[jayleaton]

What this changes

Under tool_choice: "auto" (or none given), a tool marked "strict": true was not held to its schema. The model could call it with arguments that fail validation. Now, when any offered tool is strict, the server compiles an xgrammar triggered_tags grammar. Free text stays free. At the template's call opener a call tag begins, in which strict tools' arguments follow their schema and non-strict tools take any JSON object. Required and named calls are unchanged. LFM2, and engines without the grammar library, stay unconstrained.

This changes default behaviour for every family that has a grammar runtime, not only DeepSeek: an auto request with a strict tool is now constrained. TENSORFOLD_TOOL_GRAMMAR=1 holds every auto tools request the same way. TF_DSV41_TOOL_GRAMMAR is read as an alias, because the published DeepSeek recipe sets it. Say if you would rather not carry that alias.

A second commit makes the non-strict fallback explicitly open (additionalProperties: true). The pinned xgrammar v0.2.8 already treats the bare {"type":"object"} as open, so this fixes no observed bug. It makes the intent explicit and adds a test against the real library: zig build test-grammar-tools -Dxgrammar=DIR compiles the mixed strict lookup(q, n) / non-strict now(tz) auto grammars through a real libtfgrammar in DeepSeek DSML, hermes JSON and Qwen XML markup. It checks that now(tz="UTC") and nested arguments are accepted, and that wrong-type, extra and missing strict lookup arguments are refused, with parallel calls on and off.

Issue: none open upstream. Stacked on: #1684 (DSML required tools), whose deepseek style this uses for the call opener. Review only e9f668512bb2..d5855e035461.

Receipt

• Environment, with the TensorFold commit, MLX or PyTorch version, machine and GPU, checkpoint and revision: head d5855e035461 on main d2ad5a7f2; Zig 0.17.0; host checks on a Linux aarch64 DGX Spark (GB10, CUDA 13.0). Real-pack boot on two GB10s, TP=2 over RoCE, checkpoint jayleaton/DeepSeek-V4.1-Flash-1017-EXL3-2.8bpw revision 859396591c34, TF_EXPERIMENTAL=deepseek.
• Exactness: n/a; this changes which outputs are admitted for strict tools, not token arithmetic.
• Decode speed before and after: n/a for requests without strict tools. With strict tools, the grammar mask runs on each token.
• Prompt speed before and after: n/a.
• Host tests (tool_grammar_auto_test.zig, tool_required_test.zig): the grammar shape for strict-only and mixed tools, the opener per style, the env override, and unchanged required and named calls. test-grammar-tools runs the real-library cases above.
• Tests run, with pass, fail and skip counts:

  zig fmt --check build.zig zig                      exit 0
  git diff --check d2ad5a7f2..HEAD                    exit 0
  python3 tools/zig/lean_check.py                     base 424 problems, head 424, new 0 (lean_new.py)
  zig build test-cuda-host --summary all              exit 0, 273/273 steps, 4086/4095 tests passed, 9 skipped (with test-server-cpu)
  zig build test-server-deepseek-chat-host test-tool-visibility-http-host test-golden test-grammar-tools -Dxgrammar=<xgrammar v0.2.8>  exit 0, 16/16 steps, 82/82 tests passed
  

• Real-pack TP=2 boot of main + every PR in this series (and the DeepSeek image path that is not yet proposed), at 83ba53298924: passed. Both ranks reached readiness; the recipe smoke, the canary and a strict tool call passed; MissingTritonVariant 0 in both rank logs; 2048 greedy tokens in 51.1 s with token_sha256 7538b20557b1…, the same as the downstream build of this series. On that boot, a required strict lookup returned {"q":"tensor parallelism","n":3}, and an auto request to the non-strict now returned {"tz":"UTC"}.
• Not run, and why: a tool-call accuracy comparison against main on a model with strict tools.
• Overlap: atc overlap reports the same additive tool_grammar.zig and reply_text.zig conflicts with minimax_m3: a CUDA backend for MiniMax-M3.1-Flash at TP over GB10s, from the release's own formats #1388, minimax_m3: serve MiniMax-M3.1-Flash (428B-A23B MoE, images) natively on Metal #1235 and minimax_m3 (CUDA): draft several streams' DSpark blocks in one pass #1426 as the PR this sits on.

Checklist

• Every token still goes through the lane rounds. No serial path, and nothing that needs drafts off.
• Drafted output equals "draft": false, a resumed prompt equals a fresh one, and concurrent equals solo. (Not measured for this change alone.)
• No precision traded for speed. If the bits change, the description says which sums change and why.
• A new or changed GPU kernel is registered with the kernel contract (docs/kernel-contract.md); each backend's SUMMARY line is in the receipt. (No kernel changes.)
• Prompt processing is no slower than the last release. (Server grammar only.)
• The description says which platforms I ran, Metal M1 to M5 and CUDA, and which I could not.
• New tests fail before the change, pass after it, and skip cleanly without their dependency. (No reverted-fix control was run for this PR.)
• Comments and docstrings are one line. No measurements or history in the source.
• No personal data, machine names, internal hosts or local paths. No AI attribution lines.

[jayleaton]

Restacked on the rebased #1684 (e9f668512bb2), with no changes of its own. The host gates, including test-grammar-tools against xgrammar v0.2.8, were re-run on the new head and pass; the description has the new range and numbers.

[ashhart]

Thank you @jayleaton. This is merged into main with your commits under your own name, so you're listed as a TensorFold contributor. It ships in the next release, credited to you in the changelog and the release notes.

[ashhart]

Thanks @jayleaton! Strict tools held to their schema under tool_choice: "auto" are in 1.2.1: https://github.com/ashhart/TensorFold/releases/tag/v1.2.1