schedule

name	schedule
description	Discover files with trigger_phrase and/or refresh frontmatter, let the user choose which to schedule for periodic re-injection via CronCreate. Prevents behavioral drift by keeping core rules and goals fresh in context.

Schedule — Periodic Trigger and Rule Injection

Overview

Scan the agent's loaded context for files with trigger_phrase and/or refresh frontmatter. Present candidates to the user in batches of 5. For each selected file, create a CronCreate job that periodically re-injects the trigger phrase (or re-reads the file) to prevent behavioral drift during long sessions.

Why This Exists

During long sessions, core rules and goals fade as context compresses. Agents stop following guidance they loaded 2 hours ago — not because they disagree, but because the content lost attention weight. Periodic re-injection of trigger phrases (5 tokens that reactivate 1000 tokens of rules) keeps agents calibrated without expensive full re-reads.

Frontmatter Schema

Files opt into scheduling with these frontmatter fields:

---
name: some-rule
description: "..."
trigger_phrase:                    # set by /total-recall:write
  opus: "broken windows zero tolerance"
  sonnet: "quality gate enforcement"
  haiku: "no broken windows"
refresh: 15m                       # suggested interval
---

• trigger_phrase — model-scoped phrases generated by total-recall. The schedule injects the phrase matching the current session's model. Cheap (~5 tokens).
• refresh — suggested interval for re-injection. User can override during scheduling. Supports: 5m, 10m, 15m, 20m, 30m, 1h.
• Both fields are optional. A file needs at least one to be discovered.

Process

1. Discover Candidates

Scan for .md files that have trigger_phrase and/or refresh in their YAML frontmatter. Search these locations:

1. .claude/rules/ — project rules
2. CLAUDE.md — project root instructions
3. Loaded skills — any skill .md files in the current plugin paths
4. .claude/*.md — any markdown files in the .claude directory

For each file found:

1. Read the YAML frontmatter
2. Check for trigger_phrase and/or refresh fields
3. If either exists, add to the candidate list

Determine the current model by checking which model this session is running. Use this to select the right trigger phrase from the model-scoped trigger_phrase field.

2. Present to User

If no candidates are found, report: "No files with trigger_phrase or refresh frontmatter found. Run /total-recall:write to add trigger phrases to files, or add refresh: 15m to any rule's frontmatter."

Present candidates in batches of up to 5 via AskUserQuestion (multi-select).

For each candidate, show:

• File path
• Trigger phrase for current model (or "none — will re-read file" if no trigger_phrase)
• Suggested interval from refresh field (or "no default — you choose")

Example:

"These files can be scheduled for periodic refresh. Select which to activate:"

1. .claude/rules/broken-windows.md — trigger: "broken windows zero tolerance" — interval: 15m
2. .claude/rules/error-handling.md — trigger: "boundary validation fail fast" — interval: 20m
3. .claude/skills/collab/SKILL.md — trigger: "collaborative session partnership" — interval: 30m
4. CLAUDE.md — no trigger, will re-read — interval: 10m

3. Confirm Intervals

After the user selects files, present the intervals for confirmation via AskUserQuestion:

"Confirm intervals for selected files (or adjust):"

For each selected file, show the default interval and let the user override. If no refresh field exists, suggest 20m as a default.

4. Create Schedules

For each confirmed file, create a CronCreate job:

If trigger_phrase exists for the current model:

The injection prompt is compact — just the trigger phrase with minimal framing:

Re-internalize this core rule: "{trigger_phrase}"

Briefly acknowledge what this rule means and confirm you're following it. One sentence.

If no trigger_phrase (only refresh):

The injection prompt re-reads the file:

Re-read and internalize {file_path}. This is a core rule/skill that must stay fresh in your working context.

Briefly confirm you've re-read it and state the key principle. Two sentences max.

CronCreate parameters:

• cron: Convert the interval to a cron expression. Avoid :00 and :30 marks (per CronCreate best practices).
• prompt: The injection prompt above.
• recurring: true

5. Report Created Schedules

After all jobs are created, summarize:

Scheduled refreshes active:

  ID        File                              Interval  Type
  ────────  ────────────────────────────────  ────────  ──────────
  abc12345  .claude/rules/broken-windows.md   15m       trigger
  def67890  .claude/rules/error-handling.md   20m       trigger
  ghi24680  CLAUDE.md                         10m       re-read

3 refresh jobs active. They'll fire while the session is idle.
Use "cancel the broken-windows refresh" or CronDelete to stop any job.
Jobs expire after 7 days (CronCreate limitation).

Managing Schedules

After initial setup, the user can manage schedules with natural language:

• "What scheduled tasks do I have?" → triggers CronList
• "Cancel the broken-windows refresh" → triggers CronDelete
• "Change the CLAUDE.md refresh to every 30 minutes" → delete + recreate

The skill does not need to handle these — Claude's built-in cron management covers them. Just inform the user of the option.

Interval Guidelines

Content type	Suggested interval	Rationale
Core rules (broken-windows, error-handling)	10-15m	High-value, small injection cost
Skill philosophy (collab, build)	20-30m	Important but less likely to drift
CLAUDE.md	10-20m	Project-wide instructions, critical
Reference docs	30m-1h	Informational, less calibration-sensitive

These are suggestions for when users ask. The refresh frontmatter in each file is the authoritative suggestion.

Session Scope

All schedules are session-scoped:

• They die when the session exits
• They only fire while Claude is idle (not mid-response)
• They auto-expire after 7 days
• Re-run /total-recall:schedule in each new session to re-activate

For cross-session persistence, users should set up Desktop scheduled tasks or Routines (outside this skill's scope). A future SessionStart-hook-based auto-restore is planned (see plan F7).

What Not to Do

• Don't schedule without asking. Every file goes through AskUserQuestion.
• Don't schedule too aggressively. Injections every 2 minutes flood context. Minimum practical interval is 5m.
• Don't re-read large files on tight intervals. If a file is > 200 lines and has no trigger_phrase, suggest the user run /total-recall:scan followed by /total-recall:write first to generate a cheap trigger.
• Don't create duplicate schedules. Check CronList before creating. If a job already exists for a file, ask the user if they want to replace it.
• Don't inject silently. The injection prompt asks the agent to briefly acknowledge — this makes the refresh visible in the conversation.