autonomous-improvement

name	autonomous-improvement
description	Platform improvement governance — what to auto-fix vs what needs approval, detect-fix-report pattern, and improvement cadence rules. Use when user says "auto improve", "platform fix", "should I ask first", "auto implement", "improvement approval", "detect fix report", "nightly fixes", "autonomous changes", or any platform self-improvement activity.

Autonomous Improvement Skill

Governance framework for platform self-improvement. Defines what maintenance agents (context-auditor, skill-optimizer, platform-manager, nightly reflection) can auto-implement vs what requires human approval.

Core Mandate

"I'm not approving anything. You should automatically improve everything." — platform owner (2026-05-05)

The pattern: Detect → Fix → Report. NOT Detect → Propose → Wait → Fix.

Auto-Implement Immediately (NO approval needed)

These categories are pre-approved for autonomous execution by any maintenance agent:

Category	Examples
Agent instruction updates	Fix stale refs, outdated context, wording improvements
Skill extraction & optimization	Create new skills, refactor agents to use skills
Memory cleanup	Trim bloated working.md, fix stale references, prune
Configuration fixes	Cron schedules, quiet hours violations, stale prompts
Standing order updates	Persist new patterns, update lessons learned
Context hygiene	Fix pregnancy→postpartum refs, outdated family status
Working memory updates	Any agent's working.md maintenance
Template & config maintenance	Fix task templates, update config files
Data cleanup	Shopping list dedup, task dedup, stale data removal
Copilot-instructions updates	Non-breaking improvements, new learned behaviors
Skill content updates	Fix outdated procedures, add missing rules
Extension config tweaks	Non-breaking configuration adjustments

Still Requires Approval (Escalate First)

These categories need explicit human sign-off before execution:

Category	Why
Creating brand-new domain agents	New .github/agents/ files change the platform topology
Deleting or disabling agents/extensions	Destructive — may lose capabilities
Architectural changes	New data models, new extension patterns, schema migrations
Security-sensitive changes	Auth flows, secret handling, permission changes
Notification behavior changes	Affects how family members receive messages
Major cross-agent refactors	Changes communication patterns between agents
Cron schedule changes	Affects when agents fire (could cause spam or gaps)

Implementation Cadence

Trigger	Agent	Timing	Action
Nightly reflection	platform-manager	9 PM CT	Auto-implement ALL Tier 1/2 fixes found
Context audit	context-auditor	On-demand / scheduled	Find issues, fix auto-fixable, queue rest
Skill optimization	skill-optimizer	On-demand / scheduled	Extract skills, refactor agents, commit
Improvement tasks	Any maintenance agent	Within 24 hours of creation	Pick up queued improvement tasks and execute

The Detect → Fix → Report Workflow

Step 1: Detect

• Scan target files/systems for issues
• Classify each finding: auto-fix category or escalation-needed

Step 2: Fix (for auto-fix category only)

• Make the change directly (edit files, create skills, update configs)
• Commit with clear conventional message
• If multi-file: ensure consistency across all affected files

Step 3: Report

• Report what was DONE (past tense), not what's proposed
• Include in nightly Telegram summary or immediate notification for significant changes
• Log changes in agent working memory / events.log
• Format: "✅ Auto-fixed: [brief description of what changed and why]"

Step 4: Track

• Update working memory with changes made
• Note any follow-up work still needed
• Record if a pattern emerged that should become a recurring check

Escalation Format

When something needs approval, create a task:

add_task(
  title: "Approval needed: [brief description]",
  category: "general",
  priority: "medium",
  assignee: "primary_user",
  surface: "human",
  notes: "What: [proposed change]\nWhy: [reason]\nRisk: [what could go wrong]\nAffects: [which agents/files]"
)

Anti-Patterns

• ❌ Proposing improvements and waiting for approval when they're clearly auto-fix category
• ❌ Making approval-required changes without escalating first
• ❌ Reporting proposed changes instead of executed changes in summaries
• ❌ Accumulating improvement debt by deferring safe changes
• ❌ Asking "should I fix this?" for things in the auto-fix category — just fix them
• ❌ Making changes without logging them (always update working memory + events.log)

Relationship to Other Skills

• correction-persistence — When a user corrects behavior, persist across all layers. That's a specific case of the detect→fix→report pattern triggered by human feedback.
• agent-skill-management — The extraction methodology for skill-specific improvements. This skill governs the DECISION to act, not the HOW of skill extraction.
• repo-workflow — For larger fixes that need isolated branches. Most auto-fixes are small enough for direct edits.