@uzysjung/agent-harness 26.96.0 → 26.98.0

package/templates/skills/harness-health-audit/SKILL.md

@@ -0,0 +1,370 @@
+---
+name: harness-health-audit
+description: >-
+  Audit the health of an AI-coding harness — the CLAUDE.md / AGENTS.md files, rules, skills, agents,
+  hooks, and commands that steer an agent in a repository — across three questions a linter cannot
+  answer: is it TRUE (does it match the real code?), is it USED (do skills actually trigger and does
+  the loop actually verify?), and is it AFFORDABLE (is it inside the budget where instructions are
+  still followed?). Then surgically correct or remove only what is proven wrong or dead. Use whenever
+  the steering layer may have rotted or may not be working: "하네스 점검해줘", "하네스 드리프트 감사",
+  "CLAUDE.md가 실제랑 맞는지 봐줘", "룰/스킬이 최신인지 확인해줘", "스킬이 제대로 활용되는지 봐줘",
+  "루프 엔지니어링 잘 되고 있는지 검토해줘", "죽은 훅/커맨드 정리해줘", "audit my harness",
+  "check the rules still match the real stack", "are my skills actually being used", "review the
+  agent loop", after a stack or tooling migration, or when onboarding a repo whose steering docs look
+  stale. Every finding must be backed by stated-versus-measured evidence; never guess the correct
+  value, and never remove a live path because it merely looks unused.
+---
+
+# Harness Health Audit
+
+A **harness** is everything that steers an agent but is not the model itself: `CLAUDE.md` /
+`AGENTS.md` at every level, `.claude/rules/*.md`, skills, subagents, hook wiring, slash commands, and
+tool/MCP descriptions — plus the `.codex` / `.opencode` equivalents when the repo targets several
+CLIs. LangChain frames it as `Agent = Model + Harness`, defining the harness as "every piece of code,
+configuration, and execution logic that isn't the model itself"
+([LangChain](https://www.langchain.com/blog/the-anatomy-of-an-agent-harness)); Addy Osmani puts the
+stakes plainly: "A decent model with a great harness beats a great model with a bad harness"
+([Osmani](https://addyosmani.com/blog/agent-harness-engineering/)).
+
+Harnesses fail in two different ways, and most audits only catch the first:
+
+- **They rot.** The stack migrates, a command is renamed, a plugin disappears upstream, a ceremony is
+  abandoned — but the steering text still asserts the old world. Nothing fails loudly, because the
+  harness is instructions, not code: no test turns red when a rule lies.
+- **They stop working while still being true.** Every word is accurate, but the skills never trigger,
+  the rules are too long to be followed, and the loop closes on the agent's own say-so instead of a
+  real check. A perfectly accurate harness that nothing obeys is still a broken harness.
+
+This audit targets both. It does not claim to catch everything — see the safety gap named below.
+
+## Run the deterministic linter first
+
+Do not spend model tokens on what a program can decide. Form-level checks — file length caps, secret
+detection, `.env` gitignoring, action SHA-pinning, resolvable links, naming conventions — are
+deterministic. Tools exist: [AgentLint](https://www.agentlint.app/) advertises 33 checks over five
+dimensions (findability, instructions, workability, continuity, safety) across `CLAUDE.md`,
+`AGENTS.md`, `.cursor/rules`, `.github/copilot-instructions.md`, CI workflows, pre-commit hooks and
+`.gitignore`; `cclint` covers syntax. Run what is available, then start here.
+
+> "Never send an LLM to do a linter's job. LLMs are comparably expensive and *incredibly* slow
+> compared to traditional linters and formatters."
+> — [HumanLayer](https://www.humanlayer.dev/blog/writing-a-good-claude-md)
+
+**This skill exists for what a linter structurally cannot decide.** A linter can measure that a rule
+is 40 lines long; it cannot know whether the stack it names is the stack this repo actually uses,
+whether the hook it wires still fires, or whether the skill it installs ever triggers. Those require
+reading the code and judging. That judgment — truth and efficacy — is this skill's whole job.
+
+## Governing principle
+
+**Keep the live skeleton, cut only the dead flesh.** Correction is surgical and evidence-based, not a
+rewrite. A harness that still works must survive the audit almost untouched. Two rules make that safe:
+
+- **Every finding needs stated-versus-measured evidence.** "Looks stale" is not evidence. Quote the
+  harness line, then quote the ground-truth source that contradicts it.
+- **Never delete a path because it looks unused.** Absence of obvious callers is not proof of death.
+  Measure it. If you cannot prove it is dead, keep it and flag it — silently stripping a live gate is
+  the one failure worse than drift.
+
+## The three questions
+
+Run them in this order: truth is cheapest, economy needs the whole picture.
+
+| | Question | Failure it catches |
+|---|---|---|
+| **A** | **Is it TRUE?** | The harness asserts a world the repo left behind |
+| **B** | **Is it USED?** | The harness is accurate but inert — nothing triggers, nothing verifies |
+| **C** | **Is it AFFORDABLE?** | The harness is accurate and well-designed but too big to be followed |
+
+**These three are not exhaustive, and the gap that matters most is safety.** A harness can be TRUE,
+USED, and AFFORDABLE and still be dangerous: a hook that pipes an unpinned remote script straight
+into a shell, or a rule that tells the agent to skip permission prompts, is accurate (it describes
+what it does), live (it fires), and cheap (one line) — so all three questions pass it. The
+deterministic linters cover *form* safety (secrets, `.env` hygiene, action SHA-pinning); what they
+cannot judge is whether a live, accurate instruction is a bad idea.
+
+So: **when a check reveals something dangerous, report it — do not discard it because it fits no
+question.** Treat safety as a standing lens over A/B/C rather than a fourth pass, and say plainly in
+the report that safety was judged ad hoc, not systematically. A dedicated safety axis is a known gap
+in this skill, not a solved problem.
+
+---
+
+## A. Is it TRUE? (accuracy)
+
+Drift is the delta between what the harness states and what the repo is. Measure the real state from
+load-bearing artifacts, never from memory or prose: dependencies and lockfile for the stack, the
+resolvable command set for namespaces, wiring files and recent history for what fires, upstream
+existence for advertised assets.
+
+### A1 — Stale-stack drift
+The harness mandates a stack the code no longer uses (a rules file requiring `pytest` when the
+lockfile shows `vitest`).
+**Measure:** diff each stack assertion against dependencies / lockfile / config files.
+**Action:** the stack section is load-bearing — **correct** it to the measured stack. Read the new
+value; never infer it. Removal applies only where the repo dropped the tooling outright.
+
+### A2 — Stranded command / namespace references
+References to commands or namespaces that were renamed or removed.
+**Measure:** resolve every referenced command against the actually-installed set.
+**Action:** **correct** if the target survives under a new name; **remove** if it is gone. Search
+**skill and agent bodies**, not just `CLAUDE.md` — stranded references hide inside asset bodies, which
+is where they are hardest to see and most often missed.
+
+### A3 — Dead ceremony
+Hooks, gates, agents, or checklists that no longer run: a hook wired to a deleted script, a gate
+abandoned mid-adoption, a status file frozen for months, an agent irrelevant to this project.
+**Measure:** check wiring and callers first, then whether the ceremony appears in recent history.
+**Action:** **remove** only when death is proven (no wiring, no callers, no recent fires, upstream
+gone). Dormant-but-unproven → **keep and flag**.
+
+### A4 — False advertising
+The harness promises assets or capabilities that no longer exist or no longer behave as claimed: a
+skill whose upstream vanished, a catalog count that overstates reality, a documented flag that
+crashes.
+**Measure:** confirm each advertised asset resolves and each documented capability runs. Advertised is
+not working.
+**Action:** bring advertised text and actual behaviour into exact agreement. An overstatement is drift
+even when nothing is "broken".
+
+---
+
+## B. Is it USED? (efficacy)
+
+A harness can be entirely true and still do nothing. These checks ask whether it actually reaches the
+agent's behaviour.
+
+### B1 — Skill and tool utilization
+Skills are selected from metadata alone: the agent "pre-loads the `name` and `description` of every
+installed skill into its system prompt", and loads the body only "if Claude thinks the skill is
+relevant to the current task"
+([Anthropic](https://www.anthropic.com/engineering/equipping-agents-for-the-real-world-with-agent-skills)).
+So a skill with a vague description is invisible no matter how good its body is — and Anthropic's own
+skill-creator guidance warns that Claude "has a tendency to *undertrigger* skills — to not use them
+when they'd be useful".
+
+**Measure:**
+- **Description triggerability** — does each description name concrete trigger phrases and contexts,
+  or does it say "when relevant"? Vague descriptions are the mechanism of undertriggering.
+- **Competition and collision** — do several skills claim overlapping triggers, or share a name with a
+  command? Overlap forces a selection the model can get wrong. Osmani's tool-design finding applies:
+  "Ten focused tools outperform fifty overlapping ones because the model can hold the menu in its
+  head" ([Osmani](https://addyosmani.com/blog/agent-harness-engineering/)).
+- **Sprawl** — count installed skills/tools. Every one spends startup context on metadata forever.
+- **Observed use** — where session transcripts exist, did the skill fire on tasks it should own?
+
+**Action:** **correct** vague descriptions toward concrete triggers and explicit negative cases
+("not for X, use Y"); **de-conflict** overlapping claims; flag sprawl for the user's decision — do not
+delete someone's skills to reduce a count.
