scheduled-run

name	scheduled-run
description	Wrapper for scheduled playbooks - handles git sync before and after execution
argument-hint	<playbook-name>
automation	autonomous
allowed-tools	Bash, Skill
user-invocable	true

Scheduled Run

Wrapper that runs any playbook with git sync. Use this for all scheduled executions to ensure:

• Latest changes pulled before running
• All changes committed and pushed after running

Purpose

Ensures scheduled playbooks on Trinity stay in sync with the GitHub repository.

State Dependencies

Source	Location	Read	Write	Description
Git repo	.git/	✓	✓	Pull before, commit+push after
Playbook output	varies		✓	Whatever the playbook produces

Arguments

$ARGUMENTS = The playbook name to run (e.g., refresh-index, coherence-sweep)

Process

Step 1: Pull Latest Changes

git pull --rebase

If conflicts, abort and log error. Do NOT run the playbook on a dirty tree.

Step 2: Run the Playbook

Scheduled playbooks run at the core read-scope by default. BRAIN_READ_SCOPE is unset here, so every run_search.sh / run_connections.sh call inside the playbook fails closed to core (the fingerprint) - which is the intended scope for incubation-loop, refresh-index, coherence-sweep, gjopen-refresh, and ai-crystallize. Perception/cross-domain playbooks that legitimately need non-core content (e.g. domain-watch) set a wider BRAIN_READ_SCOPE=... inline on their own search commands; do not widen it here. (Setting export BRAIN_READ_SCOPE=core here would be a redundant fail-safe - the default is already core - and would not propagate across the separate Bash invocations a playbook makes anyway.)

Invoke the specified playbook using the Skill tool:

Skill: $ARGUMENTS

Wait for completion.

Step 3: Check for Changes

git status --porcelain

If no changes, skip to completion.

Step 4: Commit Changes

git add -A && git commit -m "$(cat <<'EOF'
Scheduled: $ARGUMENTS $(date '+%Y-%m-%d %H:%M')

Automated execution via /scheduled-run

Co-Authored-By: Claude <[e-mail]>
EOF
)"

Step 5: Push to Remote

git push

Error Handling

Error	Recovery
Pull fails (conflicts)	Abort, log error, do not run playbook
Playbook fails	Still attempt commit/push of partial changes
Push fails	Log error, changes remain local

Autonomy Guide

What should be scheduled (mechanical, no creative decisions):

Schedule	Command	Cron (UTC)	Timeout	Purpose
Incubation Loop	/scheduled-run incubation-loop	0 * * * *	45m	Hourly thinking on active topics (self-seeds via starvation floor when empty) — the :00 anchor
Domain Watch	/scheduled-run domain-watch	30 */2 * * *	30m	Every 2h at :30 — perception layer; auto-activates topics (incl. starvation floor)
GJ Open Refresh	/scheduled-run gjopen-refresh	30 1 * * *	30m	01:30 daily — external forecasting data [disabled]
Refresh Index	/scheduled-run refresh-index	30 3 * * *	40m	03:30 daily — FAISS + BDG bootstrap
Coherence Sweep	/scheduled-run coherence-sweep	30 5 * * 1	30m	Mon 05:30 — staleness/health briefing [disabled]
AI Crystallize	/scheduled-run ai-crystallize	30 7 * * 1	30m	Mon 07:30 — AI synthesis of converged topics → AI Crystallizations/ (ai-inferred)
Git Sync	(raw git)	45 */6 * * *	10m	:45 every 6h — redundant with scheduled-run git hygiene [disabled]

Spacing rule: incubation holds the hourly :00 slot; every other job runs at :30 of an odd hour (domain-watch only runs even hours), so no two jobs start in the same minute and their git commits never collide. Timeouts match each skill's work budget (the 15-min platform default is too short for FAISS rebuilds and multi-domain scans).

What should NOT be scheduled (requires human judgment):

Skill	Why manual
/auto-discovery	Generates new connections - that's intellectual work
/deep-research	Creates new content from external sources
/detect-tensions	Productive contradictions need human interpretation
/compute-lifecycle	Promotions to framework status need human decision
/integrate-recent-notes	Integration choices are creative decisions
/analyze-kb	Useful on-demand, not worth weekly automation

The principle: Automate measurement and maintenance. Never automate creation or judgment into the endorsed knowledge base. The refinement (2026-06): autonomous synthesis is permitted when its output is segregated and tagged ai-inferred - the starvation floor self-seeds topics, and ai-crystallize writes AI conclusions to Brain/05-Meta/AI Crystallizations/ (never 02-Permanent/). The gate that stays human is the endorsement act: only the user promotes an ai-inferred conclusion into permanent, endorsed knowledge via /manage-thinking-topics crystallize.

The coherence sweep is the one weekly artifact that earns automation - it's a Monday briefing that says "here's what changed, here's what might be stale, here's what's evolving." You glance at it, act on what matters, ignore the rest.

Completion Checklist

• Git pull completed successfully
• Playbook executed
• Changes detected and committed (if any)
• Changes pushed to remote