feat(domain): 0.11.0 — unify skill-train under /domain; retire skillopt + skillopt-hook

.claude-plugin/marketplace.json

@@ -401,8 +401,8 @@
     {
       "name": "domain",
       "source": "./skills/domain",
-      "description": "0.10.1 (2026-06-03) `_is_name` now accepts a DIGIT-LEADING NAME — first char may be an UPPERCASE letter OR a digit 0-9, so `/domain init 8VERB` works (previously rejected: digit-leading start); subsequent chars unchanged (UPPERCASE/digit/`-`/`+`). 0.10.0 — bare `/domain` + `/domain list` now LINT for scattered report-like `.md` (a basename token audit|report|verdict|gap|finding|impl OR an ISO date like `2026-06-02`) sitting loose at the repo root or under `docs/` WHILE a domain is active, and emit a `⚠ 분리보관 감지` advisory naming each stray with its `/domain absorb <file>` runbook (excludes the domain's own pair / CHANGELOG / README / `.log.md` / already-absorbed pointer files; read-only). 0.9.0 — new `absorb <file> [--state]` verb enforces combined-storage (합산보관) over scattered docs/*.md: folds an existing separate report/finding `.md` into the ACTIVE domain's single doc pair (default → <NAME>.log.md as a dated `## <ISO> — absorbed <rel>` entry embedding the body; `--state` → <NAME>.md snapshot), then replaces the source with a one-line `<!-- absorbed into … -->` pointer (idempotent — an already-absorbed pointer file is skipped; refuses to absorb the domain's own pair). Resolves the active domain via the same _get_active / _snap_path / _log_path logic the other verbs use. Maintain UPPERCASE <NAME>.md (snapshot = final-goal milestone checkboxes) + sister <NAME>.log.md (append-only step log) at project root — auto-scaffolds both, defaults NAME to uppercase basename of git root. Verbs: /domain <NAME> (SELECT the session's active domain → later verbs default to it + show a final-goal progress bar `▓▓▓░░ NN% · done/total` from the snapshot's checkboxes) · /domain (show active) · /domain list (alias ls — repo-wide index; with a DOMAINS.tape roster it is AUTHORITATIVE: tables each registered domain ★ = active · @goal · progress · location, progress/@goal stay DERIVED so the checked-in roster holds only NAME→path and never churns; flags ghosts + unregistered; no roster → legacy disk-scan) · /domain list --sync (reconcile DOMAINS.tape with disk — bootstraps the roster) · /domain <task> (append `- [x]` to log) · /domain todo <task> (`- [ ]`) · /domain done <match> (flip [ ]→[x]) · /domain title <text> (alias subtitle — OPTIONAL `@title:` display header: icon · name · alias, e.g. `🧠 IIT4 — 의식 측정자(尺)`, rendered in place of plain `active domain: NAME`; absence keeps current output, no lint warning) · /domain new <header>. Progress is final-goal-based (snapshot completion), not log-based — it doesn't drift as the log grows. Path resolution: a DOMAINS.tape roster row (`@domain NAME := \"path\"`) wins FIRST, so a domain can live at ANY path (e.g. `domains/RUNTIME/RUNTIME.md`); absent a roster entry it falls back to root → folder-nested → root-default-scaffold (log follows the snapshot's dir), so folder-nested meta-domains no longer regenerate an empty root scaffold each call. The folder-nested fallback only matches when the nested file's first line is the domain header `# <NAME>` (guards against false-matching an unrelated same-named file like a digest on case-insensitive filesystems). 0.8.8: /domain set now SELF-HEALS the DOMAINS.tape roster (auto-registers an on-disk-but-unregistered domain — the gap that ghosted anima's LIFE) and warns on a stale SSOT shadow (the working-tree <NAME>.md is UNTRACKED or BEHIND origin/main → progress/closure may read a stale local copy; reference_domain_init_untracked_ssot); a perpetual `@goal` (no-termination marker like `종료 조건 없음`/`perpetual`/`open horizon`) renders a ♾️ perpetual badge so the progress bar reads as frontier-depletion, NOT a 100% completion countdown (per feedback-closure-is-physical-limit).",
-      "version": "0.10.1"
+      "description": "0.11.0 (2026-06-07) ML subverbs unify the retired `skillopt` + `skillopt-hook` plugins under `/domain`: `/domain doctor` (skill-train readiness) · `/domain train [--bg]` (NATIVE pip-free SkillOpt loop — rollout → reflect → edit → held-out gate via `claude -p` = your Claude Code subscription, NOT a metered API; only a SEPARATE skill.md is optimized — the domain's own DOMAIN.md / verdicts / log are NEVER touched) · `consume <skill.md>` · `activate`/`deactivate` (write ~/.sidecar/domain-skill/active-skill.md) · `status`/`log`. Bundled `examples/toyqa` (6 format-sensitive QA items) proves the loop end-to-end with NO external data/pip. ML verbs route to `bin/_domain_ml.sh` (+ `bin/_domain_ml.py` engine); ALL existing domain verbs route to `_domain.hexa` UNCHANGED. 