+
+**Honest limit:** Anthropic's guidance offers no quantitative trigger metric — it recommends running
+agents "on representative tasks and observing where they struggle" and monitoring "how Claude uses
+your skill in real scenarios". So utilization findings are **observational, not measured**. Say so;
+do not invent a trigger rate.
+
+### B2 — Loop integrity
+The loop is the cycle of reason → act → observe → verify. What makes it work is not the model but the
+**quality of the feedback signal** it closes on.
+
+**Measure:**
+- **Is there a real verify step?** Does something deterministic (tests, typecheck, lint, build) run and
+  feed failures back — or does the loop end on the agent's self-report? The hook pattern to look for:
+  "Success is silent, failures are verbose. If typecheck passes, the agent hears nothing. If it fails,
+  the error text gets injected into the loop and the agent self-corrects"
+  ([Osmani](https://addyosmani.com/blog/agent-harness-engineering/)).
+- **Is generation separated from evaluation?** Osmani reports that Anthropic's long-running harness
+  work is explicit that "separating generation from evaluation into distinct agents outperforms
+  self-evaluation, because agents reliably skew positive when grading their own work"
+  ([Osmani](https://addyosmani.com/blog/agent-harness-engineering/)). A harness where the author is
+  the only reviewer has a known bias, not a check.
+- **Are there stopping conditions?** A loop must break on *no progress*, not only on completion.
+- **Is the checked proxy the real goal?** The canonical failure is an agent deleting a failing test to
+  turn CI green — the gate passes, the objective is lost.
+
+Known loop failure modes to look for: early stopping; hallucinated success (done without
+verification); compounding errors across a long trajectory; incoherence as work stretches across
+context windows; poor decomposition ([Osmani](https://addyosmani.com/blog/agent-harness-engineering/)).
+
+**Action:** **correct** loops that close on self-report by wiring a deterministic verifier; **add** a
+no-progress stop where a loop can spin; flag proxy-goal mismatches. Do not add gates speculatively —
+see the Ratchet tension below.
+
+### B3 — Rule adherence
+Presence is not obedience. HumanLayer states it directly: "Claude will ignore the contents of your
+`CLAUDE.md` if it decides that it is not relevant to its current task"
+([HumanLayer](https://www.humanlayer.dev/blog/writing-a-good-claude-md)).
+
+**Run this check only when transcripts or recent work are available to read.** Adherence cannot be
+measured from files alone — without that evidence the check has nothing to say, so skip it and record
+it as **unverified** rather than manufacturing a verdict. (Its one file-visible residue — "a rule a
+program could enforce belongs in the program" — is C2's job; leave it there rather than duplicating
+the finding here.)
+
+**Measure:** pick a few load-bearing rules and check whether the work actually followed them. A rule
+routinely violated without consequence is either mis-placed or dead.
+**Action:** **flag** rules that are stated but never observed to bind. Route mis-placed ones to C2.
+
+---
+
+## C. Is it AFFORDABLE? (economy)
+
+Instructions compete for a finite attention budget. Past a point, adding correct text makes the
+harness *worse*.
+
+### C1 — Instruction budget
+The evidence:
+- **Context rot.** Chroma tested 18 frontier models (Claude, GPT, Gemini, Qwen families) and found
+  "model performance degrades as input length increases, often in surprising and non-uniform ways" —
+  even on trivially simple tasks where only length changed. Distractors compound it: "even a single
+  distractor reduces performance relative to the baseline, and adding four distractors compounds this
+  degradation further" ([Chroma](https://www.trychroma.com/research/context-rot)).
+- **Position matters.** Liu et al. found a U-shaped curve — performance is highest when relevant
+  information sits at the very start or end of the context and degrades in the middle — and it holds
+  even for the strongest model tested ([Liu et al., TACL](https://aclanthology.org/2024.tacl-1.9/)).
+- **Length norms.** Reported consensus is "less than 300 lines is best, and shorter is even better";
+  HumanLayer's own root `CLAUDE.md` is "less than sixty lines"
+  ([HumanLayer](https://www.humanlayer.dev/blog/writing-a-good-claude-md)). Osmani's framing: a
+  "pilot's checklist, not style guide" ([Osmani](https://addyosmani.com/blog/agent-harness-engineering/)).
+
+**Measure:** length of each always-loaded file; how many discrete instructions the harness asserts in
+total across every always-loaded layer (global + project + directory); how much is irrelevant to a
+typical task (those are the distractors).
+**Action:** **correct** by moving detail behind pointers — state what is true every session, link the
+rest. Prefer removing duplication over shortening prose.
+**Honest limit:** Chroma publishes performance *curves*, not a percentage-per-token figure. Do not
+quote a precision the research does not contain. Line norms are practitioner consensus, not a measured
+threshold — treat them as a smell test, not a gate.
+
+### C2 — Right tool for the job
+Instructions asking the model to do what a program should do are both expensive and unreliable.
+**Measure:** scan the harness for rules a linter, formatter, hook, or CI check could enforce
+deterministically (formatting, import order, banned calls, secret hygiene).
+**Action:** **move** them into the deterministic layer and delete the prose. This buys budget back for
+the judgment-level guidance only the model can use.
+
+---
+
+## The Ratchet test (cross-cutting)
+
+Osmani proposes: "Every line in a good `AGENTS.md` should be traceable back to a specific thing that
+went wrong" — earn each line, add constraints only after observing real failures
+([Osmani](https://addyosmani.com/blog/agent-harness-engineering/)).
+
+This is a sharp audit lens: for each line, ask *what failure does this prevent?* If nobody can name
+one, it is speculative bloat paying rent in the budget from C1.
+
+**It also conflicts with this skill's removal rule, and the conflict must not be blurred.** The
+Ratchet is a rule for **adding** — it justifies refusing new lines. This skill's rule governs
+**removing** — proof of death before deletion. An untraceable line is *not* proven dead; it may be the
+only thing preventing a failure nobody has hit recently.
+
+**Resolution:** an untraceable line is a **flag**, never an auto-delete. Surface it, name the missing
+rationale, and let the user decide. Apply the Ratchet at full strength to anything the audit proposes
+to *add*.
+
+## Classification
+
+Route every finding through one decision:
+
+| Decision | When | Guard |
+|---|---|---|
+| **Correct** | Still used, stated value wrong | Use the measured value, never a guess |
+| **Remove** | Provably dead: no wiring, no callers, no fires, upstream gone | Proof required; name it |
+| **Move** | Right intent, wrong layer (a linter's job stated as prose) | Verify the new layer actually fires |
+| **Keep (flag)** | Looks stale or untraceable, but not proven wrong or dead | Surface it; do not touch it |
+
+When torn between remove and keep, keep. A false "keep" costs a little noise; a false "remove"
+silently strips a live gate.
+
+## Workflow
+
+1. **Run the deterministic linter** if one is available. Do not re-do its work by hand.
+2. **Inventory the surface.** Every always-loaded file, rule, skill, agent, hook + its wiring, command,
+   and the multi-CLI equivalents. Missing a surface is how a rename gets fixed in `CLAUDE.md` and left
+   broken inside a skill body.
+3. **Measure ground truth.** Real stack (dependencies/lockfile), real commands (what resolves), real
+   wiring (what fires), real assets (what exists upstream).
+4. **Run A → B → C.** For each check, produce stated/measured pairs. No pair, no finding.
+5. **Classify** each finding: correct / remove / move / keep-flag. For every "remove", write the death
+   evidence explicitly.
+6. **Apply surgically.** Only drifted lines. Do not "improve" adjacent, correct text. Match existing
+   conventions. Keep mechanical renames in a separate change from semantic corrections — they carry
+   different risk and review differently.
+7. **Verify nothing live broke.** Re-check callers and wiring after editing; confirm no hook, command,
+   or gate that still fires was removed.
+8. **Report** the table below.
+
+## Report format
+
+```markdown
+## Harness Health Audit — <repo> @ <commit>
+Deterministic linter: <tool + score, or "none available">
+
+| # | Q | Check | Stated | Measured | Action | Evidence |
+|---|---|-------|--------|----------|--------|----------|
+| 1 | A | stale-stack | "pytest for all tests" | lockfile → vitest | correct | package.json |
+| 2 | A | stranded ref | body calls `/oldns:review` | resolves as `agent-skills:review` | correct | skill body |
+| 3 | A | dead ceremony | 6-gate pre-commit hook | not wired, 0 fires/3mo | remove | settings.json, git log |
+| 4 | B | utilization | skill desc: "when relevant" | no concrete trigger | correct | SKILL.md frontmatter |
+| 5 | B | loop integrity | "verify before commit" | no verifier wired; self-report only | correct | hook wiring |
+| 6 | C | budget | root CLAUDE.md 700 lines | norm <300 | correct | wc -l |
+
+**Unverified (could not measure):** <list, or "none">
+**Kept-and-flagged (not proven dead/wrong):** <list, or "none">
+**Safety concerns (ad-hoc lens, not a systematic pass):** <list, or "none seen — not a clean bill">
+```
+
+Always print the *Unverified*, *Kept-and-flagged*, and *Safety* rows even when empty. A clean table
+that hides a blind spot is itself a drift. The safety row's empty value is "none seen", never "none"
+— the audit has no systematic safety pass, so silence there is absence of evidence, not evidence of
+absence.
+
+## Anti-patterns
+
+- **Rewriting the whole harness.** "Improving" adjacent, correct text buries the real corrections.
