GRAPHIFY_DISABLE_THINKING is ignored by the claude backend, so thinking eats max_tokens and extraction truncates #4546

Summary

GRAPHIFY_DISABLE_THINKING is honoured by the OpenAI-compatible and Ollama paths but not by the native Claude backend (--backend claude). _call_claude in graphify/llm.py sends no thinking parameter at all, so on Claude models that think by default the thinking tokens count against the response's max_tokens (16384 from BACKENDS["claude"]), and extraction truncates or comes back hollow.

Version

graphifyy 0.9.92, anthropic 1.15.0, Python 3.14, graphify extract . --backend claude --model claude-haiku-5-5 --force (headless, in CI).

What happens

Claude Haiku 5.5 runs adaptive thinking by default. With _call_claude sending no thinking setting, a full extraction of a docs-heavy repo (347 semantic files) logged:

• 38 chunks truncated at depth N, splitting into halves, 11 of them still truncated at recursion depth 3 (max 3) — partial result kept;
• 3 hollow responses, one of them content=no nodes/edges, output_tokens=13347 — almost all of the 16384 spent on thinking;
• 64/347 dispatched file(s) produced no nodes, after which the shrink guard (correctly) refused to overwrite the existing graph.

Setting GRAPHIFY_DISABLE_THINKING=1 changes nothing on this path, because _thinking_disabled_via_env() is read in the OpenAI-compatible and Ollama request builders but not in _call_claude (nor in the claude branch of the labelling call).

Expected

GRAPHIFY_DISABLE_THINKING=1 sends thinking={"type": "disabled"} on the Claude backend too, in both _call_claude and the labelling call, as it does elsewhere. Claude Haiku 5.5 accepts disabled thinking at its default effort. Optionally, a way to set output_config.effort (for example GRAPHIFY_CLAUDE_EFFORT) would let users keep thinking on but bounded.

Workaround we are using

GRAPHIFY_MAX_OUTPUT_TOKENS=21000 plus --token-budget 15000, so each smaller chunk's response, thinking included, fits. It works around the symptom; it does not stop thinking being billed on a structured-extraction task that does not need it.

Related

#730 (hardcoded max_tokens), #2880 (hollow responses), #3671 (labelling discards finish_reason='length').