0.10.1 (2026-06-03) `_is_name` now accepts a DIGIT-LEADING NAME — first char may be an UPPERCASE letter OR a digit 0-9, so `/domain init 8VERB` works (previously rejected: digit-leading start); subsequent chars unchanged (UPPERCASE/digit/`-`/`+`). 0.10.0 — bare `/domain` + `/domain list` now LINT for scattered report-like `.md` (a basename token audit|report|verdict|gap|finding|impl OR an ISO date like `2026-06-02`) sitting loose at the repo root or under `docs/` WHILE a domain is active, and emit a `⚠ 분리보관 감지` advisory naming each stray with its `/domain absorb <file>` runbook (excludes the domain's own pair / CHANGELOG / README / `.log.md` / already-absorbed pointer files; read-only). 0.9.0 — new `absorb <file> [--state]` verb enforces combined-storage (합산보관) over scattered docs/*.md: folds an existing separate report/finding `.md` into the ACTIVE domain's single doc pair (default → <NAME>.log.md as a dated `## <ISO> — absorbed <rel>` entry embedding the body; `--state` → <NAME>.md snapshot), then replaces the source with a one-line `<!-- absorbed into … -->` pointer (idempotent — an already-absorbed pointer file is skipped; refuses to absorb the domain's own pair). Resolves the active domain via the same _get_active / _snap_path / _log_path logic the other verbs use. Maintain UPPERCASE <NAME>.md (snapshot = final-goal milestone checkboxes) + sister <NAME>.log.md (append-only step log) at project root — auto-scaffolds both, defaults NAME to uppercase basename of git root. Verbs: /domain <NAME> (SELECT the session's active domain → later verbs default to it + show a final-goal progress bar `▓▓▓░░ NN% · done/total` from the snapshot's checkboxes) · /domain (show active) · /domain list (alias ls — repo-wide index; with a DOMAINS.tape roster it is AUTHORITATIVE: tables each registered domain ★ = active · @goal · progress · location, progress/@goal stay DERIVED so the checked-in roster holds only NAME→path and never churns; flags ghosts + unregistered; no roster → legacy disk-scan) · /domain list --sync (reconcile DOMAINS.tape with disk — bootstraps the roster) · /domain <task> (append `- [x]` to log) · /domain todo <task> (`- [ ]`) · /domain done <match> (flip [ ]→[x]) · /domain title <text> (alias subtitle — OPTIONAL `@title:` display header: icon · name · alias, e.g. `🧠 IIT4 — 의식 측정자(尺)`, rendered in place of plain `active domain: NAME`; absence keeps current output, no lint warning) · /domain new <header>. Progress is final-goal-based (snapshot completion), not log-based — it doesn't drift as the log grows. Path resolution: a DOMAINS.tape roster row (`@domain NAME := \"path\"`) wins FIRST, so a domain can live at ANY path (e.g. `domains/RUNTIME/RUNTIME.md`); absent a roster entry it falls back to root → folder-nested → root-default-scaffold (log follows the snapshot's dir), so folder-nested meta-domains no longer regenerate an empty root scaffold each call. The folder-nested fallback only matches when the nested file's first line is the domain header `# <NAME>` (guards against false-matching an unrelated same-named file like a digest on case-insensitive filesystems). 0.8.8: /domain set now SELF-HEALS the DOMAINS.tape roster (auto-registers an on-disk-but-unregistered domain — the gap that ghosted anima's LIFE) and warns on a stale SSOT shadow (the working-tree <NAME>.md is UNTRACKED or BEHIND origin/main → progress/closure may read a stale local copy; reference_domain_init_untracked_ssot); a perpetual `@goal` (no-termination marker like `종료 조건 없음`/`perpetual`/`open horizon`) renders a ♾️ perpetual badge so the progress bar reads as frontier-depletion, NOT a 100% completion countdown (per feedback-closure-is-physical-limit).",
+      "version": "0.11.0"
     },
     {
       "name": "domain-doc-guard",
@@ -434,18 +434,6 @@
       "description": "/gap — multi-axis gap exploration. 42 breakthrough-strategy lenses · 8 families (F4 + F6 each have 6 lenses; `occams-razor` lives in both — hypothesis side and design side). Bare /gap = inline-triage + deep-dive only hot families. /gap full = exhaustive 8-subagent fan-out. /gap <scope> targets the sweep. /gap list prints the catalogue. Surfaces + prioritises gaps; never fixes.",
       "version": "0.2.0"
     },
-    {
-      "name": "skillopt",
-      "source": "./commands/skillopt",
