Claude Opus 4.5 (latest) eventually fails: thinking block cannot be modified

[Jonathanm10]

Description

When using Claude Opus 4.5 (latest) from Anthropic in OpenCode, conversations eventually fail with a validation error indicating that previously returned thinking / redacted_thinking blocks were modified.

Exact error:

messages.7.content.9: `thinking` or `redacted_thinking` blocks in the latest assistant message cannot be modified. These blocks must remain as they were in the original response.

This appears to happen consistently after some conversation length/tool usage (not immediately on first turn).

Impact

• Session becomes unrecoverable for that turn/model state
• Prevents continued use of Opus 4.5 in longer real-world sessions

Steps to reproduce

1. Configure OpenCode to use Anthropic Claude Opus 4.5 (latest)
2. Start a normal coding session with multiple turns (including tool calls)
3. Continue until context grows and OpenCode sends follow-up turns
4. Observe Anthropic validation failure above

Expected behavior

OpenCode should preserve any previously returned thinking blocks exactly as required by Anthropic and continue the session without validation errors.

Actual behavior

Request eventually fails with:

... `thinking` or `redacted_thinking` blocks in the latest assistant message cannot be modified ...

Notes

• Likely related to message reconstruction/compaction/replay logic for Anthropic thinking blocks.
• Possibly related prior reports: Claude Models Thinking block error #10970 and Expected thinking or redacted_thinking but found tool_use #8010 (similar thinking-block ordering/validation failures), but this report is the explicit "cannot be modified" variant on latest Opus 4.5.

Environment

• Provider: Anthropic
• Model: Claude Opus 4.5 (latest)
• OS / terminal / OpenCode version: not yet collected from reporter

[github-actions]

Thanks for the detailed report! This appears to be related to how thinking blocks are handled during message reconstruction/compaction with Anthropic's Opus 4.5 model.

This is related to similar thinking block issues with Claude models:

• Claude Models Thinking block error #10970: Claude Models Thinking block error
• Expected thinking or redacted_thinking but found tool_use #8010: Expected thinking or redacted_thinking but found tool_use

The specific "cannot be modified" validation error suggests this is happening during message replay/reconstruction when OpenCode resends previous assistant messages. This is distinct from the other thinking block ordering issues, but the root cause may be related to how thinking blocks are being preserved and replayed in conversation history.

[Jonathanm10]

Adding reporter environment details:

• OpenCode version: 1.1.60
• OS: macOS 26.2 (25C56)
• Provider/auth path: Anthropic OAuth subscription (direct Anthropic account auth)

This is reproducible in real usage over longer multi-turn sessions with Claude Opus 4.5 (latest).

[shrimpwtf]

thinking or redacted_thinking blocks in the latest assistant message cannot be modified.

getting very frequently

[gnadaban]

Anyone running into this, there are now 2 pending PRs fixing this issue, it's only a matter of time hopefully. :)

[tongsh6]

I'm also experiencing this same issue.

Error message:

messages.5.content.51: `thinking` or `redacted_thinking` blocks in the latest assistant message cannot be modified. These blocks must remain as they were in the original response.

This occurs during normal usage with extended thinking enabled. The conversation becomes unrecoverable once this error appears.

Environment:

• OS: Windows 11
• Model: Claude Opus 4.6 (via Anthropic)

+1 for a fix — this seems to be a persistent issue across multiple Claude model versions and is a significant blocker for longer sessions.

[pietdaniel]

Same Experience, this happens when I get to ~120k tokens and try to run a compaction.

undefined: The model returned the following errors: messages.1.content.1: `thinking` or `redacted_thinking` blocks in the latest assistant message cannot be modified. These blocks must remain as they were in the original response.

---

Current model information

Session ID: ses_35b0c2b0bffeAhmbLv3c6TYj3J
Created: 2/28/2026, 10:52:49 AM
Updated: 2/28/2026, 10:53:20 AM

---

User

list your current model

---

Assistant (Build · anthropic.claude-opus-4-6-v1 · 5.2s)

Thinking:

The user is asking about my current model. I already have this information from the system prompt - I'm powered by anthropic.claude-opus-4-6-v1 with the exact model ID amazon-bedrock/anthropic.claude-opus-4-6-v1.

My current model is:

• Model name: anthropic.claude-opus-4-6-v1
• Exact model ID: amazon-bedrock/anthropic.claude-opus-4-6-v1

---

User

list the current machine specification you are running on

---

Assistant (Build · anthropic.claude-opus-4-6-v1 · 7.1s)

Based on the environment information available to me, here's what I know:

• Platform: macOS (darwin)

---

$ opencode --version 2>/dev/null || opencode version 2>/dev/null || which opencode 2>/dev/null
1.2.10

[paul4156]

Hitting this consistently with Claude Sonnet 4.6 + high thinking budget. Once thinking blocks accumulate in history, compaction fails and the session is dead. The only escape is starting a new session.

A configurable compaction threshold (e.g. trigger at 50%) would help a lot — even with some context loss, getting a working session at 50% is far better than a broken one at 85%. Related: #11314.