Module evolution 

Self-evolution: agent evaluates its own runs and proposes improvements.

After each task, the agent can analyze telemetry and identify bottlenecks. Improvements are stored as tasks in .tasks/ for the next run.

Evolution loop

Run task → Collect telemetry → Evaluate → Propose improvements →
→ Pick improvement → Patch code → Test → Commit → Rebuild → Restart →
→ Run task (with improvement) → ...

Structs

CircuitBreaker

Circuit breaker: stops after N consecutive identical failures.

EvolutionEntry

One entry in evolution.jsonl — records a single self-improvement attempt.

Improvement

A proposed self-improvement.

LoopOptions

Loop configuration (mirrors solo-dev.sh flags).

LoopState

Loop state — tracks progress across iterations.

RunStats

Telemetry from a single agent run.

SessionPattern

Pattern found across multiple sessions.

Enums

ControlAction

Control commands via file. Compatible with solo-dev.sh. Write “stop”, “pause”, or “skip” to the control file.

LoopMode

SoloSignal

Solo-compatible signals in agent output. <solo:done/> = stage complete, move to next <solo:redo/> = go back to previous stage (e.g. review found issues)

Functions

analyze_sessions

Analyze recent session logs for recurring issues. Reads last max_sessions JSONL files from agent home dir.

baseline_score

Get the latest score (baseline for comparison).

check_control

Check control file for commands. Reads and deletes (except pause which persists).

evaluate

Analyze run stats and propose improvements.

evolution_log_path

Default evolution log path.

evolution_prompt

Build a prompt that asks the agent to improve itself.

evolution_prompt_with_history

Build an evolution prompt that includes session history analysis.

evolution_summary

Count how many “keep” vs “discard” in history.

format_improvements

Format improvements as a markdown task list for the agent.

load_evolution

Load evolution history.

log_evolution

Append an entry to evolution.jsonl.

parse_signal

Parse solo signals from agent output.

score

Efficiency score: 0.0 (terrible) to 1.0 (perfect). successful_calls / steps weighted by completion.