fix(structured): report an output-limit stop instead of a JSON parse error

[senamakel]

Summary

extract_provider_schema parses the response text without first checking whether there is any. When a reasoning model spends its whole output budget thinking and returns empty content, the error is:

schema 'tinysweeper_critique': response text is not valid JSON: expected value at line 1 column 1

That points at a malformed response which was never sent. The reader goes looking for bad JSON — the actual cause is truncation, and the fix is to raise max_tokens or cap reasoning.

How it was found

Reviewing a large pull request with moonshotai/kimi-k3 through OpenRouter. Calling the API directly with the same 49k-token prompt:

finish_reason     : length
completion_tokens : 8000      ← the entire budget
reasoning         : 17166 chars
content           : ""        ← nothing left for the answer

The provider had already said exactly what happened. ModelResponse::finish_reason carried it, and the extractor ignored it.

Worth noting for anyone hitting the same thing: OpenRouter's reasoning: {max_tokens: N} is ignored by this model — reasoning still grew past a 4096 cap. reasoning: {enabled: false} does work.

The change

Empty content is now reported as empty content. When finish_reason is length the message names the remedy; otherwise it reports the stop reason it did get. Genuinely malformed JSON still reports a parse error — there is a test pinning that, so the new branch cannot swallow the case it was not written for.

Tests

Three added in src/harness/structured/test.rs. cargo test --lib is green at 1405 passed; cargo clippy --all-targets -- -D warnings is clean.

Summary by CodeRabbit

• Bug Fixes
  • Improved error messages when structured responses are empty because the output limit was reached.
  • Added a clear missing-content error for other empty responses.
  • Preserved accurate parsing errors for malformed JSON responses.

[coderabbitai]

No actionable comments were generated in the recent review. 🎉

ℹ️ Recent review info

⚙️ Run configuration

Configuration used: Organization UI

Review profile: CHILL

Plan: Pro

Run ID: 5163b542-7da4-4489-8b38-8ac73e27f8ee

📥 Commits

Reviewing files that changed from the base of the PR and between 6655b6e and e1f8637.

📒 Files selected for processing (2)

• src/harness/structured/mod.rs
• src/harness/structured/test.rs

---

📝 Walkthrough

Walkthrough

The structured-output extractor now detects empty model responses before JSON parsing. It reports output-limit truncation separately from missing content. Tests cover both diagnostics and preserve parse errors for malformed non-empty JSON.

Changes

Structured output diagnostics

Layer / File(s)	Summary
Empty response handling and validation src/harness/structured/mod.rs, src/harness/structured/test.rs	extract_provider_schema reports a max_tokens error when finish_reason is "length" and a missing-content error otherwise. Tests cover both cases and retain malformed-JSON parse errors.

Estimated code review effort: 2 (Simple) | ~10 minutes

Poem

I’m a rabbit with JSON in my paws,
Truncation now explains its cause.
Empty hops get errors clear,
Bad JSON keeps its parse-error cheer.
Carrots celebrate the fix! 🐇

🚥 Pre-merge checks | ✅ 5

✅ Passed checks (5 passed)

Check name	Status	Explanation
Description Check	✅ Passed	Check skipped - CodeRabbit’s high-level summary is enabled.
Title check	✅ Passed	The title clearly and concisely describes the main change: reporting output-limit stops instead of JSON parse errors.
Docstring Coverage	✅ Passed	No functions found in the changed files to evaluate docstring coverage. Skipping docstring coverage check.
Linked Issues check	✅ Passed	Check skipped because no linked issues were found for this pull request.
Out of Scope Changes check	✅ Passed	Check skipped because no linked issues were found for this pull request.

✨ Finishing Touches

📝 Generate docstrings

• Create stacked PR
• Commit on current branch

---

_{Comment @coderabbitai help to get the list of available commands.}