+- **Guessing the correct value.** Correcting a stack assertion from memory replaces old drift with new.
+- **Fixing only `CLAUDE.md`.** A rename that misses skill and agent *bodies* strands the references.
+- **Deleting on "looks unused".** Removing a gate without proving death can strip a live safety check.
+- **Treating advertised as working.** A promised asset that no longer resolves is drift, not a footnote.
+- **Reporting a trigger rate you did not measure.** Utilization is observational; invented metrics are
+  worse than an honest "unverified".
+- **Doing a linter's job by hand** — and leaving prose in the harness that should have been a hook.
+- **Adding gates because gates feel safe.** Every speculative gate is future dead ceremony (A3) paid
+  for out of the budget (C1). Earn it with a real failure.
+- **One unreviewable diff.** Blending mechanical renames with semantic edits defeats verification.
+
+## Honest limitations
+
+- **No systematic safety pass.** A/B/C can all pass on a harness that is accurate, live, cheap, and
+  unsafe (an unpinned `curl | sh` hook; a rule waiving permission prompts). Safety is a lens applied
+  ad hoc, and the linters only cover form-level safety — so a clean audit is not a safety clearance.
+  Say so in the report.
+- **Presence ≠ adherence.** A correct rule can still be ignored; the model decides relevance. Only
+  observation reveals whether a rule binds — files alone cannot.
+- **No published utilization metric.** There is no authoritative trigger-rate measure to cite; report
+  utilization qualitatively.
+- **Gitignored harness.** Much of a harness can live in gitignored files (locally installed skills,
+  machine-specific settings). Git history then cannot reconstruct the live workflow — read the actual
+  files, and state plainly when a surface could not be measured.
+- **Unmeasurable live-fire.** Whether a hook *actually* fired in past sessions is usually not
+  recoverable from the repo alone. "Absent from history" is a signal, not proof; confirm with wiring.
+- **Cross-CLI asymmetry.** A harness can be correct for one target CLI and drifted for another. Audit
+  each surface it claims to support; report per-CLI when they diverge.
+- **Norms are not thresholds.** The line-count and tool-count figures here are practitioner consensus.
+  Use them to start a conversation, not to fail a build.
+
+## References
+
+- [LangChain — The Anatomy of an Agent Harness](https://www.langchain.com/blog/the-anatomy-of-an-agent-harness) — `Agent = Model + Harness`; harness components.
+- [Addy Osmani — Agent Harness Engineering](https://addyosmani.com/blog/agent-harness-engineering/) — loop failure modes, the Ratchet Principle, hook pattern, evaluator separation.
+- [Anthropic — Equipping agents for the real world with Agent Skills](https://www.anthropic.com/engineering/equipping-agents-for-the-real-world-with-agent-skills) — metadata-based triggering; progressive disclosure.
+- [HumanLayer — Writing a good CLAUDE.md](https://www.humanlayer.dev/blog/writing-a-good-claude-md) — length norms; "never send an LLM to do a linter's job"; rules can be ignored.
+- [Chroma — Context Rot](https://www.trychroma.com/research/context-rot) — 18 models; degradation with input length; distractor effects.
+- [Liu et al. — Lost in the Middle (TACL)](https://aclanthology.org/2024.tacl-1.9/) — U-shaped position curve.
+- [AgentLint](https://www.agentlint.app/) — deterministic harness linter; run it before this skill.
+
+## Related skills
+
+- **model-orchestration** — picks the model for the audit versus the corrections.
+- **compaction-handoff** — persists findings so a long audit survives a context boundary.
+- **no-false-ship (rule)** — check A4 is the same advertised-equals-actual discipline.

package/dist/chunk-WKXYOTUV.js.map

@@ -1 +0,0 @@
-{"version":3,"sources":["../node_modules/tsup/assets/esm_shims.js","../src/external-assets.ts","../src/track-match.ts","../src/categories.ts"],"sourcesContent":["// Shim globals in esm bundle\nimport path from 'node:path'\nimport { fileURLToPath } from 'node:url'\n\nconst getFilename = () => fileURLToPath(import.meta.url)\nconst getDirname = () => path.dirname(getFilename())\n\nexport const __dirname = /* @__PURE__ */ getDirname()\nexport const __filename = /* @__PURE__ */ getFilename()\n","/**\n * External asset matrix — bash setup-harness.sh L791~1067 등가 데이터.\n *\n * SPEC: docs/specs/cli-rewrite-completeness.md F3\n * Source: setup-harness.sh@911c246~1 (v27.18 직전, bash cutover 전)\n *\n * Track 또는 옵션 조건이 충족되면 install pipeline에서 method를 호출.\n * 실패는 \"warn-skip\" — 종료 시 누락 자산 보고 (OQ1 결정).\n *\n * code-style 800줄 cap 예외: 본 파일은 ~85% 가 큐레이션 자산 카탈로그(43 entry × ~13줄)\n *   = 데이터. 로직(interface·condition 평가·shouldInstallAsset)은 소량. 데이터/로직 분리는\n *   별도 사이클(Phase R) 후보 — 현재는 카탈로그 단일 SSOT 가독성 우선. (v26.79.0 기준 802줄)\n */\n\nimport type { Category, Source } from \"./categories.js\";\nimport { hasDevTrack } from \"./track-match.js\";\nimport type { OptionFlags, Track } from \"./types.js\";\n\nexport type ExternalAssetMethod =\n  /** `npx skills add <source>[ --skill <name>] --yes` */\n  | { kind: \"skill\"; source: string; skill?: string }\n  /** `claude plugin marketplace add <marketplace>` + `claude plugin install <pluginId>` */\n  | { kind: \"plugin\"; marketplace: string; pluginId: string }\n  /**\n   * `npm install --save-dev <pkg>@<version>` (global scope 시 `-g`).\n   * v26.80.0 — `version` 필수 (pinning). vetting 은 시점 검증인데 unpinned 는 미래 코드\n   * 실행 = hijacked vetted repo 직행 구멍 (ADR-021 \"지속 검증\" 주장과 모순). pkg 는 bare\n   * 이름 유지 (detectVersion 이 `<npm root>/<pkg>/package.json` 경로로 사용). bump 정책:\n   * A2 자산 audit 주기에 Docker 검증 후 갱신 (docs/COMPATIBILITY.md §pinning).\n   */\n  | { kind: \"npm\"; pkg: string; version: string }\n  /** `npx <cmd>@<version> [args...]` — fire-and-forget 실행. v26.80.0 — version 필수 (위와 동일 근거). */\n  | { kind: \"npx-run\"; cmd: string; version: string; args?: string[] }\n  /** `bash <script> <args...>` — 로컬 스크립트 (예: prune-ecc.sh) */\n  | { kind: \"shell-script\"; script: string; args: string[] }\n  /**\n   * v26.81.0 (ADR-022) — 내부 템플릿 자산. external-installer 가 spawn 하지 않음 —\n   * installer Phase 1 의 manifest/transform 게이팅이 `isAssetSelected(key)` 로 읽는다.\n   * (이전 OptionFlags.withTauri/withUzysHarness 자리. wizard/CLI 표면은 일반 자산과 동일)\n   */\n  | {\n      kind: \"internal\";\n      key:\n        | \"tauri-desktop\"\n        // v26.87.0 — dev-method skills (uzys 1st-party, repo-bundled templates).\n        | \"multi-persona-review\"\n        | \"gap-analysis-e2e\"\n        | \"ultracode-service-audit\"\n        | \"asis-tobe-decision\"\n        | \"compaction-handoff\"\n        | \"northstar-roadmap\"\n        // v26.93.0 — Orchestration & Model Policy (사용자 확정 2026-07-04) 스킬화.\n        | \"model-orchestration\"\n        // v26.95.0 — gemini-consult (opt-in, NOT dev-method): Gemini advisor via Antigravity agy.\n        | \"gemini-consult\";\n    };\n\nexport type ExternalAssetCondition =\n  /** Track 중 1개 이상이 set와 일치 */\n  | { kind: \"any-track\"; tracks: Track[] }\n  /** dev track (executive 외 모두) */\n  | { kind: \"has-dev-track\" }\n  /** OptionFlags 의 특정 플래그 true (잔존 동작 옵션용 — ADR-022 후 자산 토글엔 사용 금지) */\n  | { kind: \"option\"; flag: keyof OptionFlags }\n  /**\n   * v26.81.0 (ADR-022) — 순수 opt-in: condition 매치 항상 false. wizard 체크 또는\n   * `--with <id>` 의 forceInclude 로만 설치. 이전의 자산 1:1 OptionFlags(`withBmad` 등\n   * 13종)를 대체 — 자산 추가 시 플래그 코드 0곳.