[Bugfix][Parser] kimi_k2: route streaming through arg_converter so schema type coercion applies

[muhammadfawaz1]

Summary

kimi_k2 was the only Streaming Parser Engine backend with tool_args_json=True and no arg_converter. Because _compute_arg_delta and _flush_arg_converter both return early when arg_converter is None (before ever reaching _fix_arg_types), streaming emitted the model's raw JSON value types unchanged, while non-streaming always coerces to the tool schema via _build_extracted_result. Same model output, different typed values depending on whether the client streams or not — silent, since the output is still valid JSON.

Full root cause and reproduction: #49316

Fix

Entirely in vllm/parser/kimi_k2.py. Zero changes to parser_engine.py.

• Added _kimi_k2_arg_converter(raw_args, partial): normalizes kimi's raw JSON to canonical json.dumps output, using partial_json_parser (already a project dependency — see requirements/common.txt, used by 7 other tool parsers) to complete incomplete objects when json.loads fails on partial input. Unparsable text is returned unchanged so no argument bytes are lost.
• Wired it in via arg_converter=_kimi_k2_arg_converter in kimi_k2_config.
• Removed the now-incompatible _handle_arg_chunk override, which emitted the first arg chunk as raw event.value, bypassing any converter. Once a converter exists this duplicated output.

Why this approach

Once kimi has a converter, it flows through the same converter-present machinery PR #48706 already hardened: _fix_arg_types coercion, _safe_arg_prefix's withholding of unstable trailing values, and the startswith prefix invariant. Because the converter re-emits canonical json.dumps output on every tick (both partial and flush), and non-streaming routes through the same converter, both paths produce byte-identical concatenated output by construction, rather than needing to be kept in sync by hand.

Alternative considered and rejected: raw-text withholding

Adapting PR #48706's non_string_clip idea directly to the no-converter path (withhold non-string field values from the raw stream, emit them corrected at flush) was the first approach tried. It doesn't work: once a field coerces, _fix_arg_types re-serializes the entire object with json.dumps spacing ({"a":"x",...} → {"a": "x", ...}), but partial raw JSON can't be run through _fix_arg_types mid-stream (it isn't valid JSON yet), so the raw no-space prefix already streamed is never a valid prefix of the eventual coerced-and-spaced output. The startswith check then fails and silently truncates — reproducing the exact #48702 failure mode in a new place. The only prefix that's safe under both outcomes turns out to be the opening {"firstkey":, which defeats progressive streaming entirely. This is unavoidable specifically because of middle (not just trailing) non-string fields — verified by direct construction, not just argued.

Alternative considered and rejected: minimal in-place coercion

Coercing values in place while preserving the model's original raw formatting (no re-spacing) would require either changing _fix_arg_types itself (shared with the converter-present path — out of scope and risks regressing #48706's behavior) or accepting that kimi's non-streaming output diverges from the canonical format everywhere else in the engine.

Known behavior change (intentional, flagged for review)

Non-streaming output for untyped/already-correct fields is now whitespace-normalized ({"city":"Tokyo"} → {"city": "Tokyo"}) — a direct consequence of routing both paths through the same canonical json.dumps converter. Already-coerced field output is unchanged ({"count": 3} stays {"count": 3}). No existing test asserts on the old compact non-streaming format for kimi_k2 (checked: tests/parser/, tests/reasoning/test_kimi_k2_reasoning_parser.py, tests/entrypoints/unit_tests/test_chat_utils.py all pass unchanged). Flagging this explicitly in case a maintainer considers exact non-streaming byte format a compatibility concern.

Testing

• New file: tests/parser/engine/test_kimi_k2_streaming.py, 27 deterministic, byte-level tests at chunk sizes 1/3/1000, covering: single non-string field (int/bool/number), middle (non-trailing) non-string field, nested non-string field, parallel tool calls, string field with a numeric-looking literal, already-correct-types control, and truncated/never-closed JSON.
• Confirmed red→green directly: git stash + run → 21/27 fail on unpatched main (every coercion scenario); git stash pop + run → 27/27 pass.
• Independently verified with a probe script written before this fix existed (not authored alongside it) — all cases now match streaming/non-streaming, including a parallel-tool-calls case.
• Full existing suite: tests/parser/engine/ — 3699 passed, 0 regressions (includes all converter-present deepseek/qwen/gemma streaming tests, confirming this change doesn't touch that path's behavior).
• tests/parser/test_parse.py, tests/parser/test_streaming.py, tests/reasoning/test_kimi_k2_reasoning_parser.py: pass, unchanged.
• tests/entrypoints/unit_tests/test_chat_utils.py: 12 failed / 7 errored, identically with and without this change applied (confirmed via stash-diff) — pre-existing environment gap (missing Torchvision), not a regression.
• ruff check + ruff format: clean.

Not yet verified

Real-tokenizer reasoning/tool-parser suites requiring moonshotai/* tokenizer download couldn't be run in this environment (no network access to HuggingFace from the dev machine). Logic exercised is the same code path covered by the mock-tokenizer tests above, but a maintainer with the model cached should run these before merge.

Not a duplicate of

• [Bug]: [Parser] Streaming tool-call args silently truncated when non-string schema field is coerced at flush time #48702 / [Bugfix][Parser] Fix streaming tool-call args truncated by schema coercion #48706 — related but different mechanism: that bug is in the arg_converter-present path (_flush_arg_converter's startswith(prev) invariant breaking when coercion shrinks JSON). This bug is in the arg_converter is None path, which never calls _fix_arg_types at all, regardless of any prefix invariant. Confirmed no code overlap: git diff of [Bugfix][Parser] Fix streaming tool-call args truncated by schema coercion #48706 contains no reference to the arg_converter is None branch this PR modifies.
• [Bugfix] Coerce Python-repr None to JSON null for tool-call args #47643 — different bug (Python-repr None coercion).

[claude]

Claude Code Review

This pull request is from a fork — automated review is disabled. A repository maintainer can comment @claude review to run a one-time review.

[mergify]

This pull request has merge conflicts that must be resolved before it can be
merged. Please rebase the PR, @professorsab.

https://docs.github.com/en/pull-requests/collaborating-with-pull-requests/working-with-forks/syncing-a-fork