fix(claude-cli): deliver extraction instructions in the user turn (fixes infinite bisection on Claude Code >= 2.1)

[javargasm]

Summary

The claude-cli backend can hang on an infinite chunk bisection under newer Claude Code CLIs (>= ~2.1, verified on 2.1.197), so graphify extract --backend claude-cli never writes graph.json.

_call_claude_cli delivered the extraction schema via --system-prompt and put only the raw file dump in the user turn, on the assumption that a replacement system prompt is the model's sole authority. Claude Code >= 2.1 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 — what would you like me to do with it?"

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. The run ends in all semantic chunks failed for backend 'claude-cli'.

Observed log:

[graphify] LLM returned invalid JSON, skipping chunk (first 200 chars:
  'I see the file ... but no actual question or task attached...')
[graphify] claude-cli returned a hollow response; treating as truncation ...
[graphify] chunk of 94 truncated at depth 0, splitting into halves of 47 and 47
... 47 -> 23 -> 11 ...

Fix

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 (the schema text is unchanged; only its delivery channel moves). Model override (GRAPHIFY_CLAUDE_CLI_MODEL), --add-dir image handling, timeout, and token accounting are untouched.

Why not other approaches

Verified empirically before landing this:

• Model isn't the lever — --model haiku fails the same way.
• cwd isn't the lever — running from a clean dir with no CLAUDE.md still returns prose.
• Isolation flags don't work — --setting-sources "" / --tools "" / --mcp-config '{}' break auth (Not logged in) or hang.

Moving instructions into the user turn is the only approach that makes the CLI emit JSON.

Verification

• Real extract on 4 repos (Python/TS/YAML) now writes graph.json with no bisection: e.g. 38 nodes / 38 edges, 3468 / 11390, 679 / 1334, 3723 / 7302.
• The pre-existing test_uses_system_prompt_not_append (which asserted the old behaviour that caused this bug) is updated to assert instructions ride in the user turn.
• Added regression tests: --system-prompt absent from argv; schema + imperative + source present in stdin; <untrusted_source> guardrails preserved.
• pytest tests/test_claude_cli_backend.py tests/test_llm_parser.py tests/test_llm_backends.py → 116 passed.

Notes

• The est. cost (~claude-cli): $0.0000 line is unaffected by this change, but note extraction runs on the Claude Code plan (Opus by default); GRAPHIFY_CLAUDE_CLI_MODEL=haiku reduces that.

[safishamsi]

Merged into v8 as 32ff6d6 (your authorship). Nicely diagnosed — the CC >=2.1 conversational-reply-to-bare-file-dump causing _response_is_hollow to bisect forever is exactly the failure. Verified --system-prompt is dropped and the extraction schema (with the <untrusted_source> guardrails intact) now rides in the user turn. High-impact for anyone on the default claude-cli backend post-2.1. Full suite 2841. Ships next release.