feat: route Qwen3-Coder + DeepSeek-V4 tool calls through dynamo-parsers-v2 (DYN_ENABLE_EXPERIMENTAL_PARSERS_V2)

[keivenchang]

Overview

Routes Qwen3-Coder and DeepSeek-V4 tool calls through the dynamo-parsers-v2 streaming parser, gated behind a new DYN_ENABLE_EXPERIMENTAL_PARSERS_V2 env flag. When the flag is set, both the batch and the streaming path bypass the v1 tool-call jail and use the v2 parser — which owns incremental tool-call emission and drops a parameter value truncated at EOF instead of guessing it. Off by default; the v1 jail / aggregate-finalize path is unchanged.

Changes

• New DYN_ENABLE_EXPERIMENTAL_PARSERS_V2 flag (const + registry in environment_names.rs), read once via a cached LazyLock<bool>.
• Generalized the Qwen3-Coder-only module qwen3_coder_v2.rs into a family-agnostic tool_parser_v2.rs: enabled(), supports_family() = {qwen3_coder, deepseek_v4}, and apply_stream / parse_complete built on dynamo_parsers_v2::create_tool_parser_for_family(family, …) returning Box<dyn ToolParser>. Adding a family is a one-line change here plus support in the crate.
• Both dispatch sites gated on enabled() && supports_family(parser): streaming (preprocessor.rs) and batch (aggregator.rs).
• Adds the published dynamo-parsers-v2 0.1.10 dep. The dynamo-parsers 2.1.2 bump merged separately in chore: bump dynamo-parsers to 2.1.2 #10899; this branch is rebased onto main on top of it.

Behavior

• Flag unset (default): no change — every family uses the v1 jail.
• Flag set: qwen3_coder and deepseek_v4 route through v2 for both batch and streaming.

Tests

• New tool_parser_v2 unit tests: Qwen3-Coder incremental stream + truncated-at-EOF drop, plus a DeepSeek-V4 DSML incremental stream (reconstructs one get_weather(location=NYC) call, no markup leak).
• cargo test -p dynamo-llm: 3 unit + 149 integration (test_jail 69, streaming 36, tool_choice 36, parallel 8) pass — the env-gating leaves the v1 path unchanged.

Enabling the experimental v2 path

Off by default. Set DYN_ENABLE_EXPERIMENTAL_PARSERS_V2 (truthy: 1/true/yes/on) at process startup to route the supported families (qwen3_coder, deepseek_v4) through the v2 parser instead of the v1 jail, for both batch and streaming. The flag is read once per process via LazyLock, so it's a startup switch (not toggleable mid-run). Other families, and non-auto tool_choice (required/named), always stay on v1.

In fixture / Rust tests — prefix the test run with the env var (applies to the whole test binary):

# v1 (default)
cargo test -p dynamo-llm --test postprocessor_parsing_stream tool_calls_qwen3_coder_auto_routes_through_experimental_gate
# v2 (experimental)
DYN_ENABLE_EXPERIMENTAL_PARSERS_V2=1 cargo test -p dynamo-llm --test postprocessor_parsing_stream tool_calls_qwen3_coder_auto_routes_through_experimental_gate

tool_calls_qwen3_coder_auto_routes_through_experimental_gate is the gate test: it routes a qwen3_coder + auto tool-call stream through the dispatch, so the same assertion exercises v1 with the flag off and v2 with it on. CI runs with the flag off.

In real Dynamo — export it for the frontend process:

DYN_ENABLE_EXPERIMENTAL_PARSERS_V2=1 python -m dynamo.frontend --http-port 8000

Related

DIS-2228 — Qwen3-Coder v2 parser + env-gated Dynamo routing

Summary by CodeRabbit

• New Features

  • Added experimental support for improved tool-call parsing in streaming and non-streaming chat responses for select model families.
  • Introduced a new opt-in setting to enable the experimental parser path.

• Bug Fixes

  • Reduced the chance of malformed tool-call output leaking into visible text.
  • Improved handling of incomplete responses at the end of a stream so tool calls are finalized more safely.

[coderabbitai]

Walkthrough

Adds the parser-v2 dependency and environment flag, introduces a new OpenAI chat-completions tool parser module, routes supported tool-call flows through it, and adds unit and integration tests for the experimental path.

Changes

Experimental parser-v2 routing

Layer / File(s)	Summary
Dependency and gate surface Cargo.toml, lib/llm/Cargo.toml, lib/runtime/src/config/environment_names.rs, lib/llm/src/protocols/openai/chat_completions.rs	Adds dynamo-parsers-v2 to the workspace and llm dependencies, defines DYN_ENABLE_EXPERIMENTAL_PARSERS_V2, and exports tool_parser_v2.
Batch parser and family support lib/llm/src/protocols/openai/chat_completions/tool_parser_v2.rs	Defines the supported families, the v2 enablement check, v1-to-v2 tool translation, and batch parsing into OpenAI tool calls plus normal text.
Streaming bypass and unit tests lib/llm/src/protocols/openai/chat_completions/tool_parser_v2.rs	Implements the streaming v2 path with per-choice parser state, terminal flushing, and passthrough fallback, plus unit tests for incremental reconstruction and EOF truncation cases.
OpenAI routing and regression coverage lib/llm/src/preprocessor.rs, lib/llm/src/protocols/openai/chat_completions/aggregator.rs, lib/llm/tests/postprocessor_parsing_stream.rs	Routes eligible tool-call streams through parser-v2 in preprocessing and final aggregation, and adds an integration test for the qwen3_coder Auto path.

Estimated code review effort

🎯 4 (Complex) | ⏱️ ~60 minutes

🚥 Pre-merge checks | ✅ 5

✅ Passed checks (5 passed)

Check name	Status	Explanation
Title check	✅ Passed	The title clearly names the main change: experimental v2 routing for Qwen3-Coder and DeepSeek-V4 tool calls.
Description check	✅ Passed	The description covers overview, changes, behavior, tests, and related issue info, though it lacks an explicit reviewer-start section.
Docstring Coverage	✅ Passed	Docstring coverage is 100.00% which is sufficient. The required threshold is 80.00%.
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.

_{✏️ Tip: You can configure your own custom pre-merge checks in the settings.}

---

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

[devin-ai-integration]

Devin Review found 1 new potential issue.