DM / Code Agent

Local-first. Open source. Inspectable.

See what your
coding agent did.

Trace code changes, diagnose failed runs, and check the test evidence. A Python coding agent with a small, readable core.

Explore the demo View the source

No sign-up. No API key. MIT licensed. · Watch the 38s walkthrough ↗

ONE BUG. TWO ENDINGS.Runnable fixture
01
Find the failing boundary

Four local checks. One failing assertion.

Fail
02
Inspect the actual edit

A content-anchored replacement in retry.py.

Edit
03
Check the verification

Compare a tested repair with an untested claim.

Audit
scripted clientReal files · real tools · real tests

Try the evidence

Two runs. One important difference.

Both runs claim completion.
Only one runs the tests.

Open full report

The agent runs the failing checks, reads and edits the file, then reruns the checks successfully.

How this demo works. The responses are scripted, not live model output. The files, edits and unittest results were produced by the real agent tools. This is a demonstration of behavior, not a model benchmark.

View provenance ↗

Built for inspection

A run you can work with.

01 / TRACE

Follow every step.

Keep plans, tool calls and results in an append-only session log. Context compaction preserves the original records.

02 / DIAGNOSE

Spot missing evidence.

Inspect failures, recovery and verification gaps. Compare two runs without reducing the story to a single success flag.

03 / SHARE

Send a readable report.

Export one standalone HTML file. Start with a safe summary; opt in to redacted content, review it locally, then share.

Start locally

Your first run.
Zero model calls.

Install the verified release wheel, then run the demo in a new directory. Python 3.10+ is required. The demo needs neither an API key nor pytest.

Full installation guide ↗
Terminal 
python -m pip install "https://github.com/hwfengcs/DM-Code-Agent/releases/download/v2.1.0/dm_code_agent-2.1.0-py3-none-any.whl"
dm-agent-demo --output demo-run
dm-agent-trace export demo-run/sessions/verified.jsonl --output repair.html --include-content

Open repair.html in your browser. The demo writes only inside its new directory; existing contents are never overwritten.

Field notes

Small failures. Useful lessons.

Read the series ↗
01 · Budget

When more steps cost more without a reliable gain.

What the recorded experiments actually support.

 02 · Observation

The word “error” isn't always an error.

Why tool provenance matters to recovery decisions.

 03 · Editing

Find the text before changing the code.

Content anchoring and the cost of stale line numbers.
A real maintenance case.

Twelve targeted checks passed, but the broader suite still failed. Inspect the live-model run ↗ or read the case study and offline reproduction ↗.

Know the boundaries.

Local-first does not mean sandboxed. Live tasks can modify files and run commands with your permissions. Review shared reports: automatic redaction cannot identify every secret. Historical benchmark results and their sample limits are documented in the repository.