Claude-Mind: Skill for Claude Code

Evolution Engine (evolve sub-command)

Trigger evolution check — the system evaluates its own strategy and generates new aspirations. Invoked by Phase 9 performance-based evolution triggers, or directly via /aspirations evolve. Covers developmental stage assessment, config parameter tuning, state reading, evolve-first aspiration review, constraint-aware rebalancing, self-driven gap analysis, novelty filter, cap enforcement, logging, profile/meta update, forge check, pattern signature calibration, and strategy archive.

Step 0: Load Conventions — Bash: load-conventions.sh with each name from the conventions: front matter. Read only the paths returned (files not yet in context). If output is empty, all conventions already loaded — proceed to next step.

evolve

Trigger evolution check — the system evaluates its own strategy and generates new aspirations:

1. Developmental Stage Assessment (competence-based):

   Read core/config/developmental-stage.yaml (stage definitions, competence_mapping)
   Read agents/<agent>/developmental-stage.yaml (current assessment, epsilon, schema log)
   old_stage = agents/<agent>/developmental-stage.yaml.overall_stage
   old_highest = agents/<agent>/developmental-stage.yaml.current_assessment.highest_capability
   
   # Stage-block computation is SCRIPT-ENFORCED (g-115-2624). The former inline
   # formula here (tree-leaf capability mean at depth >= 2, competence_mapping
   # EXPLORE=0.15/CALIBRATE=0.45/EXPLOIT=0.70/MASTER=0.90, stage bands
   # 0.30/0.55/0.80, exploration_budget = clamp(1 - tree_maturity, 0.15, 0.85))
   # now lives in core/scripts/_competence.py::assess_stage — the SAME math,
   # deterministic and refreshed at every curriculum-evaluate chokepoint too.
   # WHY: LLM-discretionary producers drift silent (rb-3171); this exact block
   # fired once in 2 months on zeta, pinning ZDS agents at 'exploring' with
   # forge gated and resolved_hypotheses contradicting its own sibling
   # evidence (ZDS omni code-read 2026-07-18). The one call below computes AND
   # writes: overall_stage, current_assessment.{stage,tree_maturity,
   # highest_capability,lowest_capability,exploration_budget,
   # resolved_hypotheses(=pipeline_resolved),average_competence,components,
   # evidence,producer,assessed_at}, exploration.epsilon. Do NOT re-inline the
   # math here.
   Bash: py -3 core/scripts/competence-assess.py
   Parse JSON stdout → stage_assessment.{stage, tree_maturity, highest_capability, exploration_budget}
   
   If stage_assessment.highest_capability != old_highest:
     Log: "DEVELOPMENTAL UPDATE: highest_capability → {stage_assessment.highest_capability}"
   
   If stage_assessment.stage != old_stage:
     # Numbers are already written by the script; the LLM applies the
     # stage-definition SIDE EFFECTS (behavioral envelope from
     # core/config/developmental-stage.yaml stages.<stage>.behaviors):
     Update allowed hypothesis_types and max_commitment (per stage definition)
     Log: "DEVELOPMENTAL TRANSITION: {old_stage} → {stage_assessment.stage}"
   
   Metacognitive self-check (every 5th goal via sq-010):
     "Based on my knowledge, accuracy, and experience — what capability level am I at?
      Does it match the computed level?"
     If divergence: log as ACCOMMODATION in schema_operations.log
   
   Schema operation detection:
     Read recent reflections and violations
     For each finding that contradicts existing framework:
       Log as ACCOMMODATION in schema_operations.log
       Set equilibration_state = "disequilibrium"
     For each finding that confirms existing framework:
       Log as ASSIMILATION in schema_operations.log
   
   Update agents/<agent>/developmental-stage.yaml:
     schema_operations.log + equilibration_state ONLY (from the two blocks above).
     # ALL numeric assessment fields (overall_stage, current_assessment.*,
     # exploration.epsilon) are script-written by competence-assess.py — the
     # single producer since g-115-2624. Evolve MUST NOT hand-write any of
     # them; hand-writing re-creates the g-115-2028 last-writer-wins collision
     # AND the rb-3171 silent-drift class this delegation eliminated. Evolve's
     # write surface here is the schema_operations narrative alone.
   
   Run active forgetting pruning:
     Read core/config/memory-pipeline.yaml forgetting config
     For each leaf node, calculate retention score
     If retention < 0.4: archive (if validated) or deprecate