Skip to content

fix: handle reasoning_content for Kimi/thinking models#7252

Merged
michaelneale merged 7 commits intoblock:mainfrom
clayarnoldg2m:fix/reasoning-content-kimi-thinking-models
Feb 16, 2026
Merged

fix: handle reasoning_content for Kimi/thinking models#7252
michaelneale merged 7 commits intoblock:mainfrom
clayarnoldg2m:fix/reasoning-content-kimi-thinking-models

Conversation

@clayarnoldg2m
Copy link
Contributor

@clayarnoldg2m clayarnoldg2m commented Feb 16, 2026

Summary

  • Preserves reasoning_content when splitting parallel tool calls in agent.rs — providers like Kimi require it on all assistant messages with tool_calls when thinking mode is enabled
  • Omits reasoning_content field entirely when empty in format_messages instead of sending "" — Kimi rejects empty reasoning_content
  • Properly accumulates reasoning_content chunks during streaming and emits as MessageContent::reasoning()

Context

Fixes #6902. Kimi (Moonshot) and other thinking models that use reasoning_content (similar to DeepSeek) were failing because:

  1. When parallel tool calls are split into separate messages, the reasoning content was lost
  2. An empty reasoning_content: "" was sent even when no reasoning occurred, which Kimi rejects
  3. Streaming reasoning_content chunks were not accumulated, so reasoning was lost in streamed responses

Reference: PR #6962 identified these issues. This PR implements the three core fixes cleanly.

Test plan

  • cargo test -p goose --lib — 666 tests pass
  • cargo clippy -p goose --all-targets -- -D warnings — clean
  • Manual verification with DeepSeek reasoning model (reasoning still works)
  • Manual verification with Kimi thinking model (tool calls + reasoning work)

🤖 Generated with Claude Code

clayarnoldg2m and others added 7 commits February 16, 2026 22:39
@clayarnoldg2m @claude
fix: relax test_basic_response assertion for providers returning reas…
14234e0 
…oning_content

xAI (and other providers like DeepSeek) can return reasoning_content alongside
text content, producing 2+ content items. The test now accepts >= 1 items and
verifies at least one is Text, instead of requiring exactly 1 item.

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Signed-off-by: clayarnoldg2m <carnold@g2m.ai>
@clayarnoldg2m @claude
fix(ui): preserve server config values on partial provider config save
f32c6dc 
The submit filter in ProviderConfiguationModal only included fields where
the user had typed a new value (entry.value), causing untouched fields like
API Host to be silently dropped. Now the filter also includes fields with
existing non-masked server values (entry.serverValue), preventing config
reversion when only some fields are edited.

Fixes block#7245

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Signed-off-by: clayarnoldg2m <carnold@g2m.ai>
@clayarnoldg2m @claude
fix: use explicit max_tokens in apps extension to prevent truncation
4748f38 
Replace provider.complete() with provider.complete_with_model() using
explicit max_tokens of 16384 in both generate_new_app_content() and
generate_updated_app_content(). This prevents app HTML from being
truncated when the provider's default max_tokens is too low.

Fixes block#7239

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Signed-off-by: clayarnoldg2m <carnold@g2m.ai>
@clayarnoldg2m @claude
style: fix cargo fmt in apps.rs complete_with_model calls
0a29b3a 
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Signed-off-by: clayarnoldg2m <carnold@g2m.ai>
@clayarnoldg2m @claude
fix: parse thinking blocks in Gemini CLI stream responses
07a2d8e 
Add a Some("thinking") arm to parse_stream_json_response() that
extracts thinking content from Gemini CLI stream events and creates
MessageContent::Thinking entries. Without this, thinking blocks were
silently dropped, causing truncated responses.

Includes tests for thinking block parsing and no-thinking fallback.

