Self Evolve

Self-Evolution

Analyze your own performance history and fix recurring issues. ONE improvement per iteration.

Process

1. Read .rust-code/evolution.jsonl + recent session logs
2. Find patterns: patch failures, loop warnings, error spikes
3. Trace to root cause in source code
4. Apply minimal fix, make check, commit
5. Signal RESTART_AGENT if you changed agent code

Reference files for diagnosis: references/diagnosis-paths.md

Key Source Files

| Symptom | Source | |---------|--------| | Patch matching failures | crates/sgr-agent/src/app_tools/apply_patch.rs | | Agent stuck in loops | crates/sgr-agent/src/loop_detect.rs | | Prompt confusion | crates/rc-cli/src/agent.rs (SGRSYSTEMPROMPT) | | Tool argument errors | crates/rc-cli/src/agent.rs (execute_action) | | Schema issues | crates/sgr-agent/src/schema.rs |

Signals

• Fixed something → finish: "RESTART_AGENT — [what you fixed]"
• Nothing to fix → finish: ""
• Stuck → finish: " — need human input"

Gotchas

1. RESTART_AGENT kills your context — everything in memory is lost. Commit first, write notes to .tasks/ for the next session to pick up.
2. Simplicity threshold — a 0.001 score improvement that adds 20 lines of hacky code is NOT worth it. Removing code with equal results is always a win.
3. make check can take 2+ minutes — don't call it in a tight loop. Run once after your change, fix issues, run once more to verify.
4. Don't optimize what you can't measure — read actual session logs, grep for error patterns. Don't guess at improvements.
5. Evolution log is append-only — don't edit evolution.jsonl. The loop engine reads it for scoring trends.

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: fortunto2
• Source: fortunto2/rust-code
• License: MIT

Install and usage instructions live in the source repository linked above.