v0.9.6

• fix: revisions (foo-rev-*) no longer blocked as duplicates of parent skill
• fix: frontmatter nesting check downgraded from error to warning (fixFrontmatterNesting handles it at generation time)

v0.9.5

• fix: skillDir undefined variable in resolveSkillActivation Match 4/5 (capture.ts)
• chore: remove OpenViking health check references from /forge doctor and /forge status

AceForge v0.9.4

Fix: behavior gap notifications rate-limited to once per 24 hours.

AceForge v0.9.3

Full manifest declaration for ClawHub scanner compliance. Declared hooks, tools, env vars, config paths, and network endpoints.

AceForge v0.9.2

Patch: generalized identityLinks lookup (no hardcoded references), added openclaw.compat.pluginApi for ClawHub registry.

AceForge v0.9.1

What's New Since v0.9.0

Full Lifecycle Integration Test (Stages 4-6)

• Stage 4: validateAndDeploy path — exercises the exact code path where the critical _skillAction scoping bug lived
• Stage 5: Activation tracking — recordActivation, getSkillStats, listActiveSkills, getSkillMaturity
• Stage 6: Evolution trigger — checkMilestone, executeEvolve, isShortCircuitCandidate
• 28/28 integration tests passing end-to-end

Multi-Agent Shared Skills

• ACEFORGE_SHARED_SKILLS=true deploys approved skills to ~/.openclaw/skills/ (visible to ALL agents on same machine)
• Follows OpenClaw's native skill precedence: workspace > shared > bundled

Composition Execution Bridge

• proposeCompositionSkills() converts co-activation detections into workflow skill proposals
• Template fallback generates valid SKILL.md without LLM keys
• sampleTraces populated from patterns.jsonl — LLM gets real execution context
• Wired into agent_end Phase 2 cycle

Notification UX

• Actionable gap notifications: every gap type now tells the user what happened, why, and what command to run
• Warming-up notification: "AceForge is watching — tracking 4 tools, exec is 2/3 to threshold" (24h rate limit)

Documentation

• PHILOSOPHY.md: mission statement on human-in-the-loop, validation-first, milestone-based evolution
• Plain-language opener: 3-sentence blockquote before badges
• "What AceForge Is (and Isn't)" section: explicitly names complementary tools (Self-Evolve, EvolveClaw, OpenClaw-RL)
• Provider table: 5 → 13 providers matching PROVIDER_DEFAULTS
• Channel section: FORMAT_MAP architecture with format types table

Stats

• 412/412 unit tests passing
• 28/28 integration tests passing
• 23/23 adversarial mutations caught

Upgrade

cd ~/.openclaw/extensions/aceforge
git pull origin main
npm install
openclaw gateway restart

AceForge v0.9.0 — Evolution Engine

What's New

Evolution Engine — 3 new modules for skill self-improvement:

• Trace distillation at milestones (500/2000/5000 activations) using K2-Agent SRLR
• Novel success capture with 6-check novelty filter (Voyager-inspired)
• /forge evolve — LLM-powered skill revision with unified diff + human approval gate

True Provider Agnosticism — OpenAI + Anthropic API formats:

• 13 built-in providers (MiniMax, DeepSeek, OpenAI, Anthropic, OpenRouter, Together, Groq, Cerebras, Hugging Face, Kimi, Ollama, LM Studio, vLLM)
• Format auto-detected from openclaw.json api field or provider name
• baseUrl config read fixed (was silently ignoring openclaw.json)

Channel-Agnostic Notifications — FORMAT_MAP architecture:

• Format types (html/markdown/mrkdwn/plain), not channel names
• Telegram HTML, Slack mrkdwn, Discord Markdown — all Tier 2
• Adding a channel = one line in FORMAT_MAP
• HTML injection safety via sanitizeTelegramHtml
• Every notification rewritten for humans — zero jargon

Maturity Fast-Track — isShortCircuitCandidate wired into evolution:

• Mature skills (100+ activations, 80%+ success) get fast-track recommendation

Code Quality — 9 review passes, 19 bugs fixed:

• 2 critical (scoping crash on /forge approve, 4 callReviewer missing apiFormat)
• 4 medium (HTML injection, channel detection divergence, watchdog regression, provider config)
• 13 low (orphaned imports, stale docs, notification formatting)

Stats

• 412/412 tests passing
• 23/23 adversarial mutations caught
• 5 active skills in production (exec-openclaw: 350+ activations, 100% success)
• Zero known bugs

Breaking Changes

None. Drop-in upgrade from v0.8.x.

Upgrade

cd ~/.openclaw/extensions/aceforge
git pull origin main
npm install
openclaw gateway restart

AceForge v0.8.2

Skill versioning, CI, dry-run mode, 10 bug fixes. See CHANGELOG.md.

AceForge v0.8.1 — Refactor & Cleanup

v0.8.1 — Codebase Refactor

Refactor

• analyze.ts split from 1,492 → 487 lines into 5 focused modules:
  • constants.ts — canonical blocklist source (eliminates drift)
  • analyze-utils.ts — shared filesystem helpers
  • analyze-native.ts — native tool domain clustering
  • analyze-chains.ts — workflow chain detection
  • analyze.ts — orchestrator only

Fix

• Deduplicated config loading — llm-judge.ts imports from llm-generator.ts

Stats

• 363/363 tests passing
• 23/23 adversarial mutations caught
• 0 open audit findings across 3 rounds
• 31 source files across 8 directories