feat(classifier): subscription-native agent-sdk backend via Haiku (Track C)

[Shahinyanm]

What

Gives subscription users without an ANTHROPIC_API_KEY a real LLM classifier. Previously, uncertain chunks just queued in pending/ forever for them. This adds an agent-sdk backend that classifies through the local, already-logged-in claude binary (pinned to Haiku) — riding the Claude subscription instead of an API key.

This resurrects the v0.7.x cli path that was dropped in v0.8.0, documented honestly this time.

Changes

• ClaudeCliClassifier (crates/tj-core/src/classifier/agent_sdk.rs) runs claude -p <prompt> --model claude-haiku-4-5 --output-format json --strict-mcp-config, parses the JSON envelope's result, and reuses a shared verdict parser. Command execution is injected via a CommandRunner trait, so the parse path is unit-tested without ever shelling out. from_env() returns None unless claude is on PATH; model overridable via TJ_AGENT_SDK_MODEL.
• Shared classifier::parse_verdict — factored http.rs's fence-strip + serde parse into one helper used by both http and agent_sdk so they can't diverge.
• Hybrid is now an ordered LLM chain: heuristic (≥ 0.7) → agent-sdk (if claude on PATH) → api (if key) → pending/. Reorder with TJ_HYBRID_LLM_ORDER (default agent-sdk,api).
• Wiring: agent-sdk added to --backend for ingest-hook and classify-worker; install-hooks gains a --backend flag that bakes the choice into the generated hook command.
• Honest docs: README "Classifier backends" + env table, --backend help, and doctor all note that since 2026-06-15 a headless claude -p draws from the separate Agent SDK monthly credit pool (~$20 Pro / $100 Max 5x / $200 Max 20x at API rates), not the interactive pool. Classification is Haiku-class and tiny, so it lasts — but it's not strictly free.

Verification

• cargo fmt --all --check clean; cargo clippy --workspace --all-targets clean
• cargo test --workspace green; cargo test classifier:: 32/32 pass:
  • agent_sdk: canned-verdict parse, model pin, decide_status at the 0.85 threshold, code-fence tolerance, is_error → Err
  • hybrid: uncertain heuristic + agent-sdk available → uses agent-sdk and never touches http
  • http: both existing tests still pass after the parse_verdict refactor
• Live: claude -p ... --output-format json returns valid {is_error:false,result:...} on a subscription with no ANTHROPIC_API_KEY (EXIT=0) — confirms the flags before hardcoding.

Flags were verified live against the installed claude (not assumed from memory). No real claude invocation in CI — the CommandRunner is faked.

Stacked on Track A (#23) conceptually but branched from main; merge A first.

🤖 Generated with Claude Code