The evidence store for coding agents.

Run it locally in five commands.

$ curl -fsSL https://sediment.so/install.sh | sh
$ sediment server
$ sediment login http://127.0.0.1:8000
$ sediment install .    # configures git hooks
$ sediment doctor .

Quickstart · Deploy for a team

Capture five signals, depending on the agent.

inference call

Capture each prompt, response, and model from your gateway.

developer decision

Record every accept and reject, human or automatic.

edit

Measure how much of each edit survives the session.

commit and push

Link commits to sessions through git notes or similarity.

ci outcome

Record check results across runs and retries.

See the architecture diagram

1. Which model works better?

   Compare each model's CI pass rate and how often its calls reach a commit.

   $ sediment report model

2. Did accepted work reach a commit?

   Follow the work you accept through commits, merged pull requests, and CI.

   $ sediment report lifecycle --json

3. How much code survived?

   Measure how much agent code remains after review and merge. SWE-chat (October 1, 2026) found that 59% of agent-produced code survived into user commits, in sessions collected from public repositories.

   $ sediment report merge-retention

4. What needs rework?

   Count rejects, retries, changed or deleted edits, outside changes, and CI failures separately.

   $ sediment report lifecycle --json

5. What went into a failed CI run?

   Trace the failure to its commit, sessions, and the model calls that likely contributed.

   $ sediment commit COMMIT_SHA

How to measure agent work

Export SFT, DPO, and RLVR datasets. Trace every label back to its evidence.

See training exports

How Sediment turns coding-agent activity into training data · How CI outcomes become RLVR training data

Check what Sediment records from each agent.

See agent integrations