sennight

n., archaic. seven nights; a week.

Your week is a git repo. You text a bot what happened; every six hours an agent reads your messages, replans the rest of the week, admits what will not fit, and commits. No app, no database, no service.

The whole system is a few hundred lines: a markdown file of standing instructions, two YAML files, one validator, one cron, one HTML page. Small enough to read in one sitting, and made to be copied, the way micrograd is a model you can read, this is a life planner you can read.

How a run works

1. Fetch. Pull whatever is new on Telegram and in the bot's Discord DM, save it verbatim to inbox/. That file is also the read cursor; there is no other state.
2. Interpret. "did 2h dsa yesterday" is a log entry. "report due Friday, ~5h" is a task. "no gym this week" is an override that dies with the week. Only "permanently" touches the base config. Anything unclear goes to questions.md; hours are never guessed.
3. Replan. Free windows = waking hours minus fixed anchors. Fill at most 80% of any window with ordered blocks, deadline tasks first, then activities by priority. Durations, never clock times.
4. Validate. check_plan.py audits the arithmetic: every quota accounted for, every constraint honored, every miss declared. A failing plan is never committed.
5. Digest. One plain message back on every transport you wired: what was understood, what changed, today's remaining blocks, where the week is at risk.
6. Commit. The repo is the memory. Every run is a diff you can read.

Telegram and Discord both ship wired up; set up either one, or both, and the agent uses whatever it has secrets for. Neither is a dependency. Any other HTTP messaging API (Slack, email, a webhook) is a bullet in CLAUDE.md and two repo secrets away, with no code to change.

The whole thing

Counts are for the starting template. A copy that has been lived in for a while runs about half as long again, as house rules accumulate in the brain and the auditor learns your constraints by name. That is the config doing its job, not rot.

The rules that make it work

• Logs are append-only. Mistakes are corrected with negative entries, never edits. The ledger is trustworthy because it cannot be rewritten.
• Shortfalls are declared, not hidden. If the week does not fit, the plan says so, with a reason. Honest defeat beats a cramming fantasy.
• Anchors cannot be silently dropped. A cancelled class is declared in overrides.yaml. A plan that quietly deletes a lecture and refills the window fails the audit, which is the easiest way for a system like this to lie to you.
• The newest message wins. Contradictions resolve forward, and the digest says so.
• The schema is frozen. The agent proposes changes in questions.md; it does not silently reshape its own contract.
• Silence is forbidden. Every run that changed anything reports back.

Copy it

gh repo create my-sennight --private \
  --template agentlab-in/sennight
# then, in your copy:
# 1. move template/* to the repo root, delete the rest
# 2. rewrite config.yaml with your real week
# 3. add secrets: CLAUDE_CODE_OAUTH_TOKEN, then
#    TELEGRAM_BOT_TOKEN + TELEGRAM_CHAT_ID and/or
#    DISCORD_BOT_TOKEN + DISCORD_CHANNEL_ID
# 4. enable Actions, enable Pages, text your bot

That board is this repo's template/ directory running on its fictional week, rendered straight from plan.json with no build step.