[PureSaber]

Workbench and Notebook assistants rejected otherwise valid model answers wrapped in a single Markdown JSON fence. Both entry points now share a bounded parser that accepts a complete JSON object or one whole JSON/unlabeled fence, then applies the existing answer, citation, quote, and cell-source hash checks.

Requests explicitly use Chat Completions response_format: json_object. Compatible providers may only prompt for this format, so client validation remains mandatory. Prose around a result, multiple fences/objects, malformed JSON, non-object content, and oversized raw responses remain failures. The 60KB/120KB limits, zero automatic retries, redacted CLI errors, and review-only Notebook behavior are preserved. No dependency declarations or lock pins change.

Validation:

• Reproduced the original failure in both entry points before the fix: 6 targeted failures, including the actual locked SDK with a local mock transport.
• 182 tests pass; total coverage 85.54% (required 80%); standard artifact branches 14/14 (required 90%); Ruff and pip check pass.
• Actual public CLI subprocesses against a synthetic local HTTP provider pass four cases: Workbench and Notebook complete fenced results accepted; surrounding prose rejected without output, response disclosure, or a second request.
• Provider tests inspect the real locked ChatOpenAI/OpenAI request for JSON mode and preserve the explicit model, timeout, temperature, and zero retries. These automated checks do not call a real model or establish model quality.