Tool execution aborted when tool call is truncated (empty input)

[cjmayoral]

snippet-01-tool-aborted.json
snippet-02-message-finish-length.json
snippet-03-session-created.log
snippet-04-llm-model.log
snippet-05-session-cancel-1.log
snippet-06-session-cancel-2.log

Tool execution aborted when tool call is truncated (empty input)

Summary

During session "Resumen T84 antes de continuar" (ses_3b7971163ffexUgX5mjSgXDfWQ), several tool calls ended with Tool execution aborted. One concrete instance shows an edit tool call with empty input ({}) that immediately fails. The same assistant message ended with finish: "length" (output length limit), which suggests the tool call may have been truncated before arguments were fully emitted. This looks like a truncation/cancel edge case rather than an LLM transport failure.

Environment

• OpenCode version: 1.1.49
• OS: Windows
• Model: claude-opus-4-6 (azure-anthropic)
• Session title: "Resumen T84 antes de continuar"
• Session ID: ses_3b7971163ffexUgX5mjSgXDfWQ

Evidence (sanitized excerpts)

The following log snippets are attached in snippets/ (redacted paths):

1. snippet-01-tool-aborted.json — tool call aborted with empty input

2. snippet-02-message-finish-length.json — same message finished due to output length

3. snippet-03-session-created.log — session creation (OpenCode 1.1.49)

4. snippet-04-llm-model.log — model/provider info

5. snippet-05-session-cancel-1.log — user cancellation (expected behavior)

6. snippet-06-session-cancel-2.log — user cancellation (expected behavior)

Notes / Hypothesis

• The tool call is emitted with empty input and immediately aborted.
• The corresponding assistant message ended with finish: "length", which is consistent with output truncation.
• This suggests OpenCode may attempt to execute a tool call even when its arguments are incomplete due to truncation.

Expected behavior

If a tool call is truncated (missing or empty input), OpenCode should either:

• retry the model to complete the tool arguments, or
• skip tool execution with a clearer error (e.g., "tool call incomplete due to output length"), instead of Tool execution aborted.

Questions for maintainers

• Is there a guard that detects incomplete tool call payloads before execution?
• Is there a known interaction between finish: "length" and tool call emission that can cause empty-input tool calls?

Related issues

• Infinite retry loop when StreamIdleTimeoutError occurs during tool input generation #12233 (closed) — Infinite retry loop when StreamIdleTimeoutError occurs during tool input generation
• Bug: Session corrupted when tool execution is aborted - missing tool_result causes API rejection #8312 (closed) — Session corrupted when tool execution is aborted - missing tool_result causes API rejection
• Tool execution aborted #1100 (closed) — Tool execution aborted (JSON/tool input parse failure)
• Invalild tool + execution aborted #11612 (open) — Invalild tool + execution aborted (invalid tool call)

[github-actions]

This issue might be related to existing issues addressing similar tool execution failures:

• Infinite retry loop when StreamIdleTimeoutError occurs during tool input generation #12234: Infinite retry loop when StreamIdleTimeoutError occurs during tool input generation — directly addresses stream truncation causing tool input failures
• Invalild tool + execution aborted #11612: Invalid tool + execution aborted — discusses tool execution abortion with validation errors

Your report provides additional detail about the truncation → empty input → abort mechanism. The maintainers may consolidate findings across these issues.

[TheOrdinaryWow]

same here

[onurdemircan-softtech]

I'm also running into this issue with a local model served via vLLM . I'd like to propose a practical approach for handling this case.

When the model is streaming a tool call (e.g. write), the argument chunks come in incrementally:

argument: "{"
argument: "\"filePath\": \"src/app.ts\","
argument: "\"content\": \"import express..."
argument: "...more code chunks..."
argument: "...more code chunks..."
← truncated here, finish_reason: "length"

At this point, the tool call arguments are incomplete JSON, so OpenCode aborts the entire tool call and the partial output is lost.

Instead of aborting, OpenCode maybe should:

Detect that finish_reason is "length" while a tool call argument is still being streamed (i.e. the JSON is unclosed).
Synthetically close the incomplete JSON argument — append the necessary closing characters (", }, etc.) to make the partial argument valid.
Treat it as a completed tool call internally, but do NOT execute it (don't actually write the file).
Show a warning to the user with the partial content that was generated so far, something like: "Tool call truncated due to output token limit. Partial content below:"
Stop the agentic loop and hand control back to the user.

This way, the user can see exactly what was generated before truncation and decide how to proceed — whether to ask the model to continue from where it left off or retry with a different approach
The key insight is: don't discard the partial streamed data. The model already generated useful content; OpenCode just needs to gracefully capture it instead of treating truncation as a fatal error.

[lucamaia9]

same issue here:

~ Updating todos...
Tool execution aborted
""