kubus/hatbot
1
0
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity Actions

Chat commands: /new, /stop, /task and /remember run in the brain #3

Merged
kubus merged 7 commits from t3/new-conversation-command into main 2026-10-13 00:20:45 +00:00
Conversation 0 Commits 7 Files changed 37 +1498 -571
kubus commented 2026-10-13 00:20:06 +00:00
Owner
Copy link

Summary

  • /new, /stop [task], /task <what> and /remember <fact> run in the brain without the model; unknown /word is an ordinary message
  • /new starts a fresh episode (no recent messages, no taint, only the running summary)
  • /task runs your words as an untainted background task
  • /remember appends to the Notes section of INDEX.md, as the owner
  • The model hears about commands in <owner_commands> with your next message
  • Approval system simplified: one unified Haiku review before any change that would ask, instead of separate "you already asked" and "is it safe" checks
  • Chart rendering improved with stacked and horizontal bars
  • System prompt clarified on quick jobs (max 2 changes per reply, past that goes to a task)

Testing

  • Web composer tested in headless Chromium
  • iOS views syntax-checked only
  • Database migration 18 required for fresh episodes
  • Brain must be redeployed with the new migration
  • Not tried live in a real conversation yet
## Summary - `/new`, `/stop [task]`, `/task <what>` and `/remember <fact>` run in the brain without the model; unknown `/word` is an ordinary message - `/new` starts a fresh episode (no recent messages, no taint, only the running summary) - `/task` runs your words as an untainted background task - `/remember` appends to the Notes section of INDEX.md, as the owner - The model hears about commands in `<owner_commands>` with your next message - Approval system simplified: one unified Haiku review before any change that would ask, instead of separate "you already asked" and "is it safe" checks - Chart rendering improved with stacked and horizontal bars - System prompt clarified on quick jobs (max 2 changes per reply, past that goes to a task) ## Testing - Web composer tested in headless Chromium - iOS views syntax-checked only - Database migration 18 required for fresh episodes - Brain must be redeployed with the new migration - Not tried live in a real conversation yet
A change that would ask now gets a single Haiku call answering both
questions that apply: whether you already asked for it, and after
untrusted content, whether it's safe without you. Either yes lets it
through, in the chat and in tasks.

A chat reply makes at most two homelab or account changes itself (the
sandbox doesn't count); a third is refused with a pointer to run_task,
and the prompt tells the chat to pick run_task up front for anything
bigger.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Every change that would ask, tainted or not, in the chat or in a task,
now goes to Haiku's review. If it isn't destructive or dangerous, it
runs and the timeline says so. The separate check for "you already
asked for this" is removed, along with its setting.

Still always asking, with no review: payments, purchases, account and
key changes, budget, suspected injection, hosts and guests marked
careful, and Home Assistant locks, the alarm, doors and gates
(askAlways in plugin policies).

Settings: one toggle, "Let Haiku allow safe changes" (reviewChanges,
carried over from reviewTaintedChanges), on web and iOS.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
/new starts a conversation that carries nothing over (no recent messages,
no taint); the last one folds under its recap. /stop ends the reply being
written, or cancels a task by number. /task runs what you typed as an
untainted background task. /remember adds the line to INDEX.md's Notes.
The model hears about commands with your next message.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Arrow keys pick, Enter or Tab chooses: /new and /stop run at once, /task
and /remember go in the composer to finish. Lines that answer your own
commands arrive already seen.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
kubus merged commit c7b2933d23 into main 2026-10-13 00:20:45 +00:00
Sign in to join this conversation.
Reviewers
No reviewers
Labels
Clear labels
No items
No labels
Milestone
Clear milestone
No items
No milestone
Projects
Clear projects
No items
No project
Assignees
Clear assignees
No assignees
1 participant
Notifications
Due date
The due date is invalid or out of range. Please use the format "yyyy-mm-dd".

No due date set.

Dependencies

No dependencies set.

Reference
kubus/hatbot!3
No description provided.