docs: add CLAUDE.md for Claude Code guidance

CLAUDE.md

@@ -0,0 +1,104 @@
+# CLAUDE.md
+
+This file provides guidance to Claude Code (claude.ai/code) when working with code in this repository.
+
+## What this is
+
+SkillSpector is a security scanner for AI agent skills (Claude Code, Cursor, etc.). It statically analyzes a
+skill directory/zip/URL for prompt injection, data exfiltration, dangerous code, and other risk patterns, then
+optionally runs an LLM semantic pass, and produces a risk score + SARIF/JSON/Markdown/terminal report. It never
+executes the scanned skill — all analysis is static (regex, Python AST, YARA) plus optional LLM evaluation of
+file contents.
+
+The engine is a **LangGraph** workflow (`from skillspector import graph`). See `docs/DEVELOPMENT.md` for the
+full architecture writeup — the summary below is enough to be productive; read that doc before adding an
+analyzer or provider.
+
+## Commands
+
+All `make` targets assume a venv is already created and activated (Makefile picks `uv` if available, else `pip`).
+
+```bash
+uv venv .venv && source .venv/bin/activate
+make install-dev          # editable install + dev/mcp/langgraph-dev extras
+
+make test-unit             # unit tests only, no LLM calls (what you want most of the time)
+make test-integration       # invokes the full graph, may call configured LLM providers
+make test                   # test-unit + test-integration
+make test-cov               # test-unit with HTML+terminal coverage report
+make test-provider [openai|anthropic|nv_build]   # live provider tests, needs real API keys
+
+pytest tests/unit/test_patterns.py -k some_test   # run a single test
+pytest -m "not integration and not provider" tests/some_dir/  # scope to a subtree, respecting default markers
+
+make lint                   # ruff check src/ tests/
+make lint-fix                # ruff check --fix
+make format                  # ruff check --fix + ruff format
+make format-check            # ruff format --check (no changes)
+
+skillspector scan ./my-skill/ --no-llm     # run the CLI directly against a skill under test
+make langgraph-dev                          # LangGraph Studio dev server, for visually inspecting/running the graph
+```
+
+Tests are marked `integration` (full-graph, may call LLMs) and `provider` (live provider endpoint tests); the
+default `pytest` addopts excludes both, so `make test-unit` == plain `pytest tests/` in practice.
+
+## Architecture
+
+**Data flow:** `resolve_input` (URL/zip/file/dir → local `skill_path`) → `build_context` (reads files into
+`components`/`file_cache`/`ast_cache`/`manifest`) → ~22 analyzer nodes run in parallel (fan-out) → `meta_analyzer`
+(fan-in; per-file LLM filter/enrich of findings when `use_llm`) → `report` (baseline suppression, SARIF build,
+risk scoring, `report_body` formatting) → END.
+
+```
+resolve_input → build_context → [analyzers: static_* / behavioral_* / mcp_* / semantic_*] → meta_analyzer → report
+```
+
+- **State**: `SkillspectorState` (`state.py`, `TypedDict, total=False`) is threaded through every node. Findings
+  accumulate via an `operator.add` reducer on `state["findings"]`. See the field table in
+  `docs/DEVELOPMENT.md` §4 before touching state shape.
+- **Findings**: analyzers emit `AnalyzerFinding` (`Location` + `Severity` enum), converted to the canonical
+  `Finding` (`models.py`) via `static_runner.analyzer_finding_to_finding`. `Finding.to_dict()` is the schema
+  boundary for JSON/SARIF output — treat it as the contract when changing finding shape.
+- **Analyzers** (`nodes/analyzers/`): registered in `ANALYZER_NODE_IDS` / `ANALYZER_NODES`
+  (`nodes/analyzers/__init__.py`); `graph.py` wires `build_context → each analyzer → meta_analyzer` in a loop, so
+  adding a node needs **no** `graph.py` change. Categories: `static_patterns_*` (regex, one `analyze(content,
+  file_path, file_type) -> list[AnalyzerFinding]` per module, built on `static_runner.run_static_patterns` +
+  `pattern_defaults` for category/remediation), `static_yara.py` (YARA), `behavioral_ast.py` (AST1-9: exec/eval/
+  subprocess/os.system/compile/dynamic-import/getattr), `behavioral_taint_tracking.py` (TT1-5: source→sink
+  dataflow over Python AST), `mcp_least_privilege.py` / `mcp_tool_poisoning.py` / `mcp_rug_pull.py`, and
+  `semantic_*.py` (LLM-only; return `{"findings": []}` when `use_llm` is False — this is also the pattern for a
+  not-yet-implemented placeholder analyzer).
+- **LLM plumbing**: `llm_utils.get_chat_model()` / `chat_completion()` dispatch on `SKILLSPECTOR_PROVIDER`.
+  `providers/<name>/` is one subpackage per provider (own `provider.py` + bundled `model_registry.yaml`);
+  `providers/registry.py` exposes context-length/max-output lookups. CLI providers (`claude_cli`, `codex_cli`)
+  implement `AgentCLICapable` and shell out through the hardened `providers/_agent_cli.py` (no shell, stdin-only
+  untrusted content, env scrubbed of API keys, tools/MCP disabled, per-call timeout). `nodes/llm_analyzer_base.py`
+  (`LLMAnalyzerBase`, `LLMMetaAnalyzer`) provides shared per-file/per-chunk token-budget-aware batching used by
+  `meta_analyzer` and the semantic analyzers.
+- **Suppression** (`suppression.py`): baseline findings can be accepted via exact fingerprint or drift-tolerant
+  glob rule; `report.py` partitions them out before scoring. See `docs/SUPPRESSION.md`.
+- **Entry points**: CLI (`cli.py`, Typer app, `skillspector scan`), programmatic (`from skillspector import
+  graph; graph.invoke({...})`), LangGraph Studio (`make langgraph-dev`, graph declared in `langgraph.json`), and
+  an MCP server (`mcp_server.py`, `skillspector mcp`, requires the `mcp` extra) exposing a single `scan_skill`
+  tool for gating installs at runtime.
+
+## Adding an analyzer
+
+1. Implement a node: input `state: SkillspectorState`, output `AnalyzerNodeResponse` (`{"findings":
+   list[Finding]}`).
+2. For a regex-pattern analyzer, write `analyze(content, file_path, file_type) -> list[AnalyzerFinding]` and run
+   it through `static_runner.run_static_patterns`, sourcing category/remediation text from `pattern_defaults`.
+3. Register the node id/callable in `nodes/analyzers/__init__.py`'s `ANALYZER_NODE_IDS` / `ANALYZER_NODES` —
+   `graph.py` wires the edges automatically.
+4. Add unit tests (and fixtures under `tests/fixtures/` if needed); see `tests/nodes/analyzers/` for the existing
+   pattern.
+
+## Conventions
+
+- New source files need the SPDX license header (copy from any existing `.py` file).
+- Commits require DCO sign-off (`git commit -s`).
+- Ruff is the only linter/formatter in CI (line-length 100, `target-version py312`); mypy config exists in
+  `pyproject.toml` but is not currently run in CI.
+- Env config for local runs lives in `.env` (copy `.env.example`); `SKILLSPECTOR_PROVIDER` selects the LLM
+  provider (default `nv_build`), credential var depends on provider (see README's provider table).