Your first session

Watch the loop close: hit a real problem, see errata capture it, and start the next session already knowing the answer.

You've installed the CLI, signed in, and wired hooks. This page verifies the loop actually closes — capture, then recall.

1. Work until something breaks

Use your agent normally on a real task. The first useful signal is a failure: a build error, a failing test, a diagnostic. You don't need to do anything — the daemon's sensors see the failing command through your agent's hooks.

2. Fix it

When a later edit makes the failing command pass, errata correlates the edit to the failure and records a problem–fix pair in your local graph. Again: no action needed.

3. See the capture

errata report

opens a self-contained HTML snapshot (.errata/report.html): what was learned, project hotspots, and the causal spine. For a single node's neighborhood straight in the terminal:

errata burst "<a few words from the error>"

burst lands near the problem and prints the surrounding graph — the fix, related problems, provenance.

4. Recall in a fresh session

Start a new agent session in the same workspace. The context block errata injects at session start now carries the problem you just solved — hit anything similar and your agent starts from the answer instead of re-deriving it.

You can also query explicitly from the agent side: the errata.burst MCP tool, or reading .errata/g/burst/<seed> as a file.

5. (When you're ready) join the network

Everything so far happened locally. To let validated, scrubbed lessons compound into your team's scope or the collective graph:

errata consent sync on
errata sync now
What actually crosses the wire

Canonicalized problem signatures and scrubbed descriptions — not your code, not raw error text, not file paths. Run errata privacy anytime to see the collection posture, and errata consent sync off to stop. Knowledge bound for the public graph is additionally generalized; organization-specific detail stays in your scope.

Where to go from here