Catches when AI-agent branches break each other's contracts before merge. 13 languages, AST-level, runs on your CI runners. The engine goes to the code; the code never goes to the engine unless you send it.
What's new since 1.7.0
- Dart is the 13th language - L1 signature analysis (free functions and class methods; named, optional-positional, and required parameters).
- Reachability - every UNSAFE finding now says whether the broken contract is actually reachable from an entrypoint or export (or test-only / unreachable / unknown), with the shortest path.
unreachablerequires closed-world proof; uncertainty degrades to unknown, never suppresses. - Enforced merge gate - policy-driven exit codes in CI. The gate fetches your effective policy at scan time; advisory fails open, blocking fails closed, never silent.
- Agent-facing next steps - when the gate blocks, the output and the PR comment tell a coding agent exactly how to fetch the SHA-keyed verdict JSON and how to load the Rosentic MCP tools, so an agent can act on a block without a human.
- Tamper-evident audit trail - hash-chained, append-only ledger of every gate verdict, replayable, JSON/CSV export.
verify-chainrecomputes from genesis. - Govern - versioned per-repo/org policy (gate mode, severity, per-layer enforce/advisory); overrides require who + reason and land in the ledger.
- Agent attribution - which agent authored each branch, multi-signal (branch + commit identity + trailers +
.rosentic/agents.yml), with confidence tiers. - MCP server - six in-loop tools (check before the write, check before the push) for Cursor / Windsurf / Claude Code / Codex and more. Offline-capable.
pip install rosentic-mcp. - Anonymized telemetry contract - branch names and commit SHAs are hashed on your runner; raw names/SHAs never leave in no-key mode.
- Cross-repo drift (experimental) - detects the same contract drifting across your repos (proto lane). Standalone; not yet in the default scan.
No workflow changes required. @v1 users get everything automatically.