[AGE-3962] fix(runner): surface Claude extended-thinking in the playground by ardaerzin · Pull Request #5372 · Agenta-AI/agenta

ardaerzin · 2026-07-17T22:19:18Z

Problem

In the agent playground, the Claude Code harness shows no thinking/reasoning for some models — Sonnet shows nothing, Haiku works — even though both 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".

Testing

New unit test for the claudeThinkingMeta helper; full runner suite green (pnpm test, pnpm run typecheck).
Verified live in the playground: Sonnet (now surfaces thinking), Opus (now surfaces thinking), Haiku (unchanged — still works).

Notes

A separate, complementary issue exists where reasoning is dropped when a turn resolves in batch mode (fold() in the SDK); that's tracked separately and not included here.

linear-code · 2026-07-17T22:19:21Z

AGE-3962

vercel · 2026-07-17T22:19:24Z

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

Project	Deployment	Actions	Updated (UTC)
agenta-documentation	Ready	Preview, Comment	Jul 17, 2026 10:57pm

coderabbitai · 2026-07-17T22:19:30Z

No actionable comments were generated in the recent review. 🎉

ℹ️ Recent review info

⚙️ Run configuration

Configuration used: Organization UI

Review profile: CHILL

Plan: Pro Plus

Run ID: 3e3b0620-dddc-492f-b8ea-c96242a971b5

📥 Commits

Reviewing files that changed from the base of the PR and between 68a7a89 and 12a5f36.

📒 Files selected for processing (3)

services/runner/src/engines/sandbox_agent.ts
services/runner/src/engines/sandbox_agent/claude-thinking.ts
services/runner/tests/unit/claude-thinking.test.ts

🚧 Files skipped from review as they are similar to previous changes (3)

services/runner/src/engines/sandbox_agent/claude-thinking.ts
services/runner/src/engines/sandbox_agent.ts
services/runner/tests/unit/claude-thinking.test.ts

📝 Walkthrough

Summary by CodeRabbit

New Features
- Claude runs now display summarized extended-thinking content in the playground.
- Thinking settings are applied consistently when starting or resuming sessions.
Bug Fixes
- Prevented newer Claude models from hiding thinking content by default.

Walkthrough

Claude sandbox sessions now send adaptive, summarized extended-thinking metadata alongside existing system-prompt metadata. A helper and unit test define and verify the configuration, while acquireEnvironment merges it into Claude session initialization.

Changes

Claude thinking display

Layer / File(s)	Summary
Define and validate thinking metadata `services/runner/src/engines/sandbox_agent/claude-thinking.ts`, `services/runner/tests/unit/claude-thinking.test.ts`	Adds the exported `ClaudeThinkingMeta` shape and `claudeThinkingMeta()` helper, with a unit test asserting adaptive summarized thinking configuration.
Merge metadata into session initialization `services/runner/src/engines/sandbox_agent.ts`	Merges Claude thinking metadata with existing system-prompt metadata in the Claude `sessionInit` payload.

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

Possibly related PRs

Agenta-AI/agenta#5268: Modifies the Claude session initialization metadata path in the sandbox runner.

🚥 Pre-merge checks | ✅ 4 | ❌ 1

❌ Failed checks (1 warning)

Check name	Status	Explanation	Resolution
Docstring Coverage	⚠️ Warning	Docstring coverage is 50.00% which is insufficient. The required threshold is 60.00%.	Write docstrings for the functions missing them to satisfy the coverage threshold.

✅ Passed checks (4 passed)

Check name	Status	Explanation
Title check	✅ Passed	The title clearly summarizes the main change: surfacing Claude extended-thinking in the playground.
Description check	✅ Passed	The description matches the code changes and explains the Claude thinking display fix in the playground.
Linked Issues check	✅ Passed	The changes address `#5355` by enabling Claude thinking output to appear while preserving final responses.
Out of Scope Changes check	✅ Passed	The diff appears focused on the Claude thinking meta helper, runner wiring, and a matching unit test.