\n   */\n  | { kind: \"opt-in\" };\n\nexport interface ExternalAsset {\n  /** 안정 식별자 — 로깅 + 누락 보고 + 테스트에서 사용 */\n  id: string;\n  /** 사람이 읽는 라벨 (한 줄) */\n  description: string;\n  /** v26.43.0 — Category-based UI 그룹화. SPEC §3.1. */\n  category: Category;\n  /** v26.43.0 — 정확한 출처 (GitHub org/user). Step 2 라벨. SPEC §3.5 R6. */\n  source: Source;\n  condition: ExternalAssetCondition;\n  method: ExternalAssetMethod;\n  /**\n   * v26.79.0 — 검증 Trust Tier. 자산 entry 의 **필수 필드** (SSOT). 이전엔 별도 `TRUST_TIER`\n   * Record 였으나 누락(컴파일러 미검출) + stale(좀비 키, v26.76.0 content-creator 제거 전례)\n   * drift 가 가능했음 → entry 에 통합: **누락은 컴파일 에러, stale 은 구조적 불가능**.\n   * `TRUST_TIER` / `assetTrustTier` 는 이 필드에서 derive (EXTERNAL_ASSETS 정의 뒤).\n   * star snapshot(2026-05~06)은 각 entry tier 라인 주석. 실 drift 판정은 trust-tier-drift 가 live fetch.\n   */\n  tier: TrustTier;\n}\n\n/**\n * v26.71.0 (PRD v26-71) — 검증 Trust Tier. North Star 세 기둥 ②.\n *   - official: Anthropic 공식 marketplace + 본 하네스 자체.\n *   - vetted: star ≥ 1000 + 활성 유지보수 (D2; 라이선스 미표기/비-OSI 는 출처 신뢰).\n *   - experimental: star < 1000 — opt-in + 경고 (pre-check 제외).\n */\nexport type TrustTier = \"official\" | \"vetted\" | \"experimental\";\n\nconst ALL_CSR_SSR_FULL: Track[] = [\n  \"csr-supabase\",\n  \"csr-fastify\",\n  \"csr-fastapi\",\n  \"ssr-htmx\",\n  \"ssr-nextjs\",\n  \"full\",\n];\n\n/** csr-*|ssr-nextjs|full per bash setup-harness.sh L1041 (ssr-htmx 제외 — htmx는 React 미사용). */\nconst CSR_SSR_NEXTJS_FULL: Track[] = [\n  \"csr-supabase\",\n  \"csr-fastify\",\n  \"csr-fastapi\",\n  \"ssr-nextjs\",\n  \"full\",\n];\n\nconst RAILWAY_TRACKS: Track[] = [\"csr-fastify\", \"csr-fastapi\", \"ssr-htmx\", \"ssr-nextjs\", \"full\"];\n\n/**\n * v0.8.1 — executive-style Track SSOT (reviewer MEDIUM-3 fix).\n *\n * 3 Track 모두 dev/UI baseline 미적용 — `.claude/agents/strategist` + project-claude/<track>.md만.\n * `track-match.ts:hasDevTrack()` 의 negation domain. 사용처:\n *   - `shouldInstallAsset` `has-dev-track` 분기 코멘트 (L458)\n *   - `tests/external-assets.test.ts` invariant\n *\n * 신규 executive-style Track 추가 시 이 상수만 수정 → 모든 사용처 자동 반영.\n */\nexport const EXECUTIVE_STYLE_TRACKS: ReadonlyArray<Track> = [\n  \"executive\",\n  \"project-management\",\n  \"growth-marketing\",\n];\n\n/**\n * v0.8.1 — `hasDevTrack` SSOT 의 array 표현 (reviewer MEDIUM-3 fix).\n *\n * `track-match.ts:hasDevTrack()` 와 동등 (TRACKS \\ EXECUTIVE_STYLE_TRACKS = 8 Track).\n * `any-track` condition 에 dev set 전체를 인라인하지 않도록 사용.\n */\nexport const DEV_TRACKS: ReadonlyArray<Track> = [\n  \"csr-supabase\",\n  \"csr-fastify\",\n  \"csr-fastapi\",\n  \"ssr-htmx\",\n  \"ssr-nextjs\",\n  \"data\",\n  \"tooling\",\n  \"full\",\n];\n\n/**\n * v0.8.1 — dev + project-management 합집합 (reviewer MEDIUM-3 fix).\n *\n * `product-skills` (PM 도메인까지 사용) 의 9-Track 인라인 배열을 SSOT 상수로 교체.\n */\nexport const DEV_PLUS_PM_TRACKS: ReadonlyArray<Track> = [...DEV_TRACKS, \"project-management\"];\n\n/**\n * 61 자산 매트릭스 (v26.95.0 gemini-consult opt-in internal + v26.93.0 model-orchestration internal + v26.92.0 frontend-design official + v26.91.0 marketingskills opt-in + v26.87.0 dev-method skills 6종 internal + v26.86.0 Visual & Media 프레젠테이션 4종 + v26.85.0 5종 + v26.81.0 internal 2종 — ADR-022). bash setup-harness.sh@911c246~1 L791~1067 + 1320~1370 동등.\n *\n * 호출 순서: data → dev-baseline → railway → supabase-cli → impeccable → dev-tools →\n * supabase-skills → react/ui → next → executive → GSD → ToB → ECC.\n */\nexport const EXTERNAL_ASSETS: ReadonlyArray<ExternalAsset> = [\n  // === data Track ===\n  {\n    id: \"polars-K-Dense\",\n    tier: \"vetted\", // K-Dense-AI 26k\n    description: \"Polars — fast Rust-based DataFrame (pandas alternative, data track)\",\n    category: \"data\",\n    source: \"K-Dense-AI\",\n    condition: { kind: \"any-track\", tracks: [\"data\", \"full\"] },\n    method: { kind: \"skill\", source: \"K-Dense-AI/scientific-agent-skills\", skill: \"polars\" },\n    // v26.56.0 — description 보강: 트랙 hint + 한 줄 의미\n  },\n  {\n    id: \"dask-K-Dense\",\n    tier: \"vetted\", // K-Dense-AI 26k\n    description: \"Dask — distributed processing (large DataFrames · cluster, data track)\",\n    category: \"data\",\n    source: \"K-Dense-AI\",\n    condition: { kind: \"any-track\", tracks: [\"data\", \"full\"] },\n    method: { kind: \"skill\", source: \"K-Dense-AI/scientific-agent-skills\", skill: \"dask\" },\n  },\n  {\n    id: \"python-resource-management\",\n    tier: \"vetted\", // wshobson/agents 36k\n    description: \"Python memory · CPU management patterns (wshobson, data track)\",\n    category: \"data\",\n    source: \"wshobson\",\n    condition: { kind: \"any-track\", tracks: [\"data\", \"full\"] },\n    method: {\n      kind: \"skill\",\n      source: \"https://github.com/wshobson/agents\",\n      skill: \"python-resource-management\",\n    },\n  },\n  {\n    id: \"python-performance-optimization\",\n    tier: \"vetted\", // wshobson/agents 36k\n    description: \"Python performance optimization (profiling · vectorize, wshobson, data track)\",\n    category: \"data\",\n    source: \"wshobson\",\n    condition: { kind: \"any-track\", tracks: [\"data\", \"full\"] },\n    method: {\n      kind: \"skill\",\n      source: \"https://github.com/wshobson/agents\",\n      skill: \"python-performance-optimization\",\n    },\n  },\n  {\n    id: \"anthropic-data-plugin\",\n    tier: \"official\", // anthropics/knowledge-work-plugins 18k\n    description: \"Anthropic data plugin (visualization, SQL exploration)\",\n    category: \"data\",\n    source: \"anthropics\",\n    condition: { kind: \"any-track\", tracks: [\"data\", \"full\"] },\n    method: {\n      kind: \"plugin\",\n      marketplace: \"anthropics/knowledge-work-plugins\",\n      pluginId: \"data@knowledge-work-plugins\",\n    },\n  },\n\n  // === Internal template assets (v26.81.0, ADR-022 — 이전 OptionFlags.withTauri/withUzysHarness) ===\n  {\n    id: \"tauri-desktop\",\n    tier: \"official\", // uzys 본 하네스 자체 템플릿\n    description: \"Tauri desktop rule template (CSR/full tracks — manifest rule mapping)\",\n    category: \"frontend\",\n    source: \"uzys\",\n    condition: { kind: \"opt-in\" },\n    method: { kind: \"internal\", key: \"tauri-desktop\" },\n  },\n\n  // === Dev-method skills (uzys 1st-party, v26.87.0) ===\n  // 본 하네스의 작업 방법론 skill 6종 (repo-bundled templates). tier official, core on dev tracks\n  // (has-dev-track → 기본 설치; wizard uncheck / --without <id> 로 제외 가능 — isAssetSelected 게이팅).\n  {\n    id: \"multi-persona-review\",\n    tier: \"official\", // uzys 본 하네스 자체 템플릿\n    description:\n      \"Multi-persona review — critique one artifact via 3-5 parallel user personas, then synthesize P0/P1/P2 fixes\",\n    category: \"dev-tools\",\n    source: \"uzys\",\n    condition: { kind: \"has-dev-track\" },\n    method: { kind: \"internal\", key: \"multi-persona-review\" },\n  },\n  {\n    id: \"gap-analysis-e2e\",\n    tier: \"official\", // uzys 본 하네스 자체 템플릿\n    description:\n      \"Gap analysis E2E — detect north-star / correctness / UX gaps, then benchmark how reference services solved each\",\n    category: \"dev-tools\",\n    source: \"uzys\",\n    condition: { kind: \"has-dev-track\" },\n    method: { kind: \"internal\", key: \"gap-analysis-e2e\" },\n  },\n  {\n    id: \"ultracode-service-audit\",\n    tier: \"official\", // uzys 본 하네스 자체 템플릿\n    description:\n      \"Ultracode service audit — multi-agent, adversarially-verified full-service audit across 7 dimensions → milestone roadmap\",\n    category: \"dev-tools\",\n    source: \"uzys\",\n    condition: { kind: \"has-dev-track\" },\n    method: { kind: \"internal\", key: \"ultracode-service-audit\" },\n  },\n  {\n    id: \"asis-tobe-decision\",\n    tier: \"official\", // uzys 본 하네스 자체 템플릿\n    description:\n      \"ASIS→TOBE decision — present an A-or-B / approval moment as context → recommendation → option table → AS-IS/TO-BE