fix: consolidate small correctness fixes

[jxnl]

Summary

Consolidates a reviewed set of small correctness and documentation fixes onto current main:

• preserve retry/cache inputs and normalize edge-case message content
• make missing tool calls retryable for OpenAI-compatible and Mistral handlers
• return all Anthropic parallel tool errors and disable parallel calls for forced single-tool requests
• harden streaming JSON, partial models, citations, and lazy handler registration
• avoid mutating cached schemas and caller-owned Gemini configuration
• fix Bedrock default model forwarding, OpenAI audio formats, empty batch objects, and CLI file access
• redact additional credential aliases and declare the runtime regex dependency explicitly
• refresh Cerebras and Google/Vertex documentation

Consolidated PRs and issues

Canonical patchsets incorporated and intentionally superseded by this PR:

#2494 #2491 #2489 #2488 #2487 #2486 #2484 #2478 #2476 #2475 #2473 #2467 #2463 #2456 #2455 #2453 #2452 #2451 #2448 #2447 #2443 #2441 #2440 #2439 #2430 #2415 #2406 #2343

Duplicate or narrower alternatives superseded by the consolidated implementation:

#2492 #2466 #2462 #2460 #2458 #2442 #2425

Issues fixed:

#2490 #2485 #2477 #2465 #2464 #2461 #2459 #2454 #2450 #2431 #2422 #2416

Validation

• uv run pytest <all 22 changed test modules> -q: 688 passed, 2 skipped
• targeted regression reruns: 34 passed
• uv run ruff format --check instructor tests: passed
• uv run ruff check instructor tests: passed
• scoped uv run ty check for every changed production and test module: passed
• uv lock --check: passed
• commit hooks: Ruff lint/format, lock check, dependency installation, requirements export, and ty all passed
• broader pytest tests/ -k 'not llm and not openai' -q: 3050 passed, 229 skipped; 48 unrelated failures remain in pre-existing docs snippets and live provider calls using unavailable/retired models
• repository-wide ty check --config-file ty-tests.toml remains baseline-red with 265 unrelated diagnostics in notebooks, optional examples, and scripts

Skipped items

• fix(retry): attach _total_usage to successful retry responses #2474 and feat: expose cumulative token usage on success + budget enforcement #2392: token-usage accumulation remains semantically unresolved by Retry usage accumulation drops newer billable token fields (Anthropic stale, OpenAI overwritten with None) #2493; merging a partial fix would make the API harder to correct later
• chore(deps): bump the poetry group across 1 directory with 38 updates #2433: broad 38-package dependency bump
• feat(requesty): add Requesty provider #2435 feat(sarvam): add Sarvam provider support #2423 feat(azure): add Azure AI Foundry provider support #2384 feat(orcarouter): add OrcaRouter provider integration #2322 feat(minimax): add MiniMax provider support #2306 feat(providers): add Manifest provider support #2283: provider additions requiring dedicated ownership and integration validation
• feat: schema complexity analyzer + automatic mode selection #2394 refactor(v2): centralize provider client contracts #2355 test(v2): validate provider-owned client architecture #2356 docs(v2): document provider-owned client architecture #2357: architectural changes and mode/client contract work
• feat(security): add MemoryGuard and sensitive_field_guard for OWASP ASI06 prompt injection defense #2321 fix(validation): harden llm_validator prompt isolation #2307 fix: preserve true partial streaming behavior #2263 feat(batch): support for mistral api #2168 feat(bedrock): add native structured outputs and strict tool use modes #2086: nontrivial security, streaming, batch, or Bedrock behavior changes
• example/resource PRs remain open for separate product/editorial review

No release, tag, PyPI publication, deployment, or social post is included.

---

Note

Medium Risk
Touches core v2 retry, caching, streaming parsers, and multiple provider handlers; changes are targeted correctness fixes with broad test coverage but wide blast radius across integrations.

Overview
This PR bundles many small runtime and integration fixes across v2, plus doc and dependency updates.

Retries, caching, and messages: Retry paths now get isolated copies of messages / contents / chat_history so reask mutations do not break cache keys or caller-owned lists. Message dumping keeps legacy function_call when content is empty; merging treats None content as empty.

Tool-call parsing and reasks: OpenAI-compatible and Mistral handlers raise retryable ResponseParsingError when there are no tool calls, and reask flows fall back to user corrections instead of iterating None. Anthropic forced single-tool requests set disable_parallel_tool_use, and parallel-tool reasks emit a tool_result for every tool_use id.

Streaming / DSL / registry: JSON stream extraction validates balanced spans and keeps scanning past non-JSON brace blocks. Partial streaming uses per-context recursion guards, builds partial instances for incomplete nested/list items, and final-validates from raw JSON so explicit null is preserved. Citation fuzzy matching escapes regex metacharacters; lazy mode-handler registration is locked against concurrent first access. OpenAI tool schemas are copied before adding strict so the LRU cache is not poisoned; Gemini generation_config is copied before in-place mapping.

Other runtime: Batch parsing accepts {} for models with defaults. CLI file listing uses SDK attribute access. Bedrock auto-client forwards default model. OpenAI audio maps WAV/MP3 only. Debug logging redacts more credential key aliases. regex is a direct dependency.

Docs: Cerebras examples use gpt-oss-120b; Google docs clarify provider prefixes; Vertex from_genai() example passes model outside the client constructor. CHANGELOG [Unreleased] entries document the above.

^{Reviewed by Cursor Bugbot for commit 3ccb56a. Configure here.}

[cloudflare-workers-and-pages]

Deploying with    Cloudflare Workers

The latest updates on your project. Learn more about integrating Git with Workers.

Status	Name	Latest Commit	Preview URL	Updated (UTC)
✅ Deployment successful! View logs	instructor	b3be98d	Commit Preview URL Branch Preview URL	Jul 29 2026, 04:30 AM