✨ 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 fix/claude-extended-thinking-display

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

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

coderabbitai

Actionable comments posted: 1

ℹ️ Review info

⚙️ Run configuration

Configuration used: Organization UI

Review profile: CHILL

Plan: Pro Plus

Run ID: b49a5598-61e8-42ef-905a-0fa6c69dc52f

📥 Commits

Reviewing files that changed from the base of the PR and between 6c0265a and 68a7a89.

📒 Files selected for processing (3)

services/runner/src/engines/sandbox_agent.ts
services/runner/src/engines/sandbox_agent/claude-thinking.ts
services/runner/tests/unit/claude-thinking.test.ts

…round Recent Claude models (e.g. Sonnet) default extended-thinking `display` to "omitted" — the API returns signature-only thinking blocks whose text is empty. claude-agent-acp emits an `agent_thought_chunk` only when that text is non-empty, so the runner never sees the reasoning and the playground shows none, while models that return thinking text (e.g. Haiku) work. For the Claude harness, request `display: "summarized"` via `_meta.claudeCode.options.thinking` so the reasoning text is returned for every model. `type: "adaptive"` leaves the think-or-not decision to the model, and "summarized" still returns the signature, so multi-turn thinking continuity is unaffected.

mmabrouk · 2026-07-18T19:15:28Z

Ported onto the decomposed runner layout in #5380 (the runner engine was split since this PR was opened; the session-init insertion now lives in engines/sandbox_agent/environment.ts). #5380 supersedes this PR — no behavior change. Leaving this open for the author to close.

mmabrouk · 2026-07-19T14:01:12Z

deprecated

vercel Bot deployed to Preview July 17, 2026 22:20 View deployment

ardaerzin marked this pull request as ready for review July 17, 2026 22:36

dosubot Bot added size:M This PR changes 30-99 lines, ignoring generated files. bug Something isn't working labels Jul 17, 2026

coderabbitai Bot reviewed Jul 17, 2026

View reviewed changes

Comment thread services/runner/src/engines/sandbox_agent.ts

ardaerzin force-pushed the fix/claude-extended-thinking-display branch from 68a7a89 to 12a5f36 Compare July 17, 2026 22:56

vercel Bot deployed to Preview July 17, 2026 22:57 View deployment

mmabrouk mentioned this pull request Jul 18, 2026

[AGE-3962] fix(runner): surface Claude extended-thinking in the playground #5380

Merged

mmabrouk closed this Jul 19, 2026

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Uh oh!

[AGE-3962] fix(runner): surface Claude extended-thinking in the playground#5372

[AGE-3962] fix(runner): surface Claude extended-thinking in the playground#5372
ardaerzin wants to merge 1 commit into
mainfrom
fix/claude-extended-thinking-display

ardaerzin commented Jul 17, 2026 •

edited

Loading

Uh oh!

linear-code Bot commented Jul 17, 2026

Uh oh!

vercel Bot commented Jul 17, 2026 •

edited

Loading

Uh oh!

coderabbitai Bot commented Jul 17, 2026 •

edited

Loading

Summary by CodeRabbit

Walkthrough

Changes

❌ Failed checks (1 warning)

Uh oh!

coderabbitai Bot left a comment

Uh oh!

Uh oh!

mmabrouk commented Jul 18, 2026

Uh oh!

mmabrouk commented Jul 19, 2026

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants

Uh oh!

Conversation

ardaerzin commented Jul 17, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Problem

Root cause

Fix

Before / after

Testing

Notes

Uh oh!

linear-code Bot commented Jul 17, 2026

Uh oh!

vercel Bot commented Jul 17, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

coderabbitai Bot commented Jul 17, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Summary by CodeRabbit

Walkthrough

Changes

❌ Failed checks (1 warning)

Uh oh!

coderabbitai Bot left a comment

Choose a reason for hiding this comment

Uh oh!

Uh oh!

mmabrouk commented Jul 18, 2026

Uh oh!

mmabrouk commented Jul 19, 2026

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants

ardaerzin commented Jul 17, 2026 •

edited

Loading

vercel Bot commented Jul 17, 2026 •

edited

Loading

coderabbitai Bot commented Jul 17, 2026 •

edited

Loading