Add hackenproof-audit-scope skill (Rust + Solidity audit scoping)

[obs441]

Summary

Addresses #7 — a pre-sales scoping skill for Sales Managers. This first version prices Rust and Solidity only; the web/mobile (API-documentation-based) part of the issue is deferred to a follow-up.

Given a repo URL or local folder (optionally a branch/commit and the client's scope folders), the skill:

• Counts billable nSLOC with a comment- and string-aware tokenizer. Tests, mocks, scripts, interfaces, examples, generated code and vendored dependencies are listed but excluded. Rust #[cfg(test)] modules inside source files are removed from the source count (30–60 % of a file in typical CosmWasm/Anchor repos).
• Ranks the most critical files from weighted security indicators (assembly, delegatecall, oracles, cross-chain, CPI, unsafe, precision math, …) evaluated on comment-stripped code, with plain-English reasons.
• Suggests a complexity tier and produces an effort estimate: auditor-days by phase (core review, reporting, retest), a range, calendar weeks for a given team size, and a sensitivity table across all three tiers. Rates are documented in references/effort-model.md and overridable via --rates; they are starting values for the audit team to calibrate.
• Screens the repository before anyone opens it: prompt-injection text in comments/docs/agent-config files, hidden Unicode (Trojan Source, homoglyphs, zero-width), and code-execution hazards (build.rs, .cargo/config overrides, proc-macro crates, ffi = true, npm lifecycle scripts, VS Code folder-open tasks, .envrc, dev-container hooks, .mcp.json, in-repo CLAUDE.md). Any HIGH finding stops the workflow with an escalation note.
• Writes a one-page scoping summary from a template that always carries "heuristic — confirm with the audit lead before quoting".

The bundled script (scripts/audit_scope.py, stdlib only, Python 3.8+) is strictly read-only: it never builds, installs, tests or executes anything from the target repo. The skill instructs the agent never to cd into the clone or load its agent-instruction files.

Verification

• Per-file counts compared with cloc 2.06: identical on OpenZeppelin Contracts (420 files) and CosmWasm cw-plus (87 files). The 5 differing files in coral-xyz/anchor are cloc truncating strings at // inside URLs, i.e. cloc undercounts there.
• Tier calibration on seven public repos: OpenZeppelin, solmate, Uniswap v2-core, Raydium AMM → high; anchor, cw-plus → medium.
• Zero HIGH false positives on those repos after avoiding domain vocabulary (Solidity override, Solana "instructions", "send tokens").
• Planted-injection fixture fires every expected HIGH exactly once without affecting line counts.
• Clone by branch/tag/commit SHA and client --include filters exercised end to end.

Files

plugins/hackenproof-audit-scope/
  .claude-plugin/plugin.json
  skills/hackenproof-audit-scope/
    SKILL.md
    agents/openai.yaml
    scripts/audit_scope.py
    references/counting-rules.md
    references/critical-code-heuristics.md
    references/effort-model.md
    references/injection-screening.md
    references/scope-report-template.md
.claude-plugin/marketplace.json   (new entry)
README.md                         (new section)

Follow-ups

• Calibrate effort-model.md rates against real engagements.
• Web/mobile scoping from API documentation (remaining bullet of Audit scope calculation #7).

🤖 Generated with Claude Code