fix(server): Claude thinking before a tool call stays a collapsed Thought

[maria-rcks]

Problem

On Claude Opus 5.5, Fable 5.1, and other Claude 5.1+ models, reasoning often renders as plain chat text instead of a collapsed Thought. #18364 promotes the last non-empty thinking block before a tool_use to assistant text, assuming it is a progress update. T3 requests display: "summarized", and Anthropic's progress updates table says that under summarized a progress update is "not distinguishable from a reasoning block". The model can also skip the update, so the block right before a tool call is often real reasoning.

Change

Revert the promotion. Every Claude thinking block stays a reasoning item and renders collapsed with the tool calls. The existing test now asserts that thinking right before a tool call stays reasoning for Opus 5.5, Fable 5.1, Opus 5, and Sonnet 4.6. This brings back #18355: a Claude 5.1+ progress update before a tool call shows as a Thought, not as chat text. Splitting the two needs display: "updates", which hides reasoning, so it is out of scope here.

Verification

Same read-only prompt on Claude Opus 5.5 in a dev server, before and after. Before, the reasoning ("Since my working directory is apps/server…") was stored as assistant_message and rendered as chat text. After, both thoughts are reasoning items shown as collapsed Thought rows.

Recordings of the same Opus 5.5 turn streaming, before and after:

https://uploads-production-47e4.up.railway.app/files/765d93ec-5cad-4ee7-a715-5445f629819b/claude-opus-reasoning-before.mp4

https://uploads-production-47e4.up.railway.app/files/385935e4-77e4-4d9a-9350-1e7f4dff4121/claude-opus-reasoning-after.mp4

Other providers with the same prompt: Grok and Muse stored reasoning as reasoning. Codex used 0 reasoning tokens on this prompt, and recent real Codex turns store reasoning summaries correctly. OpenCode with Claude Fable 5.1 got an empty reasoning part from OpenCode itself, since it does not request summarized thinking. Pi hit an Anthropic 401 and Cursor was unavailable in the dev environment, so neither is verified. ClaudeAdapterV2.test.ts: 252 passed.

Implemented with Claude Opus 5.5 through the Claude Code harness (T3 Code).

🤖 Generated with Claude Code

[macroscopeapp]

Approvability

Verdict: Approved at d1e02cc

Macroscope's review found this PR approvable — This is a small, self-contained Claude adapter fix that keeps ambiguous thinking blocks collapsed as reasoning instead of emitting them as chat text. The behavior is covered by focused tests across the affected model variants, with no schema, deployment, security, or configuration changes.

Notes:

• Diff unchanged. Approvability was decided on eligibility alone.

^{You can add or adjust custom eligibility rules. Learn more.}

Dismissing prior approval to re-evaluate d1e02cc

[coderabbitai]

No actionable comments were generated in the recent review. 🎉

ℹ️ Recent review info

⚙️ Run configuration

• Configuration used: Path: .coderabbit.config.ts
• Review profile: CHILL
• Plan: Advanced
• Run ID: 78785d67-f263-4feb-827f-6120e3cf5b28

📥 Commits

Reviewing files that changed from the base of the PR and between e24ea1d and d1e02cc.

📒 Files selected for processing (2)

• apps/server/src/orchestration-v2/Adapters/ClaudeAdapterV2.test.ts
• apps/server/src/orchestration-v2/Adapters/ClaudeAdapterV2.ts

Included review availability: This review used your included allowance. Your plan provides up to 10 included reviews per hour; 9 remain after this review.

📝 Walkthrough

ClaudeAdapterV2 no longer uses model-version detection or saved thinking snapshots to emit pre-tool assistant text. Tests now expect summarized thinking to remain a reasoning item across the listed models.

Changes: Claude thinking item classification

Layer / File(s)	Summary
Keep thinking snapshots as reasoning apps/server/src/orchestration-v2/Adapters/ClaudeAdapterV2.ts, apps/server/src/orchestration-v2/Adapters/ClaudeAdapterV2.test.ts	The adapter removes model-version detection and lastThinking state. The snapshot loop processes thinking blocks, and tests expect reasoning items for Opus 5.5, Fable 5.1, Opus 5, and Sonnet 4.6.

Priority: ⬇️ Low

Estimated code review effort: 2 (Simple) | ~8 minutes

Change: Bug fix

Suggested reviewers: juliusmarminge

Merge Risk: ⚪ Minimal · up to d1e02

Thinking before tool calls remains a Thought rather than assistant text. No actionable merge risk remains.

### Architecture Summary

Architecture risk: 🔵 Low · up to d1e02

The change affects 1 system.

Changed systems: apps/server

Architecture concerns
No architecture-level concerns identified.

Review details

Systems and components

• observed — apps/server (service) was modified; 2 changed files map to changed impact.

Before / after behavior

• observed — Modified behavior in apps/server/src/orchestration-v2/Adapters/ClaudeAdapterV2.test.ts: The test now describes summarized thinking as reasoning and runs the same expectation for Opus 5.5, Fable 5.1, Opus 5, and Sonnet 4.6; it replaces the per-model expected-type cases.
• observed — Modified behavior in apps/server/src/orchestration-v2/Adapters/ClaudeAdapterV2.test.ts: The assertion now expects the pre-tool reply text to be a reasoning item for all tested models, replacing the model-dependent expected type.
• observed — Modified behavior in apps/server/src/orchestration-v2/Adapters/ClaudeAdapterV2.ts: Removed claudeWritesUpdatesAsThinking, which parsed Claude model versions and returned true for versions 5.1 and later.
• observed — Modified behavior in apps/server/src/orchestration-v2/Adapters/ClaudeAdapterV2.ts: Removed lastThinking from the reasoning state, which previously retained the latest non-empty thinking snapshot’s message ID, item ID, and text for possible reuse as assistant output.

Pre-merge checks · 4

Passed checks (4 passed)

Check name	Status	Explanation
Linked Issues check		Check skipped because no linked issues were found for this pull request.
Out of Scope Changes check		Check skipped because no linked issues were found for this pull request.
Title check		The title is concise, specific, and accurately describes the primary change: Claude thinking before tool calls remains a collapsed Thought.
Description check		The description clearly covers the problem, change, and verification, including test results and limitations. The Scope and approval section is not explicit, but the focused fix and related issue refe…

✨ Finishing Touches

🧪 Generate unit tests (beta)

• Create a new PR

• Autofix · Keep fixing CodeRabbit findings and required CI, and resolving merge conflicts

_{Comment @coderabbitai help to get the list of available commands.}