Module Index

One doc per component, each with a plain paragraph, a diagram, a technical half, and inputs/outputs/model-selection.

Module Index

One document per component. Each has a plain-English half (with a diagram) and a technical half (with a file/responsibility table). Suggested reading order top-to-bottom follows the closed loop, then the runtimes.

Component Kind What it is Prod files Doc
vulnhunt/ Prompt-only skill The scanner — /vulnhunt 13 md (SKILL + 11 phases) vulnhunt.md
vulnhunter-fix/ Prompt + Python The fixer — /vulnhunter-fix (TDD → PR) 39 py · 21 prompts/refs vulnhunter-fix.md
vulnhunt-fix-verify/ Prompt-only skill The verifier — /vulnhunt-fix-verify (read-only) 7 md (SKILL + 4 phases) vulnhunt-fix-verify.md
vulnhunter-agent/ Python runtime Headless: clone→scan→publish→issues 35 py (~13.2k LOC) vulnhunter-agent.md
harness/ Python (dev-only) Batch scan + accuracy benchmark 13 py (~2.4k LOC) harness.md

Each component doc ends with a Part 3 — Inputs, outputs, model selection & deterministic calls section. At a glance:

Component Inputs → Outputs How the model version is chosen Deterministic (non-LLM) calls
vulnhunt/ code checkout → results dir (report, PoCs, tests) No routing — inherits the session model; SKILL.md only checks it is Opus 4.7+ Grep/Glob/Read; (bash mode) dep install + run exploit tests; /cost. No schema/graph engine.
vulnhunter-fix/ labelled issues / results dir → a PR (A) Step-0/0b self-check: stops if session isn't Opus 4.x (interactive only; skipped headless). (B) Per-finding routing (deterministic): opus/sonnet/haiku by severity + CWE + multi-file; default sonnet. Completeness-tier classifier escalates to a bounded LLM only when no signal matches. git worktrees/commits, gh PR, pytest, Graphify AST graph, jsonschema, 6 delivery gates, language/mode detect.
vulnhunt-fix-verify/ repo + report + fixed= → verdict JSON No routing — session model (Opus recommended) Almost none by design (no Bash/network): only jsonschema at emit + file-existence sync.
vulnhunter-agent/ repo URL + TOML/env → manifest + issues Config value [anthropic].model (default claude-opus-4-8, --model/env override); issues stage tiers haiku→sonnet config load, OAuth/SigV4 auth, git clone, GitHub API, jsonschema, retry/backoff, sandbox, audit.
harness/ REPO_LIST / ground truth → results + metrics Single MODEL constant for scan and judge git clone, subprocess claude -p /vulnhunt, thread pool, JSON parse, tally, state persistence; judge is the one LLM call.

The pattern: the two prompt-only skills (vulnhunt, vulnhunt-fix-verify) make no model choice — they inherit the session model and reason with search tools. The three Python components add the deterministic machinery: the fixer routes the model per finding and wraps LLM work in schema-validated gates; the runtime and harness pick the model from config and do all the git/GitHub/validation plumbing.

Repo totals: 177 Python files / ~50k LOC + 71 Markdown files. Prod-file counts above exclude test suites (VulnHunter is heavily tested — 90 test files across the two runtimes and the fixer). Per-file, per-function detail is in tool-outputs/codeviz/vulnhunter-codeviz.html and the skill explorer.