Fixes block#7203

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Signed-off-by: clayarnoldg2m <carnold@g2m.ai>
@clayarnoldg2m @claude
style: fix cargo fmt in gemini_cli.rs test assertions
a0538e1 
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Signed-off-by: clayarnoldg2m <carnold@g2m.ai>
@clayarnoldg2m @claude
fix: handle reasoning_content for Kimi/thinking models (block#6902)
efb7744 
Three fixes for reasoning_content handling:

1. agent.rs: Preserve reasoning_content when splitting parallel tool calls.
   Providers like Kimi require reasoning_content on all assistant messages
   with tool_calls when thinking mode is enabled.

2. openai.rs format_messages: Omit reasoning_content field entirely when
   empty instead of sending empty string. Kimi rejects empty
   reasoning_content ("").

3. openai.rs streaming: Properly accumulate reasoning_content chunks
   across streaming deltas and emit as MessageContent::reasoning().

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Signed-off-by: clayarnoldg2m <carnold@g2m.ai>
@michaelneale michaelneale mentioned this pull request Feb 16, 2026
Closed
@michaelneale
Copy link
Collaborator

michaelneale commented Feb 16, 2026

Full CI passed on branch PR: #7262

michaelneale
michaelneale approved these changes Feb 16, 2026
View reviewed changes
Copy link
Collaborator

@michaelneale michaelneale left a comment

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

nice - tested it in a branch

@michaelneale michaelneale added this pull request to the merge queue Feb 16, 2026
Merged via the queue into block:main with commit 05d2af4 Feb 16, 2026
36 checks passed
@github-actions github-actions bot mentioned this pull request Feb 17, 2026
Open
michaelneale added a commit that referenced this pull request Feb 17, 2026
@michaelneale
Merge branch 'main' into gateway-abstraction
990e73c 
* main:
  fix: handle reasoning_content for Kimi/thinking models (#7252)
  feat: sandboxing for macos (#7197)
  fix(otel): use monotonic_counter prefix and support temporality env var (#7234)
  Streaming markdown (#7233)
  Improve compaction messages to enable better post-compaction agent behavior (#7259)
  fix: avoid shell-escaping special characters except quotes (#7242)
tlongwell-block added a commit that referenced this pull request Feb 17, 2026
@tlongwell-block
Merge remote-tracking branch 'origin/main' into tui_gem
1e293fc 
* origin/main:
  docs: playwright CLI skill tutorial (#7261)
  install node in goose dir (#7220)
  fix: relax test_basic_response assertion for providers returning reasoning_content (#7249)
  fix: handle reasoning_content for Kimi/thinking models (#7252)
  feat: sandboxing for macos (#7197)
  fix(otel): use monotonic_counter prefix and support temporality env var (#7234)
  Streaming markdown (#7233)
  Improve compaction messages to enable better post-compaction agent behavior (#7259)
  fix: avoid shell-escaping special characters except quotes (#7242)
  fix: use dynamic port for Tetrate auth callback server (#7228)
  docs: removing LLM Usage admonitions (#7227)
  feat(otel): respect standard OTel env vars for exporter selection (#7144)
  fix: fork session (#7219)
  Bump version numbers for 1.24.0 release (#7214)
  Move platform extensions into their own folder (#7210)
  fix: ignore deprecated skills extension (#7139)

# Conflicts:
#	Cargo.lock
#	Cargo.toml
zanesq added a commit that referenced this pull request Feb 17, 2026
@zanesq
Merge branch 'main' of github.com:block/goose into zane/validate-bund…
9ccc8eb 
…led-extensions-cmd

