Expose standalone tool healing utilities

[alkinun]

What changed

• Added a lightweight unsloth_core.tool_healing module for parsing and stripping model-emitted tool-call markup without importing the training stack.
• Added compatibility exports under unsloth.core.tool_healing for existing Unsloth import paths.
• Updated Studio llama.cpp inference to use the shared helpers when available.
• Added a backend-local fallback so studio/backend remains importable when launched as a standalone tree without the repo root on PYTHONPATH.
• Added regression tests for JSON tool calls, Ollama-style keys, function-style markup, no-stack imports, and standalone Studio backend imports.
• Updated the Studio Playwright default-model expectation to match the current DEFAULT_MODELS_GGUF[0] (unsloth/Qwen3.6-27B-MTP-GGUF).

Why

External inference servers can reuse tool healing without importing torch, transformers, or the full Unsloth training runtime.

Closes #5502

Tests

• python -m pytest tests/python/test_tool_healing_public_api.py -q
• python - <<'PY' ... assert EXPECTED_DEFAULT == DEFAULT_MODELS_GGUF[0] ... PY
• python -m py_compile unsloth_core/tool_healing.py unsloth_core/__init__.py unsloth/core/tool_healing.py unsloth/core/__init__.py studio/backend/core/tool_healing.py studio/backend/core/inference/llama_cpp.py tests/python/test_tool_healing_public_api.py tests/studio/playwright_chat_ui.py

[gemini-code-assist]

Warning

You have reached your daily quota limit. Please wait up to 24 hours and I will start processing your requests again!

[alkinun]

Supersedes #5571. The branch has been rebased onto current main and squashed to one clean commit.

[danielhanchen]

Pushed a follow-up commit on top of your branch addressing the review:

1. Rebased onto current main. The playwright_chat_ui.py hunk in this PR landed at roughly the same time as studio: read Playwright default model from defaults.py without importing it #5595, which added an ast.literal_eval helper for exactly the no-torch / no-structlog Playwright lane. Resolved the conflict by keeping studio: read Playwright default model from defaults.py without importing it #5595's helper rather than reintroducing the hardcoded literal.
2. tests/python/test_tool_healing_public_api.py
   • The test_studio_backend_llama_cpp_imports_without_unsloth_core test imported core.inference.llama_cpp, which transitively pulls structlog via core/inference/__init__.py -> orchestrator. That made the test fail in any minimal CI lane. Renamed to test_studio_backend_local_tool_healing_fallback and switched to importing core.tool_healing directly. That tests the actual claim (the fallback module works when unsloth_core is blocked) without dragging in the orchestrator.
   • Replaced cwd="/tmp" with tempfile.gettempdir() so the test runs on Windows.
3. studio/backend/core/tool_healing.py now mirrors heal_tool_call_content. The fallback's public surface is identical to the canonical module so anyone reaching the fallback path does not silently lose that helper.
4. tests/python/test_tool_healing_duplicates_in_sync.py pins TOOL_CALL_MARKUP_SIGNALS, parse_tool_calls_from_text, strip_tool_call_markup, and heal_tool_call_content outputs between the canonical and fallback modules so the duplicate cannot drift.
5. pyproject.toml: dropped the redundant "unsloth_core*" from setuptools.packages.find since the existing "unsloth*" glob already matches it.

One thing worth calling out in the description: I ran the new module against the previous inline parser in studio/backend/core/inference/llama_cpp.py on a corpus of 40 inputs. 31 are bit-for-bit identical; the other 9 are improvements (the old code emitted arguments as raw None/list/int/bool in some paths, which violates the OpenAI Chat Completions spec, and it appended tool calls with empty/null name). The Ollama-style tool_name / tool_arguments support is the documented new behaviour. None of the diffs are regressions, but it is worth a sentence in the PR body so consumers know to expect strictly-string arguments and that empty/null-name calls are now dropped instead of forwarded.

Local run (Python 3.14, no torch / transformers / structlog): all 11 tests pass.

[alkinun]

Thanks!

[danielhanchen]

Reworked and squashed the branch to a single commit.

Dropping the new top-level unsloth_core namespace was the main goal here. Sibling packages we ship live under unsloth-zoo or under studio.backend.*; introducing a brand new top-level Python package on top of that just for one stdlib helper file is more surface area than it is worth, especially when the helper is conceptually a piece of llama.cpp inference logic.

The new home is studio/backend/core/tool_healing.py. The full import chain studio.backend.core is already stdlib-only at import time (the __init__.py there is lazy via __getattr__), so from studio.backend.core.tool_healing import parse_tool_calls_from_text, strip_tool_call_markup pulls zero heavy modules and satisfies #5502 without the new package.

What this commit does:

1. Verbatim extraction. The regexes (_TOOL_CLOSED_PATS, _TOOL_ALL_PATS, _TC_JSON_START_RE, _TC_FUNC_START_RE, _TC_END_TAG_RE, _TC_FUNC_CLOSE_RE, _TC_PARAM_START_RE, _TC_PARAM_CLOSE_RE) and the parser / stripper bodies are byte-for-byte identical to the inline code that previously lived in llama_cpp.py. The 9 behavioural deltas the previous revision introduced (Ollama-style tool_name / tool_arguments keys, JSON-stringify-any-non-string arguments, empty/null-name drop, heal_tool_call_content) are all reverted. This PR is a pure code move; no behaviour change.
2. llama_cpp.py consumes the helpers via a normal relative import. _parse_tool_calls_from_text becomes a one-line delegate; the _strip_tool_markup closure keeps the if not auto_heal_tool_calls: return text fast path and delegates the substantive work. The inline _TOOL_XML_SIGNALS = ("<tool_call>", "<function=") if auto_heal_tool_calls else () and the any(s in ... for s in _TOOL_XML_SIGNALS) guards stay exactly as they were on main.
3. AST-deterministic verification test. tests/python/test_tool_healing_extraction_is_exact.py asserts the regex literals are byte-identical to the OLD inline source via direct re.compile(...) equality, asserts the function shapes via ast.parse walks, and runs a 16-input corpus through both the extracted helper and an in-test reference implementation copied verbatim from the OLD source. Any future edit must keep them in lock-step or this test fails. 5/5 tests pass on Python 3.14 in a minimal env with no torch / transformers / structlog / httpx / numpy installed.
4. Cleanup. unsloth_core/, unsloth/core/, tests/python/test_tool_healing_public_api.py, and tests/python/test_tool_healing_duplicates_in_sync.py are gone. The redundant "unsloth_core*" glob in pyproject.toml is gone too. The Playwright change from the original branch was dropped during the rebase (it would have undone studio: read Playwright default model from defaults.py without importing it #5595).

Final diff: 3 files, +485 / -136. PR is single-commit and mergeable.

[danielhanchen]

Thanks a lot! I removed unsloth_core since it'll sadly cause more folders / files - folks can still now use the tool healing though from a separate file