[AGE-3962] fix(runner): surface Claude extended-thinking in the playground

[mmabrouk]

Problem

In the agent playground, the Claude Code harness shows no thinking for some models. Sonnet and Opus show nothing; Haiku works — even though all three stream fine. Closes #5355 (AGE-3962).

Root cause

Recent Claude models default extended-thinking display to "omitted": the Anthropic API returns signature-only thinking blocks whose text is empty. @agentclientprotocol/claude-agent-acp emits an agent_thought_chunk only when that text is non-empty, so on those models the runner never sees any reasoning and nothing renders — in streaming or batch. Models that return thinking text (e.g. Haiku) are unaffected. The runner set no thinking config, so recent models fell to the omitted default.

This is emission-side (the reasoning never reaches us), not a rendering bug.

Fix

For the Claude harness, request visible ("summarized") thinking display via _meta.claudeCode.options.thinking — the same _meta channel the system-prompt append already uses, which claude-agent-acp spreads over its env-derived thinking config (so it wins):

{ claudeCode: { options: { thinking: { type: "adaptive", display: "summarized" } } } }

• type: "adaptive" leaves the think-or-not decision to the model (no forced budget).
• "summarized" still returns the thinking signature, so multi-turn thinking continuity is unaffected.
• Guarded by plan.acpAgent === "claude" — Pi harness is untouched (it has its own reasoning path).

The change is model-agnostic within the Claude harness: it fixes Sonnet/Opus (same omitted-default cause) and leaves Haiku/default behavior unchanged (summarized is their existing default, now explicit).

Before / after

• Before: Claude Code + Sonnet → no "Thought" in the playground.
• After: Claude Code + Sonnet → reasoning streams and renders as "Thought".

Supersedes #5372

This PR supersedes #5372 (same author's fix), ported onto the decomposed runner layout. The runner engine was split since #5372 was opened: the monolithic sandbox_agent.ts is now a thin facade over engines/sandbox_agent/*. #5372's new claude-thinking.ts module and its test are byte-identical here; its one session-init insertion (the import and the claudeThinking / claudeMeta merge) now lands in engines/sandbox_agent/environment.ts, next to where claudeSystemPromptMeta and the sessionInit payload are already assembled — the block that moved out of the monolith. No behavior change from #5372. #5372 can be closed in favor of this one.

Testing

• New unit test for the claudeThinkingMeta helper.
• pnpm run typecheck clean; pnpm test green — 1205 tests (was 1204, +1 for the new test).

https://claude.ai/code/session_01DnWRxU3dCJ11hgDidm26vW

[linear-code]

AGE-3962

[vercel]

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

Project	Deployment	Actions	Updated (UTC)
agenta-documentation	Ready	Preview, Comment	Jul 18, 2026 7:16pm

[mmabrouk]

@coderabbitai review

[coderabbitai]

✅ Action performed

Review finished.

Note: CodeRabbit is an incremental review system and does not re-review already reviewed commits. This command is applicable only when automatic reviews are paused.

[coderabbitai]

📝 Walkthrough

Walkthrough

Adds a Claude extended-thinking metadata helper, requests adaptive summarized reasoning for Claude sessions, merges it with existing system-prompt metadata, and adds unit coverage for the configuration.

Changes

Claude thinking configuration

Layer / File(s)	Summary
Thinking metadata contract and test services/runner/src/engines/sandbox_agent/claude-thinking.ts, services/runner/tests/unit/claude-thinking.test.ts	Exports ClaudeThinkingMeta and claudeThinkingMeta(), which configure adaptive thinking with summarized display; the unit test verifies the nested options.
Session initialization integration services/runner/src/engines/sandbox_agent/environment.ts	Generates Claude thinking metadata for Claude agents and shallow-merges it with existing system-prompt metadata before attaching _meta to sessionInit.

Estimated code review effort: 2 (Simple) | ~10 minutes

Possibly related PRs

• Agenta-AI/agenta#5372: Adds the same Claude thinking helper and session initialization metadata merge.

🚥 Pre-merge checks | ✅ 5

✅ Passed checks (5 passed)

Check name	Status	Explanation
Docstring Coverage	✅ Passed	No functions found in the changed files to evaluate docstring coverage. Skipping docstring coverage check.
Linked Issues check	✅ Passed	Check skipped because no linked issues were found for this pull request.
Out of Scope Changes check	✅ Passed	Check skipped because no linked issues were found for this pull request.
Title check	✅ Passed	The title clearly matches the main change: surfacing Claude extended-thinking in the playground.
Description check	✅ Passed	The description accurately explains the Claude thinking metadata fix and its playground impact.

✨ Finishing Touches

📝 Generate docstrings

• Create stacked PR
• Commit on current branch

🧪 Generate unit tests (beta)

• Create PR with unit tests
• Commit unit tests in branch port/claude-thinking-decomp

---

Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out.

❤️ Share

• X
• Mastodon
• Reddit
• LinkedIn

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

[github-actions]

Railway Preview Environment

Status	Destroyed (PR closed)

Updated at 2026-07-18T23:40:47.777Z

[mmabrouk]

lgtm