-      "description": "0.5.0 (2026-06-07) AUTO-USE (activate) — `/skillopt activate <skill.md>` copies a trained skill to the SSOT ~/.sidecar/skillopt/active-skill.md so the companion `skillopt-hook` auto-injects it at every SessionStart (no command needed); `/skillopt deactivate` removes it, `/skillopt agent-active on|off` toggles a one-line nudge that tells the agent to PROPOSE (never auto-run) `/skillopt train` for repeatable scored tasks. The USE-vs-TRAIN split: USE is automatic + opt-in + cheap; TRAIN stays a command/agent decision (cost-bearing). 0.4.0 (2026-06-07) BACKGROUND TRAIN + HARDER EXAMPLE — `/skillopt train --bg` detaches the run (nohup → log under ~/.sidecar/skillopt/), returns immediately; `/skillopt status` shows running-state + score/step progress, `/skillopt log` tails. The bundled `examples/toyqa` dataset swapped to 6 format-sensitive QA items (chemical symbols, ISO codes, rounding…) that an EMPTY skill answers in a full sentence → STRICT exact-match fails → a real learning gradient (the optimizer learns a 'reply with only the value' rule, then the held-out gate rises); train_size 6 · batch 3 · sel 5 for stronger signal. 0.3.0 (2026-06-07) EXECUTABLE CLI — `/skillopt` now runs directly via a `!`-exec dispatcher (`bin/skillopt.sh`, prefs-style: resolves its own cached install dir, so the user never types a long path). `/skillopt train` runs the loop in one token; `/skillopt doctor|ckpts|consume <skill.md>|help` all dispatch. Honest 0-edit runs reported as such; claude -p clarified = subscription (NOT metered). 0.2.0 (2026-06-07) sidecar-OWNED env adapter — the only domain code (run a task + score it) now ships IN the plugin at `commands/skillopt/examples/<domain>/`, NOT a clone of the upstream package; `bin/skillopt_run.py` injects sidecar adapters into the upstream hard-coded `_ENV_REGISTRY` at runtime (additive · survives `_register_builtins`) then runs the upstream trainer. Bundled reference `examples/toyqa/` = 4-item exact-match QA proving the loop runs end-to-end on local `claude -p` (no API key, no external data — target+optimizer both shell out to the Claude Code CLI = subscription, NOT metered); `examples/_base/default.yaml` vendors the upstream base config so examples are self-contained (plain `pip install skillopt` ships no `configs/`). Verified: baseline→rollout→reflect(on failures)→gate→test execute against real claude calls; an edit lands only when the optimizer judges a failure generalizable AND the held-out gate improves (no forced edit). 0.1.0 (2026-06-07) initial — /skillopt drives SkillOpt (microsoft/SkillOpt · `pip install skillopt`, arXiv:2605.23904), the text-space optimizer that trains a natural-language SKILL DOCUMENT for a frozen Claude Code agent via rollout → reflect → edit → held-out gate (DL analogy: skill.md = weights · rollout = forward · reflect = backprop · gate = validation early-stop). Subverbs — doctor (pip pkg + claude CLI + harness wiring readiness) · ckpts (list the package's bundled pretrained skill.md artifacts) · consume <skill.md> (load a trained skill into THIS session as additive system guidance) · train <config.yaml> (run the loop with Claude Code as the target harness — env TARGET_BACKEND=claude_code_exec · CLAUDE_SETTING_SOURCES=user,project so the sidecar tapes ride along as FIXED scaffolding; refuses to run without a real scoring env adapter — no fabricated scores) · help. HARD INVARIANT: only the skill.md is optimized — the model, the governance tapes (= SkillOpt's fixed prompts/*_system.md), and every *-guard safety hook stay UNCHANGED; the trained skill is a SEPARATE per-domain file injected via --append-system-prompt; /skillopt never edits a .tape or a guard (a held-out UTILITY gate is not a SAFETY check — kept orthogonal). Wraps the upstream pip CLI; does not vendor or fork it. Companion: microsoft/SkillLens (arXiv:2605.23899).",
-      "version": "0.5.0"
-    },
-    {
-      "name": "skillopt-hook",
-      "source": "./hooks/skillopt-hook",
-      "description": "0.1.0 (2026-06-07) SessionStart hook — auto-USES a trained skill without a command. If the user activated a learned skill (`/skillopt activate <skill.md>` → ~/.sidecar/skillopt/active-skill.md), injects it as additionalContext at session start so the agent applies it automatically (prefs-hook split: the command writes the SSOT, the hook auto-injects). OPT-IN + safe: silent when nothing is activated; NEVER trains (training is cost-bearing — stays a command/agent decision); always exits 0 (fail-open). Optional `~/.sidecar/skillopt/agent-active` marker adds a one-line nudge to PROPOSE (never auto-run) `/skillopt train` for repeatable auto-scorable tasks. Hook half of the skillopt USE-vs-TRAIN split; the `skillopt` command plugin is the TRAIN/activate half. NO env opt-out — the SSOT files are the switch.",
-      "version": "0.1.0"
-    },
     {
       "name": "sidecar",
       "source": "./commands/sidecar",

