Session Bootstrap

Session Bootstrap

Purpose

Rebuild just enough environment to act. Orientation is a prelude, not the deliverable.

Workflow

1. Probe the workspace. Determine cwd, whether it is a git repo, dirty state, and the operator's current request. Use references/environment-probe.md for the command set.
   • Done when you can name the working tree and whether uncommitted work exists.
2. Select the live instructions. Read the nearest AGENTS.md / project instructions only if they apply to this tree. Load one specialist skill for the request, not a survey.
   • Done when the next action is pointed at a real file or command.
3. State the map in five lines or fewer. cwd, repo/branch/dirty, relevant skill, constraint, first action.
   • Done when the map fits in five lines and contains no capability advertisement.
4. Start the request. The bootstrap is finished only after the first request-advancing tool call.
   • Done when inspection or mutation of the actual target has begun.

Rules

• Do not list installed skills, models, or policies unless the request is about those things.
• Prefer local evidence over asking the operator to restate files you can read.
• If context was lost, recover from a handoff, receipt, or git state before asking for a recap.

Completion Gate

• cwd and repo dirty-state are known.
• One relevant skill or none, with a reason.
• First request-advancing action happened in the same turn.
• No inventory dump.

Common Failure Modes

• Capability brochure instead of a map.
• Stopping after ls.
• Reloading every adjacent skill.