Agent conversations#

Agent conversations combine a provider session with a project root, model, permission posture, optional memory and an explicit set of MCP tools.

Before you start#

• A provider reports ready in Settings.
• The project root is correct and its current changes are understood.
• You have chosen whether the turn may write files or run commands.

Start safely#

1. Create a conversation from the workspace you intend to use.
2. Confirm the provider, model, local or SSH target and project path.
3. Begin read-only for unfamiliar repositories.
4. Review each permission request and the complete command, not only its label.
5. Inspect the diff before accepting generated edits.

Memory and MCP access are separate controls. A conversation with memory enabled can receive project context; a conversation with MCP enabled can call the selected server tools. Disable either when it is unnecessary.

Continue and recover#

Conversation transcripts remain readable after a provider is removed. A provider change may require a new conversation when the original runtime cannot resume its session format. If streaming stops, cancel the turn before retrying to avoid overlapping work.

Verify#

After the turn completes:

git status --short
git diff --check
git diff

Confirm that only intended files changed, then run the affected project's own tests. PacketBench supervises the interaction; it does not replace repository review or an operating-system sandbox.

If it does not work#