Summary

Introduce a skill lifecycle management system with two complementary capabilities: (1) automatic skill optimization when execution fails or underperforms, and (2) a usage-based retirement mechanism that archives and eventually removes inactive skills.

Problem to solve

OpenClaw skills are treated as static artifacts. Once created, they never evolve in response to real-world execution results. This creates two pain points:

No feedback loop on failure. When a skill execution fails, produces incorrect output, or underperforms, the user must manually diagnose the issue and edit the skill definition. The system has all the information needed (error traces, user corrections, task outcome) but discards it. Over time, skills drift out of sync with the tools, APIs, or workflows they target.

Skill bloat with no cleanup path. Users accumulate skills over months. Active skills get buried under hundreds of rarely-used ones, inflating skill search noise and context overhead. There is no built-in mechanism to identify, archive, or retire stale skills. The existing skills.workshop.autonomous.enabled can create proposals but has no corresponding retirement pathway.

Proposed solution

A. Auto-optimization on failure

After a skill execution, detect signals of poor performance:

• Hard failure: skill threw an error or produced an unparseable result
• Soft failure: user provided a correction or override ("no, do it this way")
• Performance degradation: skill took significantly longer than historical average, or required retries

When detected, automatically generate a Skill Workshop proposal (not a direct edit — safety first) that:

1. Captures the failure context (error type, relevant input/output, user correction)
2. Proposes a concrete modification to the skill definition
3. Tags the proposal as auto-optimization with the failure signal as provenance
4. Presents it for user approval through the existing Workshop review flow

This is distinct from #95477's "post-task self-reflection" (which focuses on discovering new skills from successful tasks). This proposal focuses on repairing existing skills from failure signals. Both are valuable and complementary.

B. Usage-based retirement mechanism

Track per-skill usage metrics:

Apply a tiered retirement policy:

Archived skills should remain readable and restorable. The archive is a visibility filter, not a data deletion.

Integration with existing systems

• Skill Workshop: proposals from auto-optimization flow through the existing Workshop approval gate (like skills.workshop.autonomous.enabled proposals)
• ClawHub: skill retirement telemetry (anonymized) could inform ClawHub skill quality rankings
• Dashboard: a openclaw skills lifecycle command showing active / archived / retiring skills with their metrics

Alternatives considered

• Manual-only curation: Users can already delete skills manually. This doesn't scale — most users won't audit 100+ skills regularly.
• Aggressive auto-delete (30d/90d as in the Hermes Curator approach referenced in Feature Request: Post-task self-reflection for skill auto-creation (inspired by Hermes + SkillClaw), + prefix cache optimization (inspired by Reasonix) #95477): Too aggressive for many workflows where skills are seasonal or project-specific. The 180d/365d thresholds balance cleanup with safety.
• Skill scoring only (no retirement): Showing quality scores without taking action helps power users but doesn't solve the bloat problem for typical users.

Impact

• Affected users: All OpenClaw skill users, especially those with 50+ skills or long-running workspaces
• Severity: Medium — skill bloat is a gradual productivity drag, not a blocker
• Frequency: Continuous — every session uses the skill registry
• Consequence of inaction: Users manually manage skill lifecycle or accept degraded skill search quality as their collection grows

Evidence/examples

• Hermes Agent (Nous Research): implements a Curator daemon that archives skills after 30d inactivity, deletes after 90d. Their approach is more aggressive; this proposal uses more conservative thresholds.
• SkillClaw (AMAP-ML): demonstrates cross-session skill refinement with success/failure pattern analysis — validates that execution outcome tracking is feasible and valuable.
• Related: Feature Request: Post-task self-reflection for skill auto-creation (inspired by Hermes + SkillClaw), + prefix cache optimization (inspired by Reasonix) #95477 (post-task self-reflection for skill auto-creation) — complementary but focused on skill creation, not optimization or retirement.

Additional information

The auto-optimization and retirement features should be independently toggleable:

skills:
  autoOptimization:
    enabled: true
    triggerOn: [hard_failure, user_correction]  # soft_failure optional
  retirement:
    enabled: true
    archiveAfterDays: 180
    retireAfterDays: 365
    autoDelete: false  # requires explicit confirmation by default

This ensures users who want only one feature (e.g., retirement without auto-optimization) can configure it independently.