apply_patch failure rate differs sharply by model (aggregate local data)

[ricatix]

Description

apply_patch error parts differ materially between two model IDs in same local OpenCode install. This may be model tool-call output quality, but current stored error data does not distinguish invalid patch input from file/context or runtime failures.

Local aggregate, 2026-07-11 through 2026-08-10:

Model ID	Attempts	Errors	Rate
cx/gpt-5.6-terra	1,223	156	12.76%
cx/gpt-5.6-luna	1,092	397	36.36%

An error is stored tool part where type=tool, tool=apply_patch, and state.status=error. Retries count separately. Attribution uses session.model.id, so it is aggregate evidence, not per-invocation proof. No raw paths, code, prompts, diffs, or tool payloads included.

Related: #18108, #20227, #27282.

Plugins

@opencode-ai/plugin 1.15.11

OpenCode version

1.18.15

Steps to reproduce

1. Use OpenCode with a session whose model.id is either model above.
2. Perform file edits that invoke apply_patch.
3. Inspect stored part records for state.status = "error".
4. Compare error-part count with apply_patch attempt count per model.

I cannot provide a deterministic minimal reproduction yet. What normalized error field or telemetry should be used to separate malformed tool input from file/context and runtime failures?

Screenshot and/or share link

Not available. Sharing local sessions would expose repository data.

Operating System

macOS

Terminal

Not relevant; aggregate data read from local SQLite session store.

[github-actions]

Thanks for updating the issue! The bug report now includes a clear description, reproduction steps, version info, and supporting data. Everything looks good.