Changelog

# August 2026 — Product ## From v0 to v1.0.0: a KISS security auditor, agent-first and CI-native machin-secure shipped its first full release this month — a security auditor where the tool owns **no model** and the calling agent *is* the LLM. The whole journey from "15 lines of bash beat a 3,000-line search engine" to a reusable GitHub Action with SARIF output, in 8 commits. ### What landed - **The core idea**: a deterministic regex rule engine over the filesystem, with stable finding IDs and a persisted verdict store. No LLM client inside the tool — the agent driving `secure` reads the JSONL, reasons with its own model, and writes back a verdict that survives across scans. Same BYOK split as grepapi's `/v1/brief`. - **Agent-managed verdicts**: every finding carries a `sha256(rule|file|line)` ID. The calling agent triages false positives with `secure verdict drop --reason "..."` (or batches via `--stdin`), and future scans suppress them automatically. The tool never calls an LLM to decide this itself. - **Performance for two different situations**: `--diff` / `--diff-base` scans only git-changed files — 0.23s on a 10k-file repo with 8 changed files (the actual fix for CI latency). `--workers N` parallelizes full audits — 1m28s with 8 workers on the same repo, down from ~4m37s single-threaded. Same findings every way. - **SARIF 2.1.0 output + reusable GitHub Action**: `--sarif` emits a schema-valid SARIF report (CWE helpUris, severity levels, GitHub security-severity 0–10) that drops into GitHub Code Scanning. The Docker action pins machin to an immutable commit, builds the binary, and writes the SARIF file — exit 0 on both clean and findings so `upload-sarif` always runs. - **84 CWE-tagged rules across 16 languages**: secrets, command injection, deserialization, XSS, SSRF, XXE, weak crypto, TLS bypass, SQL injection, path traversal, CORS, open redirect, buffer overflow — across Python, JS/TS/Vue, Go, Rust, C/C++, C#, Java/Kotlin, Ruby, Shell, Swift, PHP, YAML, Dockerfiles, XML. The rule pack is the actual product; the engine is interchangeable. - **Hart reporting (guidance, not generation)**: `--hart` prints a hint pointing the agent at the default hart instance and the publish contract. The tool deliberately does not author or publish the report itself — that would bake prose-generation judgment into a binary that owns none. ### Verified against Synthetic fixtures across 11 languages (all 84 rules fire); a 10.4k-file production Node.js monorepo (1,950 findings including a real hardcoded GitHub PAT); a 36k-file Vue/TS monorepo; and a full non-interactive `devin -p` smoke test that scanned, authored an HTML report, and published it to hart.intrane.fr end-to-end with zero manual steps.