Feature request: `--backend claude-cli` (route through Claude Code CLI, no `ANTHROPIC_API_KEY` needed)

[spindle79]

Feature request: --backend claude-cli — route through the local Claude Code CLI instead of requiring ANTHROPIC_API_KEY

Package: graphifyy v0.7.16
Use case: Claude Code Pro/Max subscribers running graphify's semantic pass on their own codebase.

What

Add a claude-cli backend that shells out to the locally-installed claude CLI (Claude Code) via claude -p --output-format json instead of calling the Anthropic API directly.

Why

--backend claude currently requires ANTHROPIC_API_KEY — pay-as-you-go API billing, separate from the Claude Code subscription. For Pro/Max subscribers, this means:

1. Provisioning a second Anthropic account (or a billing key on the same one) just to run graphify's one-off semantic pass.
2. Paying twice — once via the subscription (which they're already using for Claude Code itself) and once via API credit (for graphify).

The local claude CLI is already authenticated via the user's subscription OAuth and supports a non-interactive -p mode with --output-format json that returns a parseable envelope including the model result, usage stats, stop reason, and modelUsage map. It can serve as a drop-in transport for graphify's _call_claude() use case.

Verified on a real codebase: ran graphify's full semantic extract (167k input / 8k output tokens, Opus 4.7 1M context) entirely through claude -p, billed against my Max plan instead of API credit.

Approach

Mirror the existing claude backend with a new claude-cli arm:

• New BACKENDS["claude-cli"] entry with env_key: None and zero pricing (calls bill against the plan, not the API — graphify's cost estimator treats this as $0).
• New _call_claude_cli(user_message, max_tokens) that:
  • Verifies shutil.which("claude") finds the binary.
  • Subprocess-calls claude -p --output-format json --append-system-prompt <_EXTRACTION_SYSTEM> with the user message piped on stdin.
  • Parses the JSON envelope: extracts result (the model output), sums usage.input_tokens + cache_read_input_tokens + cache_creation_input_tokens for the input token total, and maps stop_reason == "max_tokens" to graphify's finish_reason == "length".
• New dispatch arm in extract_files_direct plus the lower _call_llm helper.
• New env-key carve-out in __main__.py extract validation: when backend == "claude-cli", check shutil.which("claude") instead of an env variable. Mirrors ollama's local-loopback carve-out.

The user must have authenticated claude interactively at least once so the OAuth tokens are stored locally; after that, -p mode runs against the subscription without further prompts.

Notes for the maintainer

• The current Claude Code CLI's --bare mode (which would strip the default system prompt) requires ANTHROPIC_API_KEY and would defeat the purpose, so we use --append-system-prompt instead. The base Claude Code prompt is appended, but the _EXTRACTION_SYSTEM prompt is strict enough that the model produces valid JSON anyway. Verified empirically — see PR for test fixture using a real Claude Code response envelope.
• Tests in the PR mock subprocess.run and shutil.which so the suite still runs on CI without the claude binary or network access.
• Performance is comparable to the API backend: the Claude Code CLI caches its system prompt server-side, so subsequent calls in a session benefit from the same cache discount as direct API calls.

PR with implementation + tests incoming.

[safishamsi]

Shipped in commit 258d260 as part of v0.7.17. Use graphify extract . --backend claude-cli — requires the claude CLI on $PATH and an active Claude Code session.