Working with your agent

You can ask your agent to record a decision, look up an earlier choice, or list the questions that still need answers. Docket keeps those records available across sessions.

Connect your agent

Use the setup prompt to let your agent install Docket and configure its integration. If you prefer manual setup, the installation guide has a route for Claude Code, Codex, Gemini CLI, GitHub Copilot CLI, Cursor, and OpenCode.

After setup, start a new session in your project. Ask:

Read the Docket context for this project and tell me which questions are still open.

The integration loads a briefing and gives the agent instructions for using Docket. If you prefer to configure it yourself, use the agent setup reference.

Different agent tools can use the same project ledger when they run in the same checkout. To share records with another checkout or computer, commit and sync .docket/ through Git.

Ask for a record when a choice is made

You can say:

Record that billing will use Postgres because we already operate it. Scope the decision to billing/**.

Give the agent the choice and the reason you want preserved. It can then record them with the Docket command.

For an unresolved issue:

Leave an open question about which database driver billing should use.

For a change of direction:

Replace the earlier driver decision with asyncpg and record why we changed it. Keep it linked to the question it answers.

The agent should use --supersedes to preserve the earlier record. You can inspect what it recorded with docket list and docket show ID.

Start work with the relevant context

Before a new task, ask:

Read the Docket context for billing and check the open questions before making changes.

For a more specific briefing, you or your agent can run:

docket context --query "billing database" --file billing/db.py

The briefing distinguishes a recorded choice from a claim that may need review. An accepted claim still needs fresh checking when its evidence could have changed.

When a feature is open on the current branch, the briefing names it above the record selection: its slug, state, and declared intent, then its highest-ranked attached records. The header draws from the same character budget as the rest of the briefing, so it takes a bounded share rather than a fixed count. See Feature tracking.

After compaction or a long break, ask the agent to read the context again. The integration's refresh behavior depends on the agent tool, and a briefing already loaded into a conversation does not update itself.

When a decision is missing

First, check that the record exists in the active ledger:

docket where
docket list --find billing

Then request a focused briefing:

docket context --query "billing" --file billing/db.py

A record may appear in the index rather than in full. Ask the agent to retrieve it by ID. If it does not appear at all, check the ledger location and whether the record's wording or scope matches the task.

To change a scope, record a replacement with --supersedes. To inspect a record immediately, use docket show ID --json without changing the ledger.

If the command works in your terminal but the agent receives no briefing, start a new session and check its integration setup. If the command itself fails, run docket check and see Maintenance.

Know who recorded a choice

Docket records an author and includes session and branch information when available. These describe who recorded the entry; they do not establish who approved the choice.

Use --decided-by when the decision owner is someone else. This is recorded attribution, not an authenticated identity.

For the hook formats and instruction files, see Agent setup reference. For how a briefing is selected, see Bounded context.

Recording and feature skills

skills/docket/SKILL.md covers recording: claims, decisions, and questions. skills/docket-feature/SKILL.md covers a piece of work in flight: starting it, resuming it with docket feature brief, and closing it with docket feature done. The harness selects between them on each skill's description, so a task that only records a decision never loads feature guidance, and a task that starts or resumes a feature never loads the recording skill unless it also records something.