contrast\",\n    category: \"workflow\",\n    source: \"uzys\",\n    condition: { kind: \"has-dev-track\" },\n    method: { kind: \"internal\", key: \"asis-tobe-decision\" },\n  },\n  {\n    id: \"compaction-handoff\",\n    tier: \"official\", // uzys 본 하네스 자체 템플릿\n    description:\n      \"Compaction handoff — persist durable state + git snapshot + resume anchor before a context /compact so nothing is lost\",\n    category: \"workflow\",\n    source: \"uzys\",\n    condition: { kind: \"has-dev-track\" },\n    method: { kind: \"internal\", key: \"compaction-handoff\" },\n  },\n  {\n    id: \"northstar-roadmap\",\n    tier: \"official\", // uzys 본 하네스 자체 템플릿\n    description:\n      \"North-star roadmap — measure current state vs the vision doc, then propose a ranked feature backlog persisted to docs/plans + memory\",\n    category: \"workflow\",\n    source: \"uzys\",\n    condition: { kind: \"has-dev-track\" },\n    method: { kind: \"internal\", key: \"northstar-roadmap\" },\n  },\n  // v26.93.0 — 사용자 확정(2026-07-04) Orchestration & Model Policy 스킬화. v26.94.0 개정\n  //   (2026-07-07): 역할분담 재편 — orchestrator 직접: 방향성·스펙리뷰(multi-persona-review)·\n  //   기능개선·성능/보안 문제발굴 / opus@xhigh+: 문서작성·핵심구현·V&V(fresh instance) /\n  //   sonnet@high+: 반복구현·E2E. effort floor 강제 3경로 + quota 핸드오프는 유지.\n  {\n    id: \"model-orchestration\",\n    tier: \"official\", // uzys 본 하네스 자체 템플릿\n    description:\n      \"Model orchestration policy — role split (orchestrator directs/reviews · opus@xhigh+ authors docs/core impl/V&V · sonnet@high+ repetitive impl/E2E) + effort floors + delegation spec + quota handoff\",\n    category: \"workflow\",\n    source: \"uzys\",\n    condition: { kind: \"has-dev-track\" },\n    method: { kind: \"internal\", key: \"model-orchestration\" },\n  },\n\n  // === Opt-in internal bundled skill (v26.95.0 — NOT dev-method) ===\n  // gemini-consult: uzys 1st-party skill wrapping Antigravity's `agy` CLI for natural Korean\n  //   phrasing + multi-persona second-opinion review. Bundled like dev-method skills\n  //   (templates/skills/gemini-consult/) so it renders across all 4 CLIs, but opt-in (condition\n  //   opt-in) — installed only on wizard check / `--with gemini-consult`. Ships a bash wrapper\n  //   (scripts/gemini-ask.sh) to Claude scope via the dir copy; non-Claude CLIs get the SKILL.md\n  //   which degrades to a direct `agy` call (graceful — no broken wrapper reference). tier\n  //   official (repo template); runtime dep on the external `agy` binary is a prereq, not a source.\n  {\n    id: \"gemini-consult\",\n    tier: \"official\", // uzys 본 하네스 자체 템플릿 (런타임 의존 agy 는 사용자 prereq)\n    description:\n      \"gemini-consult — consult Gemini (via Antigravity agy CLI) for natural Korean phrasing + multi-persona second-opinion review (opt-in; requires agy)\",\n    category: \"dev-tools\",\n    source: \"uzys\",\n    condition: { kind: \"opt-in\" },\n    method: { kind: \"internal\", key: \"gemini-consult\" },\n  },\n\n  // === Option-gated (v26.42.0 — opt-in, BREAKING vs prior has-dev-track auto-install) ===\n  {\n    id: \"addy-agent-skills\",\n    tier: \"vetted\", // addyosmani 47k\n    description: \"addy agent-skills (general dev)\",\n    category: \"workflow\",\n    source: \"addyosmani\",\n    condition: { kind: \"opt-in\" },\n    method: {\n      kind: \"plugin\",\n      marketplace: \"addyosmani/agent-skills\",\n      pluginId: \"agent-skills@addy-agent-skills\",\n    },\n  },\n  {\n    id: \"superpowers\",\n    tier: \"official\", // anthropics/claude-plugins-official 공식 배포 (소스 obra 213k)\n    // 저자 = obra (190k★ github.com/obra/superpowers). 호스팅 = Anthropic 공식\n    // marketplace github.com/anthropics/claude-plugins-official (\"Official,\n    // Anthropic-managed directory of high quality Claude Code Plugins\").\n    // source/marketplace 분리는 의도적 — source=저자, marketplace=registry.\n    description: \"Superpowers — agentic skills framework (obra, Anthropic official marketplace)\",\n    category: \"workflow\",\n    source: \"obra\",\n    condition: { kind: \"opt-in\" },\n    method: {\n      kind: \"plugin\",\n      marketplace: \"anthropics/claude-plugins-official\",\n      pluginId: \"superpowers@claude-plugins-official\",\n    },\n  },\n  {\n    // v26.75.0 (ADR-021) — wshobson/agents marketplace.json name = \"claude-code-workflows\"\n    // (84 plugins). 대표 = full-stack-orchestration. 다른 orchestrator(agent-orchestration/\n    // tdd-workflows/ship-mate 등): `claude plugin install <name>@claude-code-workflows`.\n    id: \"wshobson-agents\",\n    tier: \"vetted\", // wshobson/agents 36k\n    description: \"wshobson agents — multi-agent orchestration workflows (full-stack/tdd/review)\",\n    category: \"workflow\",\n    source: \"wshobson\",\n    condition: { kind: \"opt-in\" },\n    method: {\n      kind: \"plugin\",\n      marketplace: \"wshobson/agents\",\n      pluginId: \"full-stack-orchestration@claude-code-workflows\",\n    },\n  },\n  {\n    // v26.75.0 (ADR-021) — `npm i --save-dev @fission-ai/openspec` 후 `openspec init` 로 슬래시 주입.\n    id: \"openspec\",\n    tier: \"vetted\", // Fission-AI/OpenSpec 53k\n    description: \"OpenSpec — spec-driven brownfield delta workflow (propose → apply → archive)\",\n    category: \"workflow\",\n    source: \"fission-ai\",\n    condition: { kind: \"opt-in\" },\n    method: { kind: \"npm\", pkg: \"@fission-ai/openspec\", version: \"1.4.1\" },\n  },\n  {\n    // v26.75.0 (ADR-021) — 비대화형 install. v26.75.1: `--directory .` 누락 시 \"Installation\n    // directory\" 프롬프트에서 hang (Docker realcli 검출). cwd(=project) 기준 `.` 지정으로 봉합.\n    id: \"bmad-method\",\n    tier: \"vetted\", // bmad-code-org/BMAD-METHOD 48k\n    description: \"BMAD-METHOD — multi-agent agile workflow (PM/Architect/Dev, 12+ agents)\",\n    category: \"workflow\",\n    source: \"bmad-code-org\",\n    condition: { kind: \"opt-in\" },\n    method: {\n      kind: \"npx-run\",\n      cmd: \"bmad-method\",\n      version: \"6.9.0\",\n      args: [\"install\", \"--directory\", \".\", \"--tools\", \"claude-code\", \"--yes\"],\n    },\n  },\n\n  // === Railway (csr-fastify|csr-fastapi|ssr-*|full) ===\n  // v0.6.3 — railway-plugin entry 제거. railwayapp/railway-plugin repo 자체 존재 안 함\n  // (404 Not Found). 공식 docs (https://docs.railway.com/ai/claude-code-plugin) 형식은\n  // marketplace add `railwayapp/railway-skills` + plugin install `railway@railway-skills`만.\n  // → 아래 railway-skills entry로 단일화.\n  {\n    id: \"railway-skills\",\n    tier: \"experimental\", // railwayapp/railway-skills 268\n    description: \"Railway agent-skills (deploy + project/service/env management)\",\n    category: \"backend\",\n    source: \"railwayapp\",\n    condition: { kind: \"any-track\", tracks: RAILWAY_TRACKS },\n    method: {\n      kind: \"plugin\",\n      marketplace: \"railwayapp/railway-skills\",\n      pluginId: \"railway@railway-skills\",\n    },\n  },\n\n  // === csr-supabase|full CLI ===\n  {\n    id: \"vercel-cli\",\n    tier: \"vetted\", // vercel/vercel 15k\n    description: \"Vercel CLI (npm)\",\n    category: \"backend\",\n    source: \"vercel\",\n    condition: { kind: \"any-track\", tracks: [\"csr-supabase\", \"full\"] },\n    method: { kind: \"npm\", pkg: \"vercel\", version: \"54.17.3\" },\n  },\n  {\n    id: \"netlify-cli\",\n    tier: \"vetted\", // netlify/cli 1.9k\n    description: \"Netlify CLI (npm)\",\n    category: \"backend\",\n    source: \"netlify\",\n    condition: { kind: \"any-track\", tracks: [\"csr-supabase\", \"full\"] },\n    method: { kind: \"npm\", pkg: \"netlify-cli\", version: \"26.1.0\" },\n  },\n  {\n    id: \"supabase-cli\",\n    tier: \"vetted\", // supabase 103k\n    description: \"Supabase CLI (npm) — first 'supabase login' requires OAuth\",\n    category: \"backend\",\n    source: \"supabase\",\n    condition: { kind: \"any-track\", tracks: [\"csr-supabase\", \"full\"] },\n    method: { kind: \"npm\", pkg: \"supabase\", version: \"2.108.0\" },\n  },\n\n  // === UI tracks (csr-*|ssr-*|full) ===\n  {\n    id: \"impeccable\",\n    tier: \"vetted\", // pbakaus 31k\n    description:\n      \"Impeccable — UI design guide + visual consistency review (pbakaus, single-skill repo)\",\n    category: \"frontend\",\n    source: \"pbakaus\",\n    condition: { kind: \"any-track\", tracks: ALL_CSR_SSR_FULL },\n    // v26.54.1 — skills cli 1.5.7 부터 `--skill <name>` 명시 필수 (single-skill repo 도)\n    method: { kind: \"skill\", source: \"pbakaus/impeccable\", skill: \"impeccable\" },\n  },\n  // v26.92.0 — frontend-design (Anthropic official, claude-plugins-official 984.5K installs).