Skip to content

Verify your first decision in a new coding session

The first useful milestone is a decision you can retrieve after the conversation that produced it has ended. This check works with the six supported setup targets. Start with the quickstart if Selvedge is not connected yet.

Run from your project directory:

Terminal window
selvedge --version
selvedge doctor

Confirm the reported database path belongs to this project. Review any configuration or executable-path warning before continuing. The agent should have access to Selvedge’s eight MCP tools. A healthy database alone does not prove that the agent has connected.

Use an actual choice from your project, with a specific entity such as src/cache.py::load_profile. Ask your agent:

Use Selvedge to record one real approach we considered and rejected in this project. Include the entity, why we rejected it, and what would change our mind. Do not invent a decision; ask me if none is available. Then show the saved entity path and record ID.

Inspect the log_change response. A successful response reports status: "logged" and an ID; inspect any warnings about the explanation. If you only want a synthetic demonstration, run selvedge demo instead. Keep demonstration records out of real project history.

3. Retrieve it without the old conversation

Section titled “3. Retrieve it without the old conversation”

Open a new session in the same project. Do not paste the old rationale into it. Replace ENTITY with the path saved above:

Use Selvedge prior_attempts to retrieve the decision for ENTITY. Show the recorded reason and outcome. Assess whether the original constraint still applies before proposing any edit.

Confirm that the tool actually ran and returned the correct record. Compare its reason and outcome with what you saved. A fluent answer without a tool result does not verify later-session retrieval.

On the next relevant coding task, inspect whether the agent consulted the decision before editing, considered current evidence, and tested its chosen approach. If the constraint has changed, reopen the intended decision explicitly.

Terminal window
selvedge stats --json

This reports observed tool calls and explanation-quality signals. Its logging ratio is not the percentage of all repository edits captured. Record retrieval and successful application as separate observations.

What you observeWhat to check next
No Selvedge toolsRestart the client, enable MCP tools, and check the server executable path in its setup guide
log_change returns an errorCorrect the reported parameter error and retry; do not treat an attempted call as a saved record
Fresh session returns no recordCompare database paths and exact entity paths; inspect filters and full history
Terminal sees the record but the agent does notCheck the agent’s working directory and MCP environment against selvedge doctor
Agent repeats the rejected approachInspect tool use and applicability; connecting memory is not a guarantee of compliance

An empty result means no qualifying record was returned under that query. It does not prove the approach was never tried.

To evaluate a second client, follow share decision memory between agents. To report your experience voluntarily, use the existing feedback pilot; read its consent details and redact private code and decisions.