Anthropic subagents fail with reasoning.effort due to thinking signature replay in multi-turn sessions

[vicary]

Description

Summary

Anthropic-backed subagents fail when reasoning.effort is enabled. The failure appears on a later model call in a multi-turn / tool-using / resumed subagent session with errors such as:

• Invalid 'signature' in 'thinking' block
• thinking.signature: Field required

Removing reasoning.effort from the subagent config makes the same flow stop failing.

Environment / shape of failure

• Affected path: Anthropic-backed subagents
• Trigger: reasoning.effort enabled in agent frontmatter
• Typical failing shape: second or later API call in the same subagent session
• Common ingredients: tool use, multi-turn execution, or session replay/resume

What I ruled out

This does not appear to be primarily caused by switching from GPT/OpenAI to Anthropic in the same session. I have:

• Anthropic-only failures with the same error
• GPT -> Anthropic sessions that do not fail this way

So the stronger correlation is specifically with Anthropic thinking mode being enabled via reasoning.effort.

Suspected root cause

This looks like a subagent history replay / truncation / normalization bug for Anthropic thinking blocks.

My current hypothesis is:

1. reasoning.effort enables Anthropic extended thinking
2. the first Claude turn succeeds and emits a thinking block with signature metadata
3. on a later subagent call, prior assistant history is replayed without preserving the thinking/signature payload exactly
4. Anthropic rejects the reconstructed conversation with a 400 signature error

This seems closely related to prior Anthropic thinking-block replay issues, but likely on the subagent path specifically.

Minimal workaround

Remove reasoning.effort from the Anthropic-backed subagent frontmatter. That reliably avoids the failure locally.

Expected behavior

Anthropic-backed subagents with reasoning.effort should be able to survive multi-turn execution, tool use, and resumed/replayed session history without signature-related API failures.

Related issues / prior art

These look related, but not identical:

• [bug] v0.8.0 Introduced thinking block bug with Claude Sonnet 4: "AI_APICallError: messages.3.content.0.type: Expected thinking or redacted_thinking, but found text." #2599
• Expected thinking or redacted_thinking, but found tool_use #3077

There is also a public community report that appears to describe the same subagent-specific failure mode:

• https://www.answeroverflow.com/m/1473362418766839918

Notes

I can provide a tighter repro if needed, but the highest-signal trigger I found is simply:

• Anthropic subagent
• reasoning.effort enabled
• multi-turn / tool-using session
• later call fails with thinking signature validation

Plugins

None relevant / default setup

OpenCode version

Current local install as of 2026-03-06 (exact version not captured in-session)

Steps to reproduce

1. Configure a subagent backed by an Anthropic Claude model.
2. Add reasoning.effort to the agent frontmatter.
3. Run a task that causes the subagent to take multiple turns, especially with tool use or resumed/replayed session history.
4. Observe that a later model call fails with a thinking-signature validation error.
5. Remove reasoning.effort and rerun the same flow.
6. Observe that the failure disappears.

Screenshot and/or share link

Related prior art:

• [bug] v0.8.0 Introduced thinking block bug with Claude Sonnet 4: "AI_APICallError: messages.3.content.0.type: Expected thinking or redacted_thinking, but found text." #2599
• Expected thinking or redacted_thinking, but found tool_use #3077
• https://www.answeroverflow.com/m/1473362418766839918

Operating System

macOS

Terminal

OpenCode CLI session

[vicary]

One extra data point that may help narrow this down:

• I specifically tested whether this was caused by mixing GPT/OpenAI messages before switching to Anthropic in the same session.
• Local evidence did not support that theory.
• I found Anthropic-only failures with the same signature error, and also GPT -> Anthropic sessions that did not fail.

So the strongest trigger I could isolate was:

• Anthropic-backed subagent
• reasoning.effort enabled
• multi-turn and/or tool-using subagent session
• later call fails with thinking-signature validation

Workaround confirmed locally: removing reasoning.effort from the Anthropic subagent config makes the failure disappear.