\n  //   impeccable(생성↔리뷰 짝)의 official 보완재 — frontend-design=distinctive UI 코드 생성,\n  //   impeccable=일관성 리뷰. 사용자 결정: has-dev-track 기본추천 (impeccable=UI 트랙보다\n  //   넓게 — 모든 개발 트랙, executive 제외). category=frontend (UI 자산, wizard 그룹).\n  //   repoForAsset=marketplace(anthropics/claude-plugins-official); official tier=drift 제외.\n  {\n    id: \"frontend-design\",\n    tier: \"official\", // anthropics/claude-plugins-official (Anthropic 저자, 984.5K installs)\n    description:\n      \"frontend-design — distinctive production-grade UI generation (Anthropic official, avoids generic AI aesthetics)\",\n    category: \"frontend\",\n    source: \"anthropics\",\n    condition: { kind: \"has-dev-track\" },\n    method: {\n      kind: \"plugin\",\n      marketplace: \"anthropics/claude-plugins-official\",\n      pluginId: \"frontend-design@claude-plugins-official\",\n    },\n  },\n\n  // === dev tools (has_dev_track) ===\n  {\n    id: \"playwright-skill\",\n    tier: \"experimental\", // testdino-hq/playwright-skill 264\n    description: \"Playwright — browser automation E2E test authoring guide (testdino-hq)\",\n    category: \"dev-tools\",\n    source: \"testdino-hq\",\n    condition: { kind: \"has-dev-track\" },\n    // v26.54.1 — skills cli 1.5.7 부터 `--skill <name>` 명시 필수\n    method: {\n      kind: \"skill\",\n      source: \"testdino-hq/playwright-skill\",\n      skill: \"playwright-skill\",\n    },\n  },\n  {\n    id: \"find-skills\",\n    tier: \"vetted\", // vercel-labs/skills 20k (license none — 출처 신뢰)\n    description: \"find-skills — search · rank all installed skills (vercel-labs, all dev tracks)\",\n    category: \"dev-tools\",\n    source: \"vercel-labs\",\n    condition: { kind: \"has-dev-track\" },\n    method: { kind: \"skill\", source: \"vercel-labs/skills\", skill: \"find-skills\" },\n  },\n  {\n    id: \"agent-browser\",\n    tier: \"vetted\", // vercel-labs/agent-browser 34k\n    description:\n      \"agent-browser — agent-friendly Playwright wrapper (screenshot · DOM search CLI, dev tracks)\",\n    // v26.78.0 — Understanding 으로 재분류: 웹 지각(screenshot·DOM). 영상/코드 지각과 같은 축.\n    category: \"understanding\",\n    source: \"vercel-labs\",\n    condition: { kind: \"has-dev-track\" },\n    method: { kind: \"npm\", pkg: \"agent-browser\", version: \"0.31.0\" },\n  },\n  // v26.78.0 — Understanding 신규 3종 (plugin, opt-in). 에이전트 인지 증강: 영상·코드 지각 + 메모리.\n  {\n    id: \"claude-video\",\n    tier: \"vetted\", // bradautomates/claude-video 1.8k\n    description:\n      \"Claude Video — /watch downloads any video, extracts frames + transcript so Claude can see + hear it (yt-dlp/ffmpeg auto on first run)\",\n    category: \"understanding\",\n    source: \"bradautomates\",\n    condition: { kind: \"opt-in\" },\n    method: {\n      kind: \"plugin\",\n      marketplace: \"bradautomates/claude-video\",\n      pluginId: \"watch@claude-video\",\n    },\n  },\n  {\n    id: \"understand-anything\",\n    tier: \"vetted\", // Lum1104/Understand-Anything 53k\n    description:\n      \"Understand Anything — multi-agent pipeline builds an interactive knowledge graph of your codebase (files/functions/deps) to explore + query\",\n    category: \"understanding\",\n    source: \"Lum1104\",\n    condition: { kind: \"opt-in\" },\n    method: {\n      kind: \"plugin\",\n      marketplace: \"Lum1104/Understand-Anything\",\n      pluginId: \"understand-anything@understand-anything\",\n    },\n  },\n  {\n    id: \"agentmemory\",\n    tier: \"vetted\", // rohitg00/agentmemory 21k\n    description:\n      \"AgentMemory — persistent memory runtime; plugin auto-wires MCP (53 tools) + hooks + skills. Runtime server: npx @agentmemory/agentmemory\",\n    category: \"understanding\",\n    source: \"rohitg00\",\n    condition: { kind: \"opt-in\" },\n    method: {\n      kind: \"plugin\",\n      marketplace: \"rohitg00/agentmemory\",\n      pluginId: \"agentmemory@agentmemory\",\n    },\n  },\n\n  // === Visual & Media (v26.85.0) — 코드-퍼스트 제작 자산. 전부 opt-in. ===\n  // Docker 실설치 검증 PASS (실 claude 2.1.177): plugin install / npx skills add resolve 확인.\n  {\n    id: \"frontend-slides\",\n    tier: \"vetted\", // zarazhangrui/frontend-slides 21k\n    description:\n      \"frontend-slides — dependency-free HTML slide decks (presets · templates · PPTX→HTML · PDF export)\",\n    category: \"visual-media\",\n    source: \"zarazhangrui\",\n    condition: { kind: \"opt-in\" },\n    method: {\n      kind: \"plugin\",\n      marketplace: \"zarazhangrui/frontend-slides\",\n      pluginId: \"frontend-slides@frontend-slides\",\n    },\n  },\n  {\n    id: \"marp-slide\",\n    tier: \"vetted\", // softaworks/agent-toolkit 2k\n    description: \"marp-slide — Marp Markdown slides (7 themes · PPTX/PDF export)\",\n    category: \"visual-media\",\n    source: \"softaworks\",\n    // softaworks plugin dir 는 plugin.json 부재 → skill 경로가 안전 (Docker 검증).\n    condition: { kind: \"opt-in\" },\n    method: { kind: \"skill\", source: \"softaworks/agent-toolkit\", skill: \"marp-slide\" },\n  },\n  {\n    id: \"mermaid-diagrams\",\n    tier: \"vetted\", // softaworks/agent-toolkit 2k\n    description:\n      \"mermaid-diagrams — Mermaid flow/sequence/ER/state diagram authoring (code · docs)\",\n    category: \"visual-media\",\n    source: \"softaworks\",\n    condition: { kind: \"opt-in\" },\n    method: { kind: \"skill\", source: \"softaworks/agent-toolkit\", skill: \"mermaid-diagrams\" },\n  },\n  {\n    id: \"gsap-skills\",\n    tier: \"vetted\", // greensock/gsap-skills 9k (GSAP 본가 공식)\n    description:\n      \"GSAP skills — official GreenSock motion/scroll animation guide (8 skills: timeline · scrolltrigger · react)\",\n    category: \"visual-media\",\n    source: \"greensock\",\n    condition: { kind: \"opt-in\" },\n    method: {\n      kind: \"plugin\",\n      marketplace: \"greensock/gsap-skills\",\n      pluginId: \"gsap-skills@gsap-skills\",\n    },\n  },\n  {\n    // remotion: star 3.6k → vetted (star 기반 tier SSOT + trust-tier-drift CI 정합, 사용자 결정 2026-06-13).\n    //   코어 BUSL(Business Source License) — opt-in + description 고지로 신중 취급 (경고 배지 대신).\n    //   --skill 값 = remotion-best-practices (dir `remotion` ≠ frontmatter name, Docker 실측 확정).\n    id: \"remotion\",\n    tier: \"vetted\", // remotion-dev/skills 3.6k (license none — 출처 신뢰; 코어 BUSL 고지)\n    description:\n      \"Remotion — programmatic MP4 video from React components (data-driven). Core license = BUSL\",\n    category: \"visual-media\",\n    source: \"remotion-dev\",\n    condition: { kind: \"opt-in\" },\n    method: { kind: \"skill\", source: \"remotion-dev/skills\", skill: \"remotion-best-practices\" },\n  },\n  // Issue #176 — 프레젠테이션 생성 스택 4종 (Docker 실설치 4/4 PASS: skills@1.5.11 add <src> --agent claude-code --skill, 2026-06-20).