feat: add Claude Code investigation skill

[tdimino]

Why a Claude Code Skill for OpenPlanter?

OpenPlanter is a recursive investigation agent with its own runtime. Claude Code is a general-purpose AI coding assistant with native file, web, and shell tools. This skill bridges them:

• Methodology portability: OpenPlanter's epistemic discipline (Admiralty System, ACH, entity resolution protocol) works inside Claude Code without installing the full agent
• Tool parity: Every OpenPlanter tool maps to a Claude Code equivalent (see tool mapping table in SKILL.md) — the methodology is what matters, not the runtime
• Delegation bridge: When investigations exceed Claude Code's comfort zone (20+ reasoning steps, 3+ datasets), delegate_to_rlm.py spawns the full OpenPlanter agent as a subprocess with session resume support
• 12 government API fetchers: SEC EDGAR, FEC, Senate LDA, USAspending, OFAC SDN, OpenSanctions + 6 specialized fetchers (Census ACS, EPA ECHO, ICIJ Offshore Leaks, OSHA, ProPublica 990, SAM.gov)
• Knowledge graph query: Read-only access to OpenPlanter wiki knowledge graphs built during delegated investigations

This is not a port. It's a bridge between two complementary investigation paradigms.

---

Summary

• 16 scripts (10 core + 6 specialized fetchers): all Python stdlib-only, zero external dependencies
• Core scripts: dataset_fetcher.py (bulk datasets), web_enrich.py (Exa neural search), scrape_records.py (SEC/FEC/LDA/USAspending), investigate.py (pipeline orchestrator), delegate_to_rlm.py (provider-agnostic RLM bridge with session management), entity_resolver.py, cross_reference.py, confidence_scorer.py, evidence_chain.py, init_workspace.py
• Specialized fetchers: fetch_census.py, fetch_epa.py, fetch_icij.py, fetch_osha.py, fetch_propublica990.py, fetch_sam.py
• Wiki graph: wiki_graph_query.py (read-only query of OpenPlanter knowledge graphs — entity lookup, BFS pathfinding, neighbor traversal, stats)
• 2 agent definitions: investigation-agent.md (decision tree: scripts vs. RLM delegation), verifier-agent.md (read-only uncorrelated verification)
• 1 reference doc: public-records-apis.md (12 API endpoints with auth, rate limits, linking keys)
• Updated docs: SKILL.md, README.md, CLAUDE.md

Three Integration Modes

Mode	When	Scripts
Methodology Only	1-2 datasets, local analysis	Core analysis scripts
Web-Enriched	Need external data, public records	+ dataset_fetcher.py, web_enrich.py, scrape_records.py, 6 specialized fetchers
Full RLM Delegation	Complex investigations, 3+ datasets	delegate_to_rlm.py → full OpenPlanter agent (session-resumable)

Provider-Agnostic RLM Delegation

delegate_to_rlm.py auto-infers provider from model name:

• claude-* → anthropic, gpt-* → openai, org/model → openrouter
• llama/qwen/mistral/gemma/phi/deepseek → ollama (local inference)
• For ambiguous models, pass --provider explicitly

Session management: --resume SESSION_ID continues a previous investigation, --list-sessions shows saved sessions, --list-models shows available models per provider. --reasoning-effort low|medium|high controls investigation depth.

6 Specialized Government API Fetchers

Script	Source	Auth	Key Linking Fields
fetch_census.py	US Census ACS 5-Year	Optional CENSUS_API_KEY	State/county/ZIP FIPS, population, income
fetch_epa.py	EPA ECHO	None	FRS Registry ID, facility name, lat/lon, SIC/NAICS
fetch_icij.py	ICIJ Offshore Leaks	None	ICIJ node ID, jurisdiction, source dataset
fetch_osha.py	OSHA DOL Enforcement	None	Activity number, establishment name, SIC/NAICS
fetch_propublica990.py	ProPublica Nonprofit Explorer	None	EIN, org name, NTEE code, officer compensation
fetch_sam.py	SAM.gov Entity Management	SAM_GOV_API_KEY	UEI, CAGE code, legal name, NAICS

All fetchers follow the same pattern: argparse CLI → urllib.request → datasets/scraped/{source}/ → provenance.json sidecar with SHA-256.

Wiki Knowledge Graph Query

wiki_graph_query.py provides read-only access to OpenPlanter wiki graphs (NetworkX node-link JSON):

• Entity lookup (case-insensitive substring)
• Neighbor traversal (inbound + outbound edges)
• BFS shortest path between entities
• Graph statistics (node/link counts, type distributions, top connected nodes)
• Full-text search across all fields

Review Fixes Applied (from original PR)

• scrape phase added to investigate.py collect group
• FEC API key stripped from error messages
• Pagination truncation warning in dataset_fetcher.py
• Cerebras examples use --provider cerebras explicitly
• investigate.py examples include required workspace positional arg
• Verifier agent docs note evidence_chain.py write side-effect
• web_enrich.py exits on parse error instead of returning misleading message

Vision for Next PR

Agent-side integration: modify agent/tools.py and tool_defs.py to expose skill scripts as native OpenPlanter agent tools, enabling the RLM agent to invoke the same entity resolution and evidence chain scripts programmatically.

Test plan

• All 17 scripts pass --help and py_compile
• Multi-agent code review — all issues resolved
• Manual: investigate.py /tmp/test --phases all --dry-run
• Manual: scrape_records.py /tmp/test --entities "Apple" --sources sec --dry-run
• Manual: delegate_to_rlm.py --check
• Manual: fetch_sam.py /tmp/test --list (verify NAICS codes)
• Manual: wiki_graph_query.py /tmp/test --stats

🤖 Generated with Claude Code

[tdimino]

Hey @ShinMegamiBoson — just pushed a major expansion to this PR:

• 6 new government API fetcher scripts: Census ACS, EPA ECHO, ICIJ Offshore Leaks, OSHA, ProPublica 990, SAM.gov — all stdlib-only, matching the upstream fetcher patterns
• wiki_graph_query.py: Read-only query tool for OpenPlanter wiki knowledge graphs (entity lookup, BFS pathfinding, neighbor traversal, stats)
• delegate_to_rlm.py upgraded: --resume session management, --list-sessions, --list-models, --reasoning-effort, Ollama provider auto-detection
• SKILL.md + public-records-apis.md: Expanded to cover all 12 API sources, 24 tool mappings

We've been using this skill to power World War Watcher, a real-time dashboard tracking the Iran-US war's impact on infrastructure and economics — entity resolution across defense contractors, government contract data (SAM.gov), and OSINT source chains all run through OpenPlanter methodology: https://x.com/IdaeanDaktyl/status/2031455436591747519

Would love your thoughts on the direction — especially the fetcher pattern and whether you'd want to expose these as native OpenPlanter tools in a future PR.