feat(qwen35): render tool calls in native XML, not legacy Hermes JSON

[nwoolmer]

Follow-on to #508 (the dropped-outer-brace grammar fix). That PR made the Hermes-JSON grammar correct; this PR stops forcing qwen3.5/3.6 into Hermes JSON at all, since those models are XML-native (<function=NAME><parameter=ARG>). Emitting a JSON example / re-rendering history as JSON fights the model's training and is what induced the drift in the first place.

Changes

1. Per-model grammar gating (cli/index.ts, registry.json): resolveToolGrammar() picks the tool format from the registry default_tool_format (added "hermes" to the 4 carnice cards) or a hermes/carnice heuristic. Non-Hermes qwen defaults the Hermes JSON grammar OFF → the model emits its native XML; Hermes-format models (carnice) keep it ON. Applied in serve() too.
2. CLI tools-block example → native XML (Part A): for qwen3_5/qwen3_5_moe with the grammar off, the injected example is <tool_call><function=><parameter=> instead of the Hermes JSON one. Keyed on the live grammar env so the example always matches what the model will actually emit.
3. Cache-miss history re-render → native XML (Part B): build_cached_history's ChatScaffold miss branch (the HIPFIRE_JINJA_CHAT=0 hand-rolled fallback) gains a ToolCallRender param; grammar-off qwen re-renders a historical tool_call as native XML via render_tool_call_qwen_xml(). Both call sites (LCP planning in plan_prompt_cache + the actual render) read the same env-derived format so their token streams stay byte-consistent.

Note on scope

On the default path qwen3.6 already renders through its bundled chat_template (jinja_active is on by default), which is native XML for both the tools-block and history — so the real forcing signal was the grammar (now off for non-Hermes qwen). Parts A/B additionally harden the HIPFIRE_JINJA_CHAT=0 hand-rolled fallback so it's consistent too.

Testing

• prompt_frame red→green tests for both XML and Hermes-JSON history render; full hipfire-runtime lib (315) + hipfire-arch-qwen35 grammar (75) green.
• Live on qwen3.6-27b.mq4: single-turn 5/5 angle-bracket-heavy writes clean (heredoc <<, #include <stdio.h>, <vector>, HTML, a < b); multi-turn (history tool_call + tool result → new call) clean; the HIPFIRE_JINJA_CHAT=0 cache path exercised (eligible=true, Assistant tool_call miss → XML render, lcp=256 forward extension).

🤖 Generated with Claude Code

[claude]

Claude finished @nwoolmer's task in 3m 4s —— View job

---

Review complete

• Gather context (PR diff, changed files)
• Review cli/index.ts (resolveToolGrammar, tools-block example)
• Review cli/registry.json (default_tool_format additions)
• Review crates/hipfire-runtime/examples/daemon.rs
• Review crates/hipfire-runtime/src/prompt_frame.rs (ToolCallRender, XML render)
• Post final review

Verdict: Minor — one real gate-coverage gap, rest is nits.

The core logic is clean and consistent: resolveToolGrammar() (cli/index.ts:1431-1443) and qwen_history_tool_render() (daemon.rs:3641-3647) both key off the same "not explicitly 0" test, so tool-call render and grammar enforcement stay in lockstep at both build_cached_history call sites (plan_prompt_cache LCP planning + the actual render). Tests (prompt_frame.rs:1700-1786) correctly assert XML-vs-JSON exclusivity in both directions.

Findings:

• Gate coverage gap — scripts/coherence-gate.sh:44 and scripts/agentic-gate.sh:158 spawn ./target/release/examples/daemon directly, never setting HIPFIRE_QWEN35_GRAMMAR. Since the new native-XML default is only applied by the CLI's applyConfigEnv (cli/index.ts:1456-1458) before it spawns/talks to the daemon, the raw daemon still defaults grammar_enabled = true (Hermes JSON) at daemon.rs:4126/5916/8456 when launched directly. That means the behavior this PR actually ships to users (native XML for non-Hermes qwen3.5/3.6 via hipfire serve) is not exercised by the existing gate scripts — they keep validating the old Hermes-JSON path. The PR's "Live on qwen3.6-27b.mq4" testing presumably went through CLI serve, not through these gates; worth confirming, and worth a follow-up to make the gates model the CLI's env resolution (or explicitly set HIPFIRE_QWEN35_GRAMMAR=0) so a future regression in the new default path gets caught.
• Minor: render_tool_call_qwen_xml (prompt_frame.rs:384-402) silently emits <function=NAME></function> with no parameters if tc.arguments isn't a JSON object (the HermesJson branch doesn't lose data in that case since it embeds the raw Value). Low risk since tool-call arguments should always parse as an object, but worth a debug_assert or comment if that invariant isn't enforced upstream.
• Minor (self-healing, noted only for awareness): the XML render doesn't escape parameter values, so a code-writing tool call whose argument literally contains </parameter> or </function> would desync the historical token stream on cache-miss replay — but this is documented as an accepted degrade-to-miss case (daemon.rs:3652-3660, LCP divergence forces a reset), and mirrors what the model's own upstream chat_template does, so not a regression.

