Self-improve loop

Arka analyzes its own codebase, proposes fixes, and optionally applies them with safety gates.

Self-improve lets Arka inspect its own repository — tests, routing, docs — and produce an improvement plan. Use --apply to run the goal agent and implement fixes.

Quick start

arka self improve              # plan only (dry-run)
arka self improve routing      # focused plan
arka self improve --apply      # implement via goal agent
arka self memory               # past improvement attempts
arka self status               # repo + memory path

Natural language:

improve arka
self improve routing
improve arka llm fallback

self improve routing is also routed as a dedicated skill so Arka can treat routing fixes as a first-class self-improvement task.

Output (normal mode)

━━━ Arka Self-Improve ━━━
Focus: routing
Checking: tests, routing, docs

✗ Tests: 1 failure (memory_detect)
✓ Routing: symbolic.py (40 route_* handlers)
○ Plan: fix import in tests/test_memory_detect.py

Next: arka self improve routing --apply

Full diagnostics (pytest traces, llm.txt context) appear only in debug mode.

Safety

Rule	Behavior
Plan-only default	No file writes without --apply
--apply requires agent mode	Blocked in ask or plan mode
Secrets blocked	Never modifies .env, credentials, node_modules
Memory	~/.config/arka/self-improve-memory.json tracks past attempts

Related

• Goal agent — multi-step autonomous tasks
• Testing — pytest and CI