James Washington
writing about archive utils free game ideas replays
Agent hook

Continuity system: autopilot + context-budget relay

Keeps a long AI task moving without James typing 'yes go ahead' at every step, and passes the work to a fresh AI before the old one gets confused.

built Sep 2026 updated Sep 2026 Upgrading the AI's workbench

how it works

Stop hook tails the transcript: blocks the stop with 'next: X' while todos remain (capped hops, reset by continuity_reset.py when James types), triggers /handoff past a soft token budget at a green boundary or a hard budget regardless; relay_spawn.sh starts a new claude in tmux seeded with /pickup; /autopilot toggles via autopilot.sh. Fails open.

Agent hook: A script that runs automatically at a moment in an AI session (before a command, when it stops) to steer or guard it.

The code for this one lives in a private repo. This page explains the idea so you can build your own.

autopilotcontext-managementhandoffhooksstop-hook

related

  • /autopilot: auto-continue toggle Slash command Turns off the 'should I continue?' interruptions for one session.
  • tmux window auto-titling Agent hook Names each terminal tab after what the AI is working on.
  • auto-handoff mod Plugin If an AI task is still unfinished just before its cheap memory cache expires, it writes the hand-over note automatically.
  • Mission Control presence/telemetry hooks Agent hook Every AI session automatically shows up live on the operations dashboard: when it starts, works, edits files and stops.
  • ~/.claude config repo: deny-by-default .gitignore + secret pre-commit Automation Keeps James's AI configuration under version control without ever committing transcripts or passwords.
← everything I've built
© 2026 James Washington
github twitter rss