Malformed input for a known tool is reported like an unknown tool call

[johanhallberg]

Description

Problem

When a model selects a valid tool but emits malformed or truncated input for that tool, the failure can be routed through the invalid tool in a way that looks similar to an unknown tool call.

This makes it harder to distinguish between two different cases:

1. the model called a tool that does not exist
2. the model called a valid tool, but the tool input was malformed or could not be parsed

Expected behavior

These cases should be distinguishable in the invalid-tool result, so users and maintainers can tell whether the problem was an unavailable tool or malformed input for an available tool.

Why this matters

The two cases require different follow-up behavior:

• unknown tool: the model/tool registry mismatch should be surfaced clearly
• known tool with malformed input: the model should retry with valid input or split a large/truncated operation into smaller chunks

Related PR

I have a small fix ready that classifies these cases separately.

Plugins

none

OpenCode version

dev branch / source build from anomalyco/opencode dev at 5720883, with fix branch based on that commit

Steps to reproduce

1. Have a model call an existing OpenCode tool, but with malformed or truncated JSON input.
2. Observe that the failure is routed through the invalid tool.
3. Compare this with a model calling a tool that does not exist.
4. Both cases are hard to distinguish from the invalid-tool result, even though they mean different things.

Screenshot and/or share link

PR: #25788

Operating System

macOS 26.4 locally; also reproduced/validated on Debian 12 test host

Terminal

iTerm2 locally; OpenCode web UI/browser for manual validation

[github-actions]

This issue might be a duplicate of existing issues. Please check:

• Truncated tool calls are misclassified and unrecoverable (finishReason: length + repairToolCall + doom loop) #18108: Directly overlaps — describes the same root cause where repairToolCall routes all parse failures to the generic invalid tool without distinguishing whether the tool name is valid. The P0 fix proposed there is essentially the same classification improvement described here.
• retry invalid tool-call diff / malformed tool input instead of aborting #15906: Related — covers malformed/truncated tool input handling and retry behavior for the same class of failures.
• Bug Report: invalid internal tool is exposed and can spam repeated calls #21900: Related — covers the invalid tool being surfaced in error messages and repeated call spam from malformed tool inputs.

If your fix in #25788 addresses a distinct aspect not covered by those issues, it may be worth cross-referencing them there.