Reflective Engineer v2 turns promotion decisions into a replayable gate

What it does

Deciding whether an agent change deserves to ship is usually a conversation. Reflective Engineer makes it a gate you can replay: import an experiment report with parent and candidate metrics, and the same deterministic evaluator decides in the browser, on the command line or through MCP.

The promotion rules are stated, not implied. Quality must improve. Safety must be at least 0.99 and must never decrease. No regressions, no cost increase, no latency increase.

What changed

Commit 3822d73, merged as pull request #4 on September 11, replaces the 2024 React prompt builder with the evidence console: strict versioned JSON with artefact hashes and unique test cases, a mobile-friendly browser console under a restrictive CSP with no network access, the same gate evaluator behind a CLI and MCP, MetaHarness profiles and an Autogenous pinned fitness gate with an explicit manual promotion boundary.

The historical React code stays in src/ for migration reference and is excluded from the supported build. Its browser provider credentials and local storage encryption were never real security boundaries and are no longer presented as such.

Get started

Node 24:

git clone https://github.com/ruvnet/reflective-engineer.git
cd reflective-engineer
npm ci
npm test
npm run build
node console/serve.mjs

Expected result: the console serves on http://127.0.0.1:4173 with a clearly synthetic example loaded. Replace it with your own evidence.

The same gate runs headless:

node console/cli.mjs status
node console/cli.mjs replay < experiment.json
node console/cli.mjs benchmark

{"mcpServers":{"reflective-engineer":{"command":"node","args":["/absolute/path/reflective-engineer/console/cli.mjs","mcp"]}}}

MCP tools are project_status, evidence_replay, project_benchmark and project_test, with policy at ruv://reflective-engineer/policy. Local test execution needs RUV_ALLOW_VALIDATION=1. Requests cannot choose commands, paths or environment variables.

Use it today

Practical case: a weekly review where someone proposes promoting an agent change. Input is the experiment JSON with parent and candidate metrics. Workflow is replay in the CLI for the record, then open the same file in the console so the room can see the numbers. Output is a downloadable review with a stated verdict.

Acceptance test: take a passing report and lower its safety score below 0.99. Replay must refuse promotion, and the CLI and browser must agree.

Push it further

Experimental commentary. Running the replay in CI against every proposed change turns promotion criteria into something a pull request can fail, rather than a judgement made in a meeting.

Limitation: the tool does not execute agents, does not authenticate imported measurements and does not deploy anything. It evaluates the numbers you give it, so a fabricated report passes a real gate. Falsifiable test: submit a report whose artefact hash does not match its content. Strict parsing should reject it before any metric is considered.