fix(opencode): recover from truncated tool calls instead of failing silently

[nickveenhof]

Issue for this PR

Closes #18108
Closes #17471
Refs #14519, #13102, #18151, #18131

Type of change

• Bug fix
• New feature
• Refactor / code improvement
• Documentation

What does this PR do?

When a model's output hits max_tokens mid-tool-call, the streamed JSON arguments get truncated. The tool parser receives undefined for required fields, producing:

The bash tool was called with invalid arguments: [
  {"expected": "string", "code": "invalid_type", "path": ["command"],
   "message": "Invalid input: expected string, received undefined"}
]

Two fixes:

1. Truncation detection in experimental_repairToolCall (llm.ts)

repairToolCall currently only handles tool name case-sensitivity. All other parse failures route to the generic invalid tool. But when toolName is a valid registered tool and the args failed to parse, the cause is truncation, not an invalid tool. This check distinguishes the two cases and returns an actionable error telling the model to split its operation into smaller pieces.

2. Auto-continue on finishReason: "length" (prompt.ts)

The session loop exits when finishReason is anything other than "tool-calls". This means "length" (token limit hit) causes the session to stop silently. Now the loop detects "length", injects a synthetic continuation message, and keeps going so the model can resume.

How did you verify your code works?

• bun run --cwd packages/opencode tsc --noEmit passes with no new type errors
• Changes are additive only: 51 lines added, 0 removed, no changes to tests or public API
• Manually tested by triggering large file writes that exceed OUTPUT_TOKEN_MAX and verifying the model receives the truncation error message and retries with smaller operations

Screenshots / recordings

N/A - no UI changes

Checklist

• I have tested my changes locally
• I have not included unrelated changes in this PR

[github-actions]

Thanks for updating your PR! It now meets our contributing guidelines. 👍