\n  {\n    id: \"ppt-master\",\n    tier: \"vetted\", // hugohe3/ppt-master 29k\n    description:\n      \"ppt-master — editable PowerPoint (.pptx) from any document (native shapes · speaker notes · custom .pptx template)\",\n    category: \"visual-media\",\n    source: \"hugohe3\",\n    condition: { kind: \"opt-in\" },\n    method: { kind: \"skill\", source: \"hugohe3/ppt-master\", skill: \"ppt-master\" },\n  },\n  {\n    // deer-flow 72k 거대 harness지만 --skill 로 ppt-generation 단일 skill만 설치 (skills/public/ 중첩, Docker 확인).\n    id: \"ppt-generation\",\n    tier: \"vetted\", // bytedance/deer-flow 72k\n    description:\n      \"ppt-generation — PPTX by generating an image per slide and composing into PowerPoint (deer-flow skill)\",\n    category: \"visual-media\",\n    source: \"bytedance\",\n    condition: { kind: \"opt-in\" },\n    method: { kind: \"skill\", source: \"bytedance/deer-flow\", skill: \"ppt-generation\" },\n  },\n  {\n    id: \"web-video-presentation\",\n    tier: \"vetted\", // ConardLi/garden-skills 8.4k\n    description:\n      \"web-video-presentation — click-driven 16:9 web decks that look like video (optional TTS narration; garden-skills)\",\n    category: \"visual-media\",\n    source: \"ConardLi\",\n    condition: { kind: \"opt-in\" },\n    method: { kind: \"skill\", source: \"ConardLi/garden-skills\", skill: \"web-video-presentation\" },\n  },\n  {\n    id: \"revealjs\",\n    tier: \"experimental\", // ryanbbrown/revealjs-skill 347 (<1000 → experimental, opt-in + 경고)\n    description:\n      \"reveal.js — polished HTML presentations (themes · multi-column · code highlight · speaker notes, no build step)\",\n    category: \"visual-media\",\n    source: \"ryanbbrown\",\n    condition: { kind: \"opt-in\" },\n    method: { kind: \"skill\", source: \"ryanbbrown/revealjs-skill\", skill: \"revealjs\" },\n  },\n  {\n    id: \"architecture-decision-record\",\n    tier: \"experimental\", // yonatangross/orchestkit 179\n    description:\n      \"ADR — Architecture Decision Record template + status flow (orchestkit, one of 80+ skills)\",\n    category: \"dev-tools\",\n    source: \"yonatangross\",\n    condition: { kind: \"has-dev-track\" },\n    method: {\n      kind: \"skill\",\n      source: \"yonatangross/orchestkit\",\n      skill: \"architecture-decision-record\",\n    },\n  },\n\n  // === Supabase agent-skills (csr-supabase|full) ===\n  {\n    id: \"supabase-agent-skills\",\n    tier: \"vetted\", // supabase/agent-skills 2.2k\n    description:\n      \"Supabase — RLS · auth · edge function · realtime guide (csr-supabase · full tracks)\",\n    category: \"backend\",\n    source: \"supabase\",\n    condition: { kind: \"any-track\", tracks: [\"csr-supabase\", \"full\"] },\n    method: {\n      kind: \"plugin\",\n      marketplace: \"supabase/agent-skills\",\n      pluginId: \"supabase@supabase-agent-skills\",\n    },\n  },\n  {\n    id: \"postgres-best-practices\",\n    tier: \"vetted\", // supabase/agent-skills 2.2k\n    description:\n      \"Postgres best practices — schema · index · query patterns (csr-supabase · full tracks)\",\n    category: \"backend\",\n    source: \"supabase\",\n    condition: { kind: \"any-track\", tracks: [\"csr-supabase\", \"full\"] },\n    method: {\n      kind: \"plugin\",\n      marketplace: \"supabase/agent-skills\",\n      pluginId: \"postgres-best-practices@supabase-agent-skills\",\n    },\n  },\n\n  // === React + Next UI tracks ===\n  // v0.6.3 — vercel-labs/agent-skills source는 short form 안 됨. full HTTPS URL 필요.\n  // 사용자 확인 형식: `npx skills add https://github.com/vercel-labs/agent-skills --skill <name>`.\n  {\n    id: \"react-best-practices\",\n    tier: \"vetted\", // vercel-labs/agent-skills 27k (license none — 출처 신뢰)\n    description:\n      \"React best practices — Vercel's hook · perf · component patterns (CSR · SSR · Next tracks)\",\n    category: \"frontend\",\n    source: \"vercel-labs\",\n    condition: { kind: \"any-track\", tracks: CSR_SSR_NEXTJS_FULL },\n    method: {\n      kind: \"skill\",\n      source: \"https://github.com/vercel-labs/agent-skills\",\n      // v0.6.5 — skills.sh registry name. GitHub dir 이름(react-best-practices)과 다름.\n      // skills.sh: 대부분 vercel- prefix (web-design-guidelines, deploy-to-vercel만 예외).\n      skill: \"vercel-react-best-practices\",\n    },\n  },\n  {\n    id: \"shadcn-ui\",\n    tier: \"vetted\", // shadcn-ui/ui 115k\n    description: \"shadcn/ui — Radix-based React component copy + Tailwind theme (shadcn official)\",\n    category: \"frontend\",\n    source: \"shadcn-ui\",\n    condition: { kind: \"any-track\", tracks: CSR_SSR_NEXTJS_FULL },\n    // v26.54.1 — shadcn/ui repo 의 실제 skill 이름은 `shadcn` (자산 id 와 다름).\n    method: { kind: \"skill\", source: \"shadcn/ui\", skill: \"shadcn\" },\n  },\n  {\n    id: \"web-design-guidelines\",\n    tier: \"vetted\", // vercel-labs/agent-skills 27k (license none — 출처 신뢰)\n    description:\n      \"Web design guidelines — Vercel's visual hierarchy · color · spacing (CSR · SSR · Next tracks)\",\n    category: \"frontend\",\n    source: \"vercel-labs\",\n    condition: { kind: \"any-track\", tracks: CSR_SSR_NEXTJS_FULL },\n    method: {\n      kind: \"skill\",\n      source: \"https://github.com/vercel-labs/agent-skills\",\n      skill: \"web-design-guidelines\",\n    },\n  },\n  // === Executive tracks ===\n  {\n    id: \"anthropic-document-skills\",\n    tier: \"official\", // anthropics/skills 144k\n    description: \"Anthropic document-skills (pptx/docx/xlsx/pdf)\",\n    category: \"business\",\n    source: \"anthropics\",\n    condition: { kind: \"any-track\", tracks: [\"executive\", \"full\"] },\n    method: {\n      kind: \"plugin\",\n      marketplace: \"anthropics/skills\",\n      pluginId: \"document-skills@anthropic-agent-skills\",\n    },\n  },\n  // alirezarezvani/claude-skills marketplace (v2.3.0) — 2026-04-25 통합 갱신.\n  // 기존 alirezarezvani/c-level-skills + alirezarezvani/finance-skills 별도 marketplace\n  // → 통합된 alirezarezvani/claude-skills marketplace (claude-code-skills 이름)로 이동.\n  {\n    id: \"c-level-skills\",\n    tier: \"vetted\", // alirezarezvani 16k\n    description: \"c-level-skills (claude-code-skills, 28 advisory)\",\n    category: \"business\",\n    source: \"alirezarezvani\",\n    condition: { kind: \"any-track\", tracks: [\"executive\", \"full\"] },\n    method: {\n      kind: \"plugin\",\n      marketplace: \"alirezarezvani/claude-skills\",\n      pluginId: \"c-level-skills@claude-code-skills\",\n    },\n  },\n  {\n    id: \"business-growth-skills\",\n    tier: \"vetted\", // alirezarezvani 16k\n    description: \"business-growth-skills (4 — customer success, sales eng, revops, contract)\",\n    category: \"business\",\n    source: \"alirezarezvani\",\n    // v0.5.0 — growth-marketing Track에서도 재사용. 합집합 조건.\n    condition: { kind: \"any-track\", tracks: [\"executive\", \"full\", \"growth-marketing\"] },\n    method: {\n      kind: \"plugin\",\n      marketplace: \"alirezarezvani/claude-skills\",\n      pluginId: \"business-growth-skills@claude-code-skills\",\n    },\n  },\n  {\n    id: \"finance-skills\",\n    tier: \"vetted\", // alirezarezvani 16k\n    description: \"finance-skills (3 — financial analyst, SaaS metrics, investment advisor)\",\n    category: \"business\",\n    source: \"alirezarezvani\",\n    condition: { kind: \"any-track\", tracks: [\"executive\", \"full\"] },\n    method: {\n      kind: \"plugin\",\n      marketplace: \"alirezarezvani/claude-skills\",\n      pluginId: \"finance-skills@claude-code-skills\",\n    },\n  },\n\n  // === Project Management Track (v0.5.0) ===\n  // SPEC docs/specs/new-tracks-pm-growth.md §3.5 — pm-skills 4/4.