v0.6.75

Highlights

• HarmonyStreamingRouter — delegates gpt-oss tool-call streaming to openai-harmony's StreamableParser (same library vLLM and SGLang use). Closes #444 #455 #468 #480 #513.
• Structured tool-call passthrough — surfaces (name, arguments) natively through the engine via GenerationOutput.tool_calls; routes bypass regex-based reconstruction. Eliminates the class of sentinel-anchored parsing bugs that produced "harmony channel markers leak into content delta" symptoms across the four upstream issues.

Architecture

The legacy custom state machine remains as a fallback (auto-selected when the strict three-layer compat gate — tokenizer-identity allowlist + 7-marker parity + body-vocab probe — fails). All other models (Gemma 4 OutputRouter, think-tag, harmony with mismatched IDs) continue on the existing path.

Streaming fast-path enforces:

• tool_calls[*].index monotonicity across router chunks (OpenAI delta-merge requirement)
• parallel_tool_calls=false external contract cap (matching non-streaming post-parse trim)

Issues closed

• #444 — harmony streaming tool calls leak as raw content deltas
• #455 — Anthropic /v1/messages streaming + tools: harmony commentary leaks as text block
• #468 — `tool_choice="required"` not enforced + harmony channel markers leak
• #480 — harmony streaming tool calls leak channel syntax into content (gpt-oss-20b)
• #513 — marker-preserving router redesign for end-to-end fix of the above

Validation

• 16 rounds of codex adversarial review (final round: 0 BLOCKINGs)
• pr_validate 3×3 stress matrix: qwen3.5-35B-A3B-8bit / qwen3.6-27B-MLX-8bit / gpt-oss-20b-MXFP4-Q8
• 4484 unit tests pass (+9 new for structured-streaming index + cap)
• Release-SOP gates G1-G9 walked. G10 N/A. G11 deferred to #516 (waiver: auto-upgrade falls back to in-production legacy router; no behavioral regression vs pre-PR baseline).

Dependencies

• New: `openai-harmony>=0.0.6` (soft-imported at runtime)

Installation

```
pip install --upgrade rapid-mlx
brew upgrade rapid-mlx
```