The claude-cli backend passed the extraction schema via --system-prompt with
only the raw file dump in the user turn, assuming a replacement system prompt
is the model's sole authority. Claude Code >= ~2.1 (verified on 2.1.197) does
not honour that: it still layers in the local coding-agent context
(CLAUDE.md/AGENTS.md in cwd, skills, MCP) and, given a user turn that is just a
file with no request, replies conversationally ("I see the file, but there's no
actual request attached"). That prose parses to zero nodes/edges, so
_response_is_hollow flags it as truncation and the adaptive-retry path bisects
the chunk indefinitely (94 -> 47 -> 23 -> ...), never converging and never
writing graph.json.
Move the full extraction schema plus an explicit imperative into the user turn
and drop --system-prompt, so the CLI emits the JSON object directly. The
<untrusted_source> prompt-injection guardrails are carried verbatim; model
override, --add-dir image handling, timeout, and token accounting are untouched.
Two Windows fixes for the claude-cli backend:
1. _call_llm (the label/community-naming path) spawned a bare ["claude", ...],
which CreateProcess cannot resolve to the npm claude.cmd shim (PATHEXT does
not apply) — every labeling batch failed with WinError 2. Mirror the
extraction path's resolution: prefer shutil.which("claude.cmd") and pass
the resolved path. Regression test included.
2. Both claude -p spawn sites now pass CREATE_NO_WINDOW on Windows. Without
it, each per-batch claude.cmd spawn allocates a console — with Windows
Terminal as the default terminal, a labeling run pops one visible window
per batch on the user's desktop for the duration of the model call.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Jeisson