Check if text contains potential API keys or tokens.

Validate and normalize fields before creating an EvalExample. Returns: Dict of validated fields, or None if the example should be skipped

Read user/assistant message pairs from Copilot sessions. Args: limit: Maximum messages to return (0 = no limit). Returns

Extract a JSON object from LLM scoring output. Strategy: 1. Try direct json.loads (handles clean LLM output) 2. Fall back to regex ex

Extract messages from external tools, filter for relevance, and save. This is the main entry point called by both the standalone CLI and evol

Filter messages by relevance and generate eval examples. Args: messages: Raw messages from importers. skill_name: Nam

Quick heuristic check if a message might be relevant to a skill. Uses keyword overlap between the message and skill description/name. This is

Load a skill file and parse its frontmatter + body. Returns: { "path": Path, "raw": str (full file content),

Check that a skill file has valid YAML frontmatter and markdown body.

Load skill text from the installed Hermes skills directory. This is used by the standalone CLI only. When called via evolve_skill.py, skill l

Load dataset splits from JSONL files.

Parse a single Copilot events.jsonl into user/assistant pairs.

Extract cwd from a Copilot workspace.yaml file.

Reassemble a skill file from frontmatter and evolved body. Preserves the original YAML frontmatter (name, description, metadata) and replaces

Return the hermes-agent repo path, honoring an explicit override. An explicit path (for example from ``--hermes-repo``) is expanded and used

Run all applicable constraints. Returns list of results.

Parse a score value, handling various LLM output formats.

Convert a split to DSPy Example objects.

Discover the hermes-agent repo path. Priority: 1. HERMES_AGENT_REPO env var 2. ~/.hermes/hermes-agent (standard install location) 3.

Save dataset splits to JSONL files.

DSPy-compatible metric function for skill optimization. This is what gets passed to dspy.GEPA(metric=...). Returns a float 0-1 score.

Best-effort hermes-agent repo discovery that never raises. Returns the discovered path, or None when no repo can be found. Used as the Evolut

Create a minimal hermes-agent-style repo with one skill.

Main evolution function — orchestrates the full optimization loop.

Find a skill by name in the hermes-agent skills directory. Searches recursively for a SKILL.md in a directory matching the skill name.

Generate a full eval dataset with train/val/holdout splits.

Import external session data into golden eval datasets for self-evolution.

Evolve a Hermes Agent skill using DSPy + GEPA optimization.

Weighted composite score.

Read user messages from Claude Code history. Args: limit: Maximum messages to return (0 = no limit). Returns:

Read user/assistant pairs from Hermes session files. Args: limit: Maximum messages to return (0 = no limit). Returns:

Load a golden dataset. If no splits exist, auto-split the single file.

Mock dspy.LM and dspy.context to avoid real LLM calls.

Run the full hermes-agent test suite. Must pass 100%.

Score an agent output using LLM-as-judge.

Verify end-to-end: import -> filter -> split -> save.

`evolve_skill --hermes-repo <path> --dry-run` must work when the default location is absent. This is the exact scenario the bug broke.

A bare EvolutionConfig() must not raise when no repo is present.

An explicitly supplied path is kept as-is, never overwritten by discovery.

Clean JSON should be parsed directly without regex fallback.

LLM returns relevant=True but empty expected_behavior -> example dropped.

Create real Copilot events, import, filter (mocked), save, reload.

Verify output matches GoldenDatasetLoader expected format.