Sonnet rejects some /undo + /compact sessions that GPT still accepts

[markthepixel]

I think the Claude auth plugin may be part of the failure path here, but I can’t prove it’s the original cause.

What I’m seeing: after using /undo and then /compact in OpenCode, some sessions start failing on Sonnet with:

messages.2: tool_use ids were found without tool_result blocks immediately after ... Each tool_use block must have a corresponding tool_result block in the next message.

If I switch the same session from Sonnet to GPT, it works.

That makes me think one of two things is happening:

• OpenCode is producing slightly malformed history after undo/compact
• this plugin is not repairing / normalizing that history enough for Anthropic’s stricter validation

The reason I’m opening this here is that opencode-claude-auth does touch the message payload and already has repairToolPairs() in dist/transforms.js.

From what I can tell, that function checks whether matching tool_use / tool_result ids both exist somewhere in the message list, but it does not enforce the Anthropic rule that the matching tool_result has to be in the very next message.

So if undo/compact leaves history in a state where:

• message N has a tool_use
• the matching tool_result still exists, but not immediately after

then GPT may still tolerate it, but Sonnet rejects it, and the plugin lets it through because the pair technically still exists.

I inspected one broken session locally and found exactly that shape: the assistant message had the tool_use, and the next message was another assistant message, not the matching tool result wrapper.

So to be clear: I’m not saying this plugin necessarily creates the bad history. I suspect OpenCode core may be the original source, especially with /undo + /compact. But I do think this plugin may be related because it’s the Claude-specific path, and its repair logic seems too shallow for Anthropic’s adjacency requirement.

If helpful, I can put together a cleaner repro next time it happens. Right now the strongest evidence is:

• broken session on Sonnet
• same session works on GPT
• happened after /undo + /compact
• plugin has tool-pair repair logic, but it appears to check existence rather than adjacency

Thanks — posting here because this feels close enough to the Claude transform layer that it seemed worth flagging.

[griffinmartin]

Closing as a duplicate of #212 — same error, same root cause. The adjacency analysis here is correct and valuable: repairToolPairs (src/transforms.ts) validates that tool_use/tool_result id pairs exist somewhere in the message list, but does not enforce Anthropic's rule that the tool_result must be in the immediately-following message, which is exactly the shape /undo + /compact can produce. Cross-linked into #212 so the fix work happens in one place.