Evaluate Mjolnir in ten minutes

Fixture and CLI surface reviewed against Mjolnir 1.0.2 on 2026-07-21. Live provider output is model- and availability-dependent and is not run in docs CI.

This evaluation uses a checked-in Python fixture in a disposable Git repository. It proves that a configured Thor session can inspect a small project, request an Eitri implementation handoff, present role-labelled work, run an explicit review, preserve a resumable session, and emit headless stream records.

It does not prove large-repository performance, consistent quality across models, predictable provider cost, safe unattended shell execution, spontaneous Loki timing, or remote-server security.

Before you start

You need:

• Mjolnir installed and mj --version working.
• Python 3 for the fixture test.
• Git.
• At least one authenticated, launchable provider route. Provider use may cost money.

Run mj, open /mjconfig, and confirm that Thor and Eitri resolve to available models. Loki is optional. Model or ACP-server changes apply to the next session, so exit and relaunch after changing them.

Prepare the disposable fixture

From a Mjolnir checkout:

EVAL_DIR="$(mktemp -d)/mjolnir-eval"
cp -R docs/fixtures/ten-minute-evaluation "$EVAL_DIR"
cd "$EVAL_DIR"
git init
git add .
git -c user.name="Mjolnir Eval" \
    -c user.email="mjolnir-eval@example.invalid" \
    commit -m "evaluation fixture"
python3 -m unittest -v

The baseline has two passing tests. The requested edge cases are intentionally absent.

If you are reading the published docs, clone the repository first:

git clone https://github.com/BrokkAi/mjolnir.git
cd mjolnir

Then run the preparation block above.

Journey 1: Council implementation

Start Mjolnir in the fixture:

mj --cwd "$EVAL_DIR"

Send this prompt:

Use your implementation agent, Eitri, for this bounded change. Update weather.py
so status(0) returns "freezing", negative values return "below freezing", values
below 20 return "cold", and all other values return "warm". Add focused tests,
run python3 -m unittest -v, and explain the result. Do not change anything else.

Expected observations:

1. Thor owns the user turn and presents the plan or handoff.
2. An Eitri-labelled implementation run appears in the normal transcript.
3. Any requested permission remains fully readable before you decide.
4. The returned change is limited to weather.py and test_weather.py.
5. python3 -m unittest -v reports four passing tests.

The exact wording and tool sequence can differ by model. If Thor ignores the explicit handoff or the result is wrong, record the selected models and adapter in your evaluation notes; that is a failed outcome, not a docs failure.

Journey 2: explicit review

Run:

/review recent

Choose the most recent change-producing turn. Review is findings-only; a clean result is valid. Loki’s background advice is best-effort and may arrive before or after this step, so spontaneous advice is not a pass condition.

Exit with Ctrl-D on an empty prompt. Mjolnir prints a command shaped like:

mj resume <session-id>

Run the printed command. The session should return through its saved ACP adapter and model provenance. When a worktree was used, pass its printed --worktree <name> value to reuse that directory.

Journey 3: headless read-only output

From another terminal:

mj --cwd "$EVAL_DIR" \
  --print \
  --permission-mode manual \
  --output-format stream-json \
  "Inspect weather.py and summarize its behavior. Do not modify files." \
  | tee /tmp/mjolnir-eval.ndjson

The output is newline-delimited JSON. It should contain connection/session records, role-labelled message or thought records, and a final result record. manual rejects any permission request rather than hanging an unattended run.

Interpret the result

A successful run proves the selected provider route can support the core Council path on one small repository. Compare providers by repeating the same fixture and recording model IDs, elapsed time, token/cost telemetry, whether the handoff occurred, test outcome, review outcome, and any manual intervention.

Before broader use, read Permissions and workspace scope and Data and trust boundaries.