CHANGELOG.md

@@ -6,6 +6,26 @@ For the full audit trail, see `git log`.
 
 ---
 
+## 2026-06-07 — 🎓 domain 0.11.0 — `/domain train` 으로 통일 · skillopt + skillopt-hook 폐기
+
+스킬 학습 능력을 별도 플러그인 대신 **`/domain` 하나로 통일**. 네이티브 · pip 무의존.
+
+- 🎓 **domain 0.11.0** — ML 서브버브 추가: `/domain doctor`(학습 준비 점검) ·
+  `/domain train [--bg]`(네이티브 pip-free SkillOpt 루프 — rollout → reflect → edit →
+  held-out gate, 백엔드 `claude -p` = 구독, 과금 API 아님) · `consume <skill.md>` ·
+  `activate`/`deactivate`(`~/.sidecar/domain-skill/active-skill.md`) · `status`/`log`.
+  ML 서브버브는 `bin/_domain_ml.sh`(+ `bin/_domain_ml.py` ~120줄 엔진)로 라우팅,
+  기존 도메인 버브는 `_domain.hexa` 로 **그대로** 유지.
+- 🔒 **HARD INVARIANT** — 학습은 **별도 skill.md 만** 최적화. 도메인 자신의
+  `DOMAIN.md` / verdicts / 로그는 절대 자동 편집 안 함(B 분리).
+- 📦 번들 `examples/toyqa`(format-sensitive QA 6 + 5) — 외부 데이터/pip 없이
+  baseline→rollout→reflect→gate 루프가 end-to-end 도는 걸 증명.
+- 🗑 **skillopt(0.5.0) · skillopt-hook(0.1.0) 폐기** — 두 디렉터리 제거,
+  marketplace 엔트리 삭제. 모든 능력이 `/domain` 으로 흡수됨(USE=activate/hook 대체는
+  domain activate + SessionStart, TRAIN=domain train).
+- 검증: `/domain doctor` smoke(claude CLI + 네이티브 엔진 감지) · py/sh 문법 · JSON 유효 ·
+  g22 lockstep(plugin 0.11.0 ↔ marketplace 0.11.0) + CHANGELOG 동일 커밋.
+
 ## 2026-06-07 — 🎓 skillopt 0.5.0 + skillopt-hook 0.1.0 — 자동 사용(하이브리드)
 
 "명령어 안 치고 자동으로" + "에이전트 적극 활용"을 USE-vs-TRAIN 분리로 구현.