write tool: filePath truncated when content is large — field order puts content first #29140

Summary

When using local OpenAI-compatible backends (Ollama, LM Studio, llama.cpp), the `write` tool frequently fails with:

```
SchemaError(Missing key at ["filePath"])
```

even though the model intended to provide a filePath.

Root cause

The `write` tool schema defines `content` before `filePath`:

```ts
o.Struct({
content: o.String.annotate(...), // first
filePath: o.String.annotate(...), // second
})
```

Local models generate tool-call arguments in schema field order. When `content` is large (e.g. writing a full source file), the model exhausts its output token budget before reaching `filePath`, producing a valid JSON object with content but no filePath.

Fix

Swap the field order so `filePath` is declared before `content`. The model then generates the path first, and only the content can be truncated — which is a recoverable error vs. a missing path.

Related

• Default agent is brittle against local OpenAI-compatible tool-call quirks #20669 (bash tool `description` missing — same class of problem with local models omitting required fields)
• PR fix(opencode): make bash tool description parameter optional #26419 (fixes bash `description` but does not address write `filePath`)

Workaround

Swap the field order in the opencode binary. Works correctly after the patch with Qwen3.6, Qwen2.5-Coder, and Qwen3-Coder via Ollama.