Session bootstrap

session-bootstrap — sessions boot from files, not memory

Chat history evaporates between sessions and summaries are lossy; the project's durable state lives in files. This is the boot and shutdown sequence.

OPEN — starting a session that continues earlier work

1. Read the state, in this order (skip what doesn't exist):
   • .claude/state/<task-slug>.md — active task state (pseudo-fable-lift long-task-state)
   • .claude/state/delegations.md — in-flight delegated work (pseudo-fable-orchestrate ledger)
   • docs/plan/<slug>/03-tickets.md — ticket statuses and traceability (pseudo-fable-blueprint)
   • git log --oneline -15 and git status — what actually changed vs. what the files claim
2. Verify drift — the files were true when written, not necessarily now:
   • Re-verify the top ~3 load-bearing facts against the actual code (files move; teammates commit).
   • Dirty git status not mentioned in the state file → investigate before continuing; someone (possibly a previous you) left uncommitted work.
   • In-flight delegations: check whether workers finished or died; reconcile the ledger before spawning anything new.
3. Restate before resuming — one short message: where the task stands (X/Y), the next single action, anything that drifted. This is the contract check; the user can correct it before you burn tokens.
4. Continue from Next action. Do not re-plan what is already planned — re-planning finished plans is how sessions lose their first hour.

CLOSE — ending the session (or the user says "done for now", or context is nearly full)

1. Update the state file(s): plan checkboxes, learned facts, failed approaches — and Next action, exactly one concrete step. That line is the single most valuable thing you leave for the next session.
2. Update ticket statuses and the delegations ledger: verdicts, in-flight items and what to check on them.
3. Park the working tree deliberately — commit or stash with a message the next session can read; never leave silent dirty state. If you are not authorized to commit, report exactly what is uncommitted and why.
4. If anything bounced, reverted, or dragged this session → run a micro-retro (≤2 min, ≤2 rules).
5. Close with a 3-line handoff in chat: done today / next action / open risks.

Anti-patterns

• Booting from your memory of the last session ("I remember we were...") — memory is the lossy copy; the files are the original.
• Re-reading the whole codebase at session start — the state file's Learned facts exist to make recon cheap. Verify them; don't re-derive them.
• Ending a session with the plan only in chat — chat is not a handoff.
• Treating OPEN as optional after compaction — a compacted context is a new session in disguise.