\n  {\n    id: \"pm-skills\",\n    tier: \"vetted\", // alirezarezvani 16k\n    description:\n      \"pm-skills (6 — senior PM, scrum master, Jira/Confluence/Atlassian admin, template creator)\",\n    category: \"business\",\n    source: \"alirezarezvani\",\n    condition: { kind: \"any-track\", tracks: [\"project-management\"] },\n    method: {\n      kind: \"plugin\",\n      marketplace: \"alirezarezvani/claude-skills\",\n      pluginId: \"pm-skills@claude-code-skills\",\n    },\n  },\n  // SPEC §3.5 — product-skills: has-dev-track + project-management 합집합 (executive/growth-marketing 제외).\n  // v0.8.1 — DEV_PLUS_PM_TRACKS 상수로 SSOT 통일 (reviewer MEDIUM-3 fix).\n  {\n    id: \"product-skills\",\n    tier: \"vetted\", // alirezarezvani 16k\n    description: \"product-skills (15 — RICE, PRD, agile PO, UX research, SaaS scaffolder ...)\",\n    category: \"dev-tools\",\n    source: \"alirezarezvani\",\n    condition: { kind: \"any-track\", tracks: [...DEV_PLUS_PM_TRACKS] },\n    method: {\n      kind: \"plugin\",\n      marketplace: \"alirezarezvani/claude-skills\",\n      pluginId: \"product-skills@claude-code-skills\",\n    },\n  },\n\n  // === Growth Marketing Track (v0.5.0) ===\n  // SPEC docs/specs/new-tracks-pm-growth.md §3.5 — 4 entries 모두 4/4.\n  {\n    id: \"marketing-skills\",\n    tier: \"vetted\", // alirezarezvani 16k\n    description:\n      \"marketing-skills (44 — content/SEO/CRO/channels/growth/intelligence/sales/twitter)\",\n    category: \"business\",\n    source: \"alirezarezvani\",\n    condition: { kind: \"any-track\", tracks: [\"growth-marketing\"] },\n    method: {\n      kind: \"plugin\",\n      marketplace: \"alirezarezvani/claude-skills\",\n      pluginId: \"marketing-skills@claude-code-skills\",\n    },\n  },\n  // v26.91.0 — coreyhaines31/marketingskills (35k★, MIT, plugin v2.5.1). 위 alirezarezvani\n  //   marketing-skills(16k) 와 병존 — id 구분(marketingskills ≠ marketing-skills). opt-in =\n  //   growth-marketing 외 전 트랙에서도 wizard 토글 + `--with marketingskills` 로 설치 가능\n  //   (\"SEO 는 일반 개발에도 활용\" 충족: dev 트랙에서 본 번들 토글). 45 스킬 중 SEO 7종\n  //   (seo-audit/schema/ai-seo/site-architecture/programmatic-seo/content/aso)은 상호참조\n  //   (product-marketing 우선 + seo-audit↔schema↔ai-seo)라 부분 추출 시 포인터 깨짐 → 번들\n  //   통째 설치만 정합. repoForAsset = marketplace(coreyhaines31/marketingskills) → drift 라이브.\n  {\n    id: \"marketingskills\",\n    tier: \"vetted\", // coreyhaines31 35k\n    description: \"marketingskills (45 — CRO/copywriting/SEO/AI-SEO/ads/growth, coreyhaines31 35k★)\",\n    category: \"business\",\n    source: \"coreyhaines31\",\n    condition: { kind: \"opt-in\" },\n    method: {\n      kind: \"plugin\",\n      marketplace: \"coreyhaines31/marketingskills\",\n      pluginId: \"marketing-skills@marketingskills\",\n    },\n  },\n  // v26.76.0 — content-creator / demand-gen 제거: alirezarezvani/claude-skills marketplace.json 에\n  // 해당 plugin 부재(Docker 실설치 검출, exit 1). 거짓 광고 0건 원칙(Promise=Implementation).\n  // growth-marketing 트랙은 business-growth-skills + marketing-skills + research-summarizer 유지.\n  {\n    id: \"research-summarizer\",\n    tier: \"vetted\", // alirezarezvani 16k\n    description: \"research-summarizer (market research summarization)\",\n    category: \"business\",\n    source: \"alirezarezvani\",\n    condition: { kind: \"any-track\", tracks: [\"growth-marketing\"] },\n    method: {\n      kind: \"plugin\",\n      marketplace: \"alirezarezvani/claude-skills\",\n      pluginId: \"research-summarizer@claude-code-skills\",\n    },\n  },\n\n  // === Code-quality enforcement (has-dev-track, v0.5.0) ===\n  // SPEC §3.5 — karpathy-coder 4/4. CLAUDE.md P1-P4 선언적 원칙의 검출 도구 layer.\n  // 4 Python tools (stdlib only) + reviewer agent + /karpathy-check + pre-commit hook.\n  {\n    id: \"karpathy-coder\",\n    tier: \"vetted\", // alirezarezvani 16k\n    description:\n      \"karpathy-coder (4 Python tool + reviewer agent + /karpathy-check + pre-commit hook)\",\n    category: \"dev-tools\",\n    source: \"alirezarezvani\",\n    condition: { kind: \"has-dev-track\" },\n    method: {\n      kind: \"plugin\",\n      marketplace: \"alirezarezvani/claude-skills\",\n      pluginId: \"karpathy-coder@claude-code-skills\",\n    },\n  },\n\n  // === Option-gated ===\n  {\n    // v26.39.2 fix — marketplace name = \"trailofbits\" (NOT \"trailofbits-skills\") +\n    // \"trailofbits-skills\" plugin 자체가 존재하지 않음. marketplace 안에 14+ 개별 plugin.\n    // 단일 대표 plugin = `differential-review` (코드 변경 보안 리뷰, 가장 보편).\n    // 추가 plugin 원하는 사용자는: `claude plugin install <name>@trailofbits` (예: audit-context-building)\n    id: \"trailofbits-skills\",\n    tier: \"vetted\", // trailofbits/skills 5.5k (CC-BY-SA — 출처 신뢰)\n    description: \"Trail of Bits differential-review plugin (security-focused code review)\",\n    category: \"dev-tools\",\n    source: \"trailofbits\",\n    condition: { kind: \"opt-in\" },\n    method: {\n      kind: \"plugin\",\n      marketplace: \"trailofbits/skills\",\n      pluginId: \"differential-review@trailofbits\",\n    },\n  },\n  {\n    id: \"ecc-plugin\",\n    tier: \"vetted\", // affaan-m/everything-claude-code 199k\n    description: \"ECC — 60 agents · 230 skills · 75 commands. Affaan's hackathon package\",\n    category: \"ecc-suite\",\n    source: \"affaan-m\",\n    condition: { kind: \"opt-in\" },\n    // v26.54.1 — upstream marketplace.json 의 name 은 \"ecc\" (plugin name 도 \"ecc\").\n    // 기존 매핑 `everything-claude-code@everything-claude-code` 는 marketplace 가\n    // 그 이름으로 등록되던 옛 버전 기준. fresh install 에서는 \"Plugin not found\" 발생.\n    method: {\n      kind: \"plugin\",\n      marketplace: \"affaan-m/everything-claude-code\",\n      pluginId: \"ecc@ecc\",\n    },\n  },\n  {\n    id: \"ecc-prune\",\n    tier: \"official\", // uzys 본 하네스 자체\n    description:\n      \"ECC prune (drop items beyond curated 89 KEEP → copy to .claude/local-plugins/ecc/)\",\n    category: \"ecc-suite\",\n    source: \"uzys\",\n    condition: { kind: \"option\", flag: \"withPrune\" },\n    method: {\n      kind: \"shell-script\",\n      script: \"scripts/prune-ecc.sh\",\n      args: [\"--apply\", \"--force\"],\n    },\n  },\n];\n\n/**\n * v26.87.0 — dev-method skill ids (uzys 1st-party, internal templates). installer 의\n * `selectedInternalSkills` 계산 + manifest copy 게이팅 + 테스트가 공유하는 SSOT.\n * 각 id 는 method.kind===\"internal\" 이며 `templates/skills/<id>/SKILL.md` 로 번들된다.\n */\nexport const DEV_METHOD_SKILL_IDS: ReadonlyArray<string> = [\n  \"multi-persona-review\",\n  \"gap-analysis-e2e\",\n  \"ultracode-service-audit\",\n  \"asis-tobe-decision\",\n  \"compaction-handoff\",\n  \"northstar-roadmap\",\n  // v26.93.0 — Orchestration & Model Policy.\n  \"model-orchestration\",\n];\n\n/**\n * v26.95.0 — ALL repo-bundled internal skill ids (dev-method + opt-in advisors). Bundling is\n * condition-agnostic: manifest Claude dir-copy, the 3 non-Claude CLI transforms, and\n * gen-compatibility iterate THIS superset so every bundled skill renders across CLIs; each entry's\n * `condition` (has-dev-track vs opt-in) still gates whether it actually installs. Kept separate\n * from `DEV_METHOD_SKILL_IDS` so \"dev-method\" keeps meaning the 7 has-dev-track methodology skills.\n */\nexport const INTERNAL_BUNDLED_SKILL_IDS: ReadonlyArray<string> = [\n  ...DEV_METHOD_SKILL_IDS,\n  // opt-in internal skills (NOT dev-method): bundled + 4-CLI rendered, installed only on opt-in.\n  \"gemini-consult\",\n];\n\n/**\n * v26.79.0 — `TRUST_TIER` 는 EXTERNAL_ASSETS.tier 에서 derive (단일 출처). 별도 Record 유지 시\n * 누락/stale drift 가능 → 제거. 기존 소비자(prompts.ts·gen-compatibility·trust-tier-drift)는\n * 이 derived map 을 그대로 import. id 키는 각 자산 id 와 1:1 (자산 추가 시 자동 