Self_Code_and_Extension

Self-Code and Extension

Orrin can write new cognitive functions, tools, and skills for itself and register them live without restarting. This is the capability the README classifies as experimental / high-risk, and it is fenced accordingly — the guardrails are structural, not honor-system.

The writer

brain/agency/code_writer.py:

• Can only write to two directories: brain/cognition/custom_cognition/ and brain/agency/skills/.
• Cannot touch the selection/repair machinery: think/, cognition/repair/, ORRIN_loop.py, registry/, core/ are blocked paths. The code that decides what runs, and the code that repairs code, cannot be self-modified.
• All generated code is validated in the sandbox before being registered (brain/think/sandbox_runner.py); code that fails validation never joins the registry.
• A manifest records everything Orrin has written, so self-authored code is enumerable and auditable rather than scattered.

Review and quality

• The Architect peer (brain/peers/architect.py) reviews self-modifications before they happen — an outside observer whose whole job is proposed code (see Peers Subsystem).
• Verified sandbox checks (produce_and_check) are recorded on the effect ledger as tool_run_effect, so proven code pays and unverified code doesn't.
• The quality standard applies to produced artifacts, and Orrin cannot edit that standard.

Registration

New functions register into the same registry the bandit selects from, so a self-written function competes for selection like any built-in: it must earn reward to keep being picked, and the rut breaker / outcome devaluation keep a mediocre self-written function from monopolizing cycles.

Operator guidance

Treat self-extension as an area of caution (see SECURITY.md): review generated code before relying on it, and remember the blocked-path list is the contract — changes to it deserve the same scrutiny as changes to the sandbox.

Code pointers

• brain/agency/code_writer.py — writer + guardrails + manifest
• brain/cognition/custom_cognition/, brain/agency/skills/ — the only writable targets
• brain/think/sandbox_runner.py — pre-registration validation
• brain/peers/architect.py — review before effect