Agent integration
Mjölnir is built to sit inside an agent's fix loop — the same loop a human maintainer runs, made mechanical. One rule governs the surface: the agent gets derived facts, never trust claims.
The loop
1. mjolnir --save-baseline # the before-state: .mjolnir/baseline.json
2. (the agent fixes findings)
3. mjolnir ci verify # before/after digest — see below
4. repeat 2–3 until clean, then commitmjolnir ci verify prints the digest and exits with the frozen contract: 0 clean · 1 new error findings · 2 partial scan or no baseline. A partial scan never masquerades as a clean verification — the agent re-runs or reports honestly, exactly like CI does.
The digest
- RESOLVED — baseline findings no longer present, each with its §15 lifecycle resolution. Only
VERIFIED-RESOLVEDis a fix claim;INCONCLUSIVE(cause)states what the scan could not prove. - NEW — findings the change introduced, with severity. Any new error fails the verb (exit 1).
- UNCHANGED — pre-existing debt, grouped by
ruleId + location. The agent's working key is the location: a finding whose message reworded but whose position persists is unchanged debt, not progress. - SCORE DELTA — before → after, with the direction named. The score is a measurement, not a reward signal: it moves because findings moved, and the digest shows which ones.
MCP
Connect from Claude Code
Run this command to register the local Mjölnir MCP server:
claude mcp add mjolnir -- npx -y mjolnir-qa@6.1.0 mcpUse Node.js 22.18 or later. For another MCP client, configure npx as the command with arguments -y mjolnir-qa@6.1.0 mcp. Provide the repository path when requesting a scan.
Ask the agent to inspect findings, explain their evidence, and compare the scan before and after a change. Run the affected tests separately. Mjölnir does not execute the test suite or automatically apply suggested fixes.
For a concrete result shape, open the JSON tab in the saved example report. For machine-readable discovery, start with llms.txt.
Verification contract
The same loop runs over the MCP transport: the verify tool is 1:1 with the verb (scan + baseline diff → the digest as data), under the same guardrails as every tool — one scan in flight, parameter size caps, filesystem boundary = scan target + .mjolnir/, zero network, plugin gate unchanged. mjolnir install writes the agent instruction surfaces (Claude Code command, agent briefs) with the loop documented inline.
Boundaries (what the agent loop must never do)
- Suppress a finding to obtain a clean digest — suppressions require a reason, live in
mjolnir.config.json, and are bounded only by an explicit ISOexpiresdate. - Read trust into the digest: a
VERIFIED-RESOLVEDline is a §15 lifecycle resolution, not a proof of correctness. - Treat the score as a reward signal to maximize — it is a measurement with a published deduction table, and gaming it is the failure mode the formula was rebuilt to close (docs/SCORING.md v2).