* 'main' of github.com:block/goose: (24 commits)
  Set up direnv and update flake inputs (#6526)
  fix: restore subagent tool call notifications after summon refactor (#7243)
  fix(ui): preserve server config values on partial provider config save (#7248)
  fix(claude-code): allow goose to run inside a Claude Code session (#7232)
  fix(openai): route gpt-5 codex via responses and map base paths (#7254)
  feat: add GoosePlatform to AgentConfig and MCP initialization (#6931)
  Fix copied over (#7270)
  feat(gemini-cli): add streaming support via stream-json events (#7244)
  fix: filter models without tool support from recommended list (#7198)
  fix(google): handle more thoughtSignature vagaries during streaming (#7204)
  docs: playwright CLI skill tutorial (#7261)
  install node in goose dir (#7220)
  fix: relax test_basic_response assertion for providers returning reasoning_content (#7249)
  fix: handle reasoning_content for Kimi/thinking models (#7252)
  feat: sandboxing for macos (#7197)
  fix(otel): use monotonic_counter prefix and support temporality env var (#7234)
  Streaming markdown (#7233)
  Improve compaction messages to enable better post-compaction agent behavior (#7259)
  fix: avoid shell-escaping special characters except quotes (#7242)
  fix: use dynamic port for Tetrate auth callback server (#7228)
  ...
katzdave added a commit to YusukeShimizu/goose that referenced this pull request Feb 17, 2026
@katzdave
Merge remote-tracking branch 'origin/main' into responses-gpt-5.2-x
f9401ba 
* origin/main: (263 commits)
  working_dir usage more clear in add_extension (block#6958)
  Use Canonical Models to set context window sizes (block#6723)
  Set up direnv and update flake inputs (block#6526)
  fix: restore subagent tool call notifications after summon refactor (block#7243)
  fix(ui): preserve server config values on partial provider config save (block#7248)
  fix(claude-code): allow goose to run inside a Claude Code session (block#7232)
  fix(openai): route gpt-5 codex via responses and map base paths (block#7254)
  feat: add GoosePlatform to AgentConfig and MCP initialization (block#6931)
  Fix copied over (block#7270)
  feat(gemini-cli): add streaming support via stream-json events (block#7244)
  fix: filter models without tool support from recommended list (block#7198)
  fix(google): handle more thoughtSignature vagaries during streaming (block#7204)
  docs: playwright CLI skill tutorial (block#7261)
  install node in goose dir (block#7220)
  fix: relax test_basic_response assertion for providers returning reasoning_content (block#7249)
  fix: handle reasoning_content for Kimi/thinking models (block#7252)
  feat: sandboxing for macos (block#7197)
  fix(otel): use monotonic_counter prefix and support temporality env var (block#7234)
  Streaming markdown (block#7233)
  Improve compaction messages to enable better post-compaction agent behavior (block#7259)
  ...

# Conflicts:
#	crates/goose/src/providers/openai.rs
zanesq added a commit that referenced this pull request Feb 17, 2026
@zanesq
Merge branch 'main' of github.com:block/goose into zane/recipe-extens…
47c601c 
…ions-fallback

* 'main' of github.com:block/goose: (43 commits)
  Added cmd to validate bundled extensions json (#7217)
  working_dir usage more clear in add_extension (#6958)
  Use Canonical Models to set context window sizes (#6723)
  Set up direnv and update flake inputs (#6526)
  fix: restore subagent tool call notifications after summon refactor (#7243)
  fix(ui): preserve server config values on partial provider config save (#7248)
  fix(claude-code): allow goose to run inside a Claude Code session (#7232)
  fix(openai): route gpt-5 codex via responses and map base paths (#7254)
  feat: add GoosePlatform to AgentConfig and MCP initialization (#6931)
  Fix copied over (#7270)
  feat(gemini-cli): add streaming support via stream-json events (#7244)
  fix: filter models without tool support from recommended list (#7198)
  fix(google): handle more thoughtSignature vagaries during streaming (#7204)
  docs: playwright CLI skill tutorial (#7261)
  install node in goose dir (#7220)
  fix: relax test_basic_response assertion for providers returning reasoning_content (#7249)
  fix: handle reasoning_content for Kimi/thinking models (#7252)
  feat: sandboxing for macos (#7197)
  fix(otel): use monotonic_counter prefix and support temporality env var (#7234)
  Streaming markdown (#7233)
  ...

# Conflicts:
#	crates/goose/src/config/extensions.rs
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Reviewers

@michaelneale michaelneale michaelneale approved these changes

Assignees

No one assigned

Labels

None yet

Projects

None yet

Milestone

No milestone

Development

Successfully merging this pull request may close these issues.

[Bug] Kimi k2.5 API Error: "thinking is enabled but reasoning_content is missing" in tool calls

2 participants

@clayarnoldg2m @michaelneale

Comments