cli/registry.json's new default_tool_format fields are additive and the CLI fallback heuristics (regex on model tag / desc) already covered the same carnice cards before this PR, so no behavior-breaking risk if registry/v1.json lags behind before the next registry-gen refresh.

GPU gates still required: scripts/coherence-gate.sh (qwen3.5/3.6 rows) recommended before merge given the tool-call format flip, but per the gap above, run it with HIPFIRE_QWEN35_GRAMMAR=0 explicitly (or via hipfire serve) to actually cover the new default path — as shipped, the plain gate invocation won't. scripts/serve-multiturn-gate.sh is also relevant since Part B changes the prefix-cache render used for cross-turn LCP byte-consistency (plan_prompt_cache + the actual render both read qwen_history_tool_render()) — worth a run to confirm no state-bleed across a multi-turn tool-call session.

[Kaden-Schutt]

/gate

[Kaden-Schutt]

/gate

[Kaden-Schutt]

/gate

[Kaden-Schutt]

/gate

[Kaden-Schutt]

/gate

[Kaden-Schutt]

/gate

[Kaden-Schutt]

/gate

[Kaden-Schutt]

/gate

[Kaden-Schutt]

/gate

[github-actions]

GPU PR Gate (Tier 3)

arch	verdict
gfx1100	REJECT
gfx1151	REJECT
gfx1201	PASS
behavior	REJECT

Bill of Debt (blocking):

• gfx1100 probe_unavailable — required box never produced a probe result; this arch is required per validated_dispatch.json.required_archs and cannot be silently skipped.
• gfx1151 probe_unavailable — same failure mode as gfx1100; required box unreachable.
• behavior qwen35-native-xml-toolcall — plain (non-carnice) qwen3.6 model did not emit native XML tool calls (<tool_call>\n<function=write_file>\n<parameter=...>) on the angle-bracket code-writing prompt; this is the PR's own stated regression trigger.
• behavior qwen35-history-cache-miss-xml-replay — under HIPFIRE_JINJA_CHAT=0, the turn-1 historical tool_call re-rendered on cache-miss did not stay native XML and/or the turn-2 follow-up tool call was not coherent.
• behavior qwen35-grammar-explicit-override — explicit HIPFIRE_QWEN35_GRAMMAR=1 override did not force Hermes JSON tool-call rendering for the plain qwen3.6 model, indicating applyConfigEnv's === undefined guard may be clobbering a user override.

Actionable next steps:

1. Restore probe availability on gfx1100/gfx1151 (hipx box) and re-run — these are required archs and cannot pass without a probe result.
2. Re-verify qwen_history_tool_render() and render_tool_call_qwen_xml actually produce XML-delimited tool calls (not Hermes JSON) for non-carnice qwen3.6 models across both daemon.rs call sites (plan_prompt_cache and the real render), per review.md's LCP byte-consistency note.
3. Re-check resolveToolGrammar's default_tool_format/regex fallback and applyConfigEnv's === undefined guard for the explicit-override precedence bug flagged by the failing hiptrx behavior test.

This PR is blocked until the above are resolved and all three required archs (gfx1100, gfx1151, gfx1201) plus all three behavior tests pass.

[nwoolmer]

/gate

[Kaden-Schutt]

Folded into #534 as a cleanly rebased/cherry-picked delta on beta for v0.3.0, preserving the newer Redline and modular-runtime line. Closing this source PR in favor of the consolidated release PR. Thank you for the contribution.