients: the thinking picker now lists the current model's declared levels (for example off / low / medium / high) instead of only an on/off toggle. Use the thinking selector in your ACP client (e.g. Zed) to pick a level; the legacy on/off values keep working.

• #1735 ce0e3ce Thanks @7Sageer! - Let custom agent files restrict which sub-agent types they may delegate to (v2 engine only).

• #1735 ce0e3ce Thanks @7Sageer! - Support custom agents defined as Markdown files with frontmatter, usable as the main agent or a sub-agent (v2 engine only).

• #1735 ce0e3ce Thanks @7Sageer! - Add global tool gating to constrain which tools agents may use, with a per-session override (v2 engine only).

• #2012 d67a200 Thanks @sailist! - Add a GET /api/v1/fs:content server endpoint that serves any file on the host by absolute path as raw content with Content-Type, ETag, and Range support.

• #1735 ce0e3ce Thanks @7Sageer! - Support overriding the default main-agent system prompt with a user-level file for every session (v2 engine only).

Patch Changes

• #1997 74da87a Thanks @sailist! - Add agent.created and agent.disposed events to the server session event stream, and expose each agent's disposal time in the transcript API.

• #2030 ec88d35 Thanks @RealKai42! - Fix a set of small correctness issues on top of the catalog metadata work: configured efforts (config or the KIMI_MODEL_THINKING_EFFORT env override) are now normalized instead of being sent upstream as invalid values; a model's declared input limit can no longer exceed its effective context window, and the clamp now copies the record instead of mutating the user's config in place; context-usage percentages share one denominator (the effective input cap) across status endpoints, clamped to 1 where the wire schema bounds it while event streams keep the documented raw overflow signal; a provider-observed smaller context window now actually wins over the catalog's declared input cap during overflow recovery; per-model endpoints declared with an unrecognized override SDK are preserved via the OpenAI-compatible fallback, while known proprietary SDKs stay refused; and the model inspector attributes input-limit fields to their actual config, override, or clamp provenance.

• #2015 b5efba7 Thanks @RealKai42! - Import many more providers from the models.dev catalog: vendor SDKs like xai and openrouter now import instead of being refused (with a "guessed" note), deprecated and alpha models are filtered out, per-model gateway protocol and endpoint overrides are honored, and context limits are correct (input limit for compaction, total window for completion). Imports lacking a usable endpoint now ask for one via --base-url or a prompt.

• #1993 37eda4e Thanks @RealKai42! - Add environment variable overrides for agent loop and background task limits. Set KIMI_LOOP_MAX_STEPS_PER_TURN, KIMI_LOOP_MAX_RETRIES_PER_STEP, or KIMI_CODE_BACKGROUND_MAX_RUNNING_TASKS to take priority over the [loop_control] and [background] config.

• #1993 37eda4e Thanks @RealKai42! - Fix config environment overrides (such as KIMI_IMAGE_MAX_EDGE_PX or KIMI_SUBAGENT_TIMEOUT_MS) being persisted into config.toml by config API writes while the env var is set, and keeping the old value after the env var is changed to an invalid value or removed.

• #1970 6dd4fd3 Thanks @sailist! - Fix cancelled model requests being wrapped as retryable provider errors, so interrupting a request no longer triggers silent retries.

• #1970 6dd4fd3 Thanks @sailist! - Send the session prompt cache key to OpenAI and OpenAI Responses providers, restoring provider-side prompt cache affinity that previously only reached Kimi and Anthropic.

• #1968 71bcfba Thanks @RealKai42! - Fix sessions getting stuck on every turn with a provider "message must not be empty" error after a content-filtered response.

• #2022 154e082 Thanks @wbxl2000! - web: Show transparent images over a checkerboard canvas so white and black content stays visible in both light and dark themes.

• #1990 115b096 Thanks @liruifengv! - Fix goal mode continuation prompts leaking into the transcript when resuming a session.

• #1970 6dd4fd3 Thanks @sailist! - Rework the model wire layer in the experimental v2 engine into a small set of protocol bases plus declarative provider trait definitions, so adding a provider no longer means copying adapter code, and per-turn request intent (cache key, thinking effort, sampling) flows as request parameters instead of cloned model objects. The never-functional [platforms] config section and the provider.platformId field are removed; credential resolution is now a two-layer model → provider lookup.

• #1976 e458323 Thanks @liruifengv! - Improve TUI performance and resume speed for long-running sessions.

• #1991 92576e4 Thanks @7Sageer! - Reconnect a dropped MCP server connection automatically when one of its tools is called, and retry the call once.

• #1970 6dd4fd3 Thanks @sailist! - Add read-only model resolution inspection and a live connectivity probe to the server's RPC surface, reporting per-field value provenance (config, override, builtin, env, synthesized) for internal debugging tools.

• #2015 b5efba7 Thanks @RealKai42! - Fix thinking levels being offered for models that do not support them (e.g. phantom levels on Kimi K3): levels now come from each model's declared capabilities. Models that cannot disable reasoning (e.g. gpt-5) no longer offer an Off option, and turning thinking Off on models that support it (e.g. xai grok) now truly disables reasoning.

• #1735 ce0e3ce Thanks @7Sageer! - Warn when a tool allow/deny list entry can never match any tool, for example a misspelled name (v2 engine only).

• #2005 a3699dd Thanks @7Sageer! - Add an active flag to each tool in the server's tool listing API.

• #1995 73eb5f8 Thanks @liruifengv! - Remove red coloring from syntax highlighting in code previews and markdown code blocks.

• #2014 576d650 Thanks @liruifengv! - Add a reminder for third-party install sources to use the official installer in the update prompt.

@moonshot-ai/agent-core-v2@0.2.0

Minor Changes

• #1735 ce0e3ce Thanks @7Sageer! - Let custom agent files restrict which sub-agent types they may delegate to (v2 engine only).

• #1735 ce0e3ce Thanks @7Sageer! - Support custom agents defined as Markdown files with frontmatter, usable as the main agent or a sub-agent (v2 engine only).

• #1735 ce0e3ce Thanks @7Sageer! - Add global tool gating to constrain which tools agents may use, with a per-session override (v2 engine only).

• #1735 ce0e3ce Thanks @7Sageer! - Support overriding the default main-agent system prompt with a user-level file for every session (v2 engine only).

Patch Changes

• #2030 ec88d35 Thanks @RealKai42! - Fix a set of small correctness issues on top of the catalog metadata work: configured efforts (config or the KIMI_MODEL_THINKING_EFFORT env override) are now normalized instead of being sent upstream as invalid values; a model's declared input limit can no longer exceed its effective context window, and the clamp now copies the record instead of mutating the user's config in place; context-usage percentages share one denominator (the effective input cap) across status endpoints, clamped to 1 where the wire schema bounds it while event streams keep the documented raw overflow signal; a provider-observed smaller context window now actually wins over the catalog's declared input cap during overflow recovery; per-model endpoints declared with an unrecognized override SDK are preserved via the OpenAI-compatible fallback, while known proprietary SDKs stay refused; and the model inspector attributes input-limit fields to their actual config, override, or clamp provenance.

• #1993 37eda4e Thanks @RealKai42! - Add environment variable overrides for agent loop and background task limits. Set KIMI_LOOP_MAX_STEPS_PER_TURN, KIMI_LOOP_MAX_RETRIES_PER_STEP, or KIMI_CODE_BACKGROUND_MAX_RUNNING_TASKS to take priority over the [loop_control] and [background] config.

• #1993 37eda4e Thanks @RealKai42! - Fix config environment overrides (such as KIMI_IMAGE_MAX_EDGE_PX or KIMI_SUBAGENT_TIMEOUT_MS) being persisted into config.toml by config API writes while the env var is set, and keeping the old value after the env var is changed to an invalid value or removed.

• #1968 71bcfba Thanks @RealKai42! - Fix sessions getting stuck on every turn with a provider "message must not be empty" error after a content-filtered response.

• #2015 b5efba7 Thanks @RealKai42! - Fix thinking levels being offered for models that do not support them (e.g. phantom levels on Kimi K3): levels now come from each model's declared capabilities. Models that cannot disable reasoning (e.g. gpt-5) no longer offer an Off option, and turning thinking Off on models that support it (e.g. xai grok) now truly disables reasoning.

• #1735 ce0e3ce Thanks @7Sageer! - Warn when a tool allow/deny list entry can never match any tool, for example a misspelled name (v2 engine only).

• Updated dependencies [ce0e3ce, ce0e3ce, a3699dd]:

  • @moonshot-ai/protocol@0.5.0

@moonshot-ai/kap-server@0.1.0

Minor Changes

• #1735 ce0e3ce Thanks @7Sageer! - Support custom agents defined as Markdown files with frontmatter, usable as the main agent or a sub-agent (v2 engine only).

• #1735 ce0e3ce Thanks @7Sageer! - Add global tool gating to constrain which tools agents may use, with a per-session override (v2 engine only).

Patch Changes

• #2005 a3699dd Thanks @7Sageer! - Add an active flag to each tool in the server's tool listing API.

• Updated dependencies [ce0e3ce, ec88d35, 37eda4e, 37eda4e, ce0e3ce, 71bcfba, ce0e3ce, ce0e3ce, b5efba7, ce0e3ce]:

  • @moonshot-ai/agent-core-v2@0.2.0

@moonshot-ai/klient@0.1.0

Minor Changes

• #1735 ce0e3ce Thanks @7Sageer! - Add global tool gating to constrain which tools agents may use, with a per-session override (v2 engine only).

Patch Changes

• Updated dependencies [ce0e3ce, ec88d35, 37eda4e, 37eda4e, ce0e3ce, 71bcfba, ce0e3ce, ce0e3ce, b5efba7, ce0e3ce]:
  • @moonshot-ai/agent-core-v2@0.2.0

@moonshot-ai/kimi-code-sdk@0.14.0

Minor Changes

• #1735 ce0e3ce Thanks @7Sageer! - Add global tool gating to constrain which tools agents may use, with a per-session override (v2 engine only).

Patch Changes

• #2030 ec88d35 Thanks @RealKai42! - Fix a set of small correctness issues on top of the catalog metadata work: configured efforts (config or the KIMI_MODEL_THINKING_EFFORT env override) are now normalized instead of being sent upstream as invalid values; a model's declared input limit can no longer exceed its effective context window, and the clamp now copies the record instead of mutating the user's config in place; context-usage percentages share one denominator (the effective input cap) across status endpoints, clamped to 1 where the wire schema bounds it while event streams keep the documented raw overflow signal; a provider-observed smaller context window now actually wins over the catalog's declared input cap during overflow recovery; per-model endpoints declared with an unrecognized override SDK are preserved via the OpenAI-compatible fallback, while known proprietary SDKs stay refused; and the model inspector attributes input-limit fields to their actual config, override, or clamp provenance.

• #2015 b5efba7 Thanks @RealKai42! - Import many more providers from the models.dev catalog: vendor SDKs like xai and openrouter now import instead of being refused (with a "guessed" note), deprecated and alpha models are filtered out, per-model gateway protocol and endpoint overrides are honored, and context limits are correct (input limit for compaction, total window for completion). Imports lacking a usable endpoint now ask for one via --base-url or a prompt.

• #1976 e458323 Thanks @liruifengv! - Improve TUI performance and resume speed for long-running sessions.

@moonshot-ai/protocol@0.5.0

Minor Changes

• #1735 ce0e3ce Thanks @7Sageer! - Support custom agents defined as Markdown files with frontmatter, usable as the main agent or a sub-agent (v2 engine only).

• #1735 ce0e3ce Thanks @7Sageer! - Add global tool gating to constrain which tools agents may use, with a per-session override (v2 engine only).

Patch Changes

• #2005 a3699dd Thanks @7Sageer! - Add an active flag to each tool in the server's tool listing API.

kimi-code@0.6.5

Patch Changes

• Updated dependencies [ec88d35, b5efba7, ce0e3ce, e458323]:
  • @moonshot-ai/kimi-code-sdk@0.14.0

@moonshot-ai/acp-adapter@0.3.5

Patch Changes

• Updated dependencies [ec88d35, b5efba7, 37eda4e, 71bcfba, ce0e3ce, e458323, b5efba7]:
  • @moonshot-ai/kimi-code-sdk@0.14.0
  • @moonshot-ai/agent-core@0.15.6

@moonshot-ai/agent-core@0.15.6

Patch Changes

• #2030 ec88d35 Thanks @RealKai42! - Fix a set of small correctness issues on top of the catalog metadata work: configured efforts (config or the KIMI_MODEL_THINKING_EFFORT env override) are now normalized instead of being sent upstream as invalid values; a model's declared input limit can no longer exceed its effective context window, and the clamp now copies the record instead of mutating the user's config in place; context-usage percentages share one denominator (the effective input cap) across status endpoints, clamped to 1 where the wire schema bounds it while event streams keep the documented raw overflow signal; a provider-observed smaller context window now actually wins over the catalog's declared input cap during overflow recovery; per-model endpoints declared with an unrecognized override SDK are preserved via the OpenAI-compatible fallback, while known proprietary SDKs stay refused; and the model inspector attributes input-limit fields to their actual config, override, or clamp provenance.

• #1993 37eda4e Thanks @RealKai42! - Add environment variable overrides for agent loop and background task limits. Set KIMI_LOOP_MAX_STEPS_PER_TURN, KIMI_LOOP_MAX_RETRIES_PER_STEP, or KIMI_CODE_BACKGROUND_MAX_RUNNING_TASKS to take priority over the [loop_control] and [background] config.

• #1968 71bcfba Thanks @RealKai42! - Fix sessions getting stuck on every turn with a provider "message must not be empty" error after a content-filtered response.

• #1735 ce0e3ce Thanks @7Sageer! - Add global tool gating to constrain which tools agents may use, with a per-session override (v2 engine only).

• #1976 e458323 Thanks @liruifengv! - Improve TUI performance and resume speed for long-running sessions.

• #2015 b5efba7 Thanks @RealKai42! - Fix thinking levels being offered for models that do not support them (e.g. phantom levels on Kimi K3): levels now come from each model's declared capabilities. Models that cannot disable reasoning (e.g. gpt-5) no longer offer an Off option, and turning thinking Off on models that support it (e.g. xai grok) now truly disables reasoning.

• Updated dependencies [ec88d35, b5efba7, ce0e3ce, 71bcfba, ce0e3ce, b5efba7, a3699dd]:

  • @moonshot-ai/kosong@0.5.5
  • @moonshot-ai/protocol@0.5.0

@moonshot-ai/kosong@0.5.5

Patch Changes

• #2030 ec88d35 Thanks @RealKai42! - Fix a set of small correctness issues on top of the catalog metadata work: configured efforts (config or the KIMI_MODEL_THINKING_EFFORT env override) are now normalized instead of being sent upstream as invalid values; a model's declared input limit can no longer exceed its effective context window, and the clamp now copies the record instead of mutating the user's config in place; context-usage percentages share one denominator (the effective input cap) across status endpoints, clamped to 1 where the wire schema bounds it while event streams keep the documented raw overflow signal; a provider-observed smaller context window now actually wins over the catalog's declared input cap during overflow recovery; per-model endpoints declared with an unrecognized override SDK are preserved via the OpenAI-compatible fallback, while known proprietary SDKs stay refused; and the model inspector attributes input-limit fields to their actual config, override, or clamp provenance.

• #2015 b5efba7 Thanks @RealKai42! - Import many more providers from the models.dev catalog: vendor SDKs like xai and openrouter now import instead of being refused (with a "guessed" note), deprecated and alpha models are filtered out, per-model gateway protocol and endpoint overrides are honored, and context limits are correct (input limit for compaction, total window for completion). Imports lacking a usable endpoint now ask for one via --base-url or a prompt.

• #1968 71bcfba Thanks @RealKai42! - Fix sessions getting stuck on every turn with a provider "message must not be empty" error after a content-filtered response.

• #2015 b5efba7 Thanks @RealKai42! - Fix thinking levels being offered for models that do not support them (e.g. phantom levels on Kimi K3): levels now come from each model's declared capabilities. Models that cannot disable reasoning (e.g. gpt-5) no longer offer an Off option, and turning thinking Off on models that support it (e.g. xai grok) now truly disables reasoning.

@moonshot-ai/pi-tui@0.80.8

Patch Changes

• #1976 e458323 Thanks @liruifengv! - Speed up the terminal renderer in long sessions by reusing processed lines across frames.