/evolve — Self-Evolution Cycle

You are executing the self-evolution cycle. This is the core loop that enables learning from feedback.

Input

The user has provided feedback about your recent performance. The feedback may be explicit ("you should have done X instead of Y") or implicit (a correction, a different approach, frustration).

Procedure

Step 1: Gather Context

Read the following files to understand current state:

1. •Recent events: Read .claude/path-kernel/event-log.jsonl (last 20 lines) to understand recent actions and modifications.
2. •Current rules: Read .claude/path-kernel/config.json for the authoritative list of mutable rule files (agentCanModify). Read each file listed there. These span three domains under .claude/rules/: world, valence, and skill.
3. •Cycle state: Read .claude/path-kernel/state.json if it exists, for cross-session context.

Step 1b: Log Feedback Event

After gathering context, emit a feedback:human event to the event log. This captures the feedback input for observability.

PROMPT_VERSION=$(jq '.cycleCount // 0' .claude/path-kernel/state.json 2>/dev/null || echo 0)
bash .claude/hooks/append-event.sh "feedback:human" \
  '{"summary":"<brief feedback summary>"}' "$PROMPT_VERSION"

Replace <brief feedback summary> with a 1-sentence summary of the user's feedback.

Step 2: Analyze the Gap

Compare what happened (the feedback) against what should have happened. Identify:

• •What went wrong (or what could be improved)
• •Root cause: Was it a misunderstanding of the situation (world), wrong priorities (valence), or poor technique (skill)?
• •Specific change that would prevent this gap in the future

Step 2b: Log Analysis Event

After analyzing the gap, emit an evolve:analysis event capturing the analysis results.

PROMPT_VERSION=$(jq '.cycleCount // 0' .claude/path-kernel/state.json 2>/dev/null || echo 0)
bash .claude/hooks/append-event.sh "evolve:analysis" \
  '{"summary":"<gap identified>","domain":"<world|valence|skill>","proposed_change":"<brief>"}' "$PROMPT_VERSION"

Replace the placeholders with the actual analysis from Step 2.

Step 3: Classify the Modification Domain

Route the change to the correct internal model:

Step 4: Propose and Apply the Modification

1. •Draft the change: Write the specific text to add, modify, or remove from the target rules file.
2. •State the rationale: Explain why this change addresses the gap.
3. •Log the proposal: Before applying the edit, emit a modification:proposal event:
   bash

   PROMPT_VERSION=$(jq '.cycleCount // 0' .claude/path-kernel/state.json 2>/dev/null || echo 0)
   bash .claude/hooks/append-event.sh "modification:proposal" \
     '{"domain":"<world|valence|skill>","target":"<file>","rationale":"<why>"}' "$PROMPT_VERSION"
   

   Replace placeholders with the actual domain, target file, and rationale.
4. •Apply the edit: Use the Edit tool to modify the target .claude/rules/{domain}/*.md file.
   • •Keep changes focused and incremental. Prefer adding a specific lesson over rewriting entire sections.
   • •Preserve existing content unless it directly contradicts the new learning.
5. •The PostToolUse hook will automatically log this modification to .claude/path-kernel/event-log.jsonl.

Step 5: Recompose CLAUDE.md

After modifying rules, update CLAUDE.md to reflect the change:

1. •Read the current CLAUDE.md.
2. •Increment the version comment at the top.
3. •The @ import references in CLAUDE.md point to the rules files, so the content updates automatically. But if the structure of CLAUDE.md itself needs updating (new sections, reordering), apply those edits now.

Step 6: Update Cycle State

Write to .claude/path-kernel/state.json:

{
  "lastTask": "<brief description of the task>",
  "lastFeedback": "<brief summary of feedback>",
  "lastModification": "<domain>/<filename>",
  "lastModificationRationale": "<why>",
  "awaitingFeedback": false,
  "cycleCount": <increment from previous>
}

Step 7: Report

Tell the user:

1. •What you analyzed
2. •What domain you classified the feedback into (world/valence/skill)
3. •What specific change you made and to which file
4. •Why this change should help in the future

Constraints

See .claude/rules/kernel/invariants.md for immutable constraints and .claude/path-kernel/config.json (agentCannotModify) for the protected file list.

• •If you're unsure whether a change is warranted, say so and ask for more feedback rather than making a speculative change.
• •Keep rule files concise. Prefer specific, actionable lessons over vague principles.
• •One evolution cycle = one focused change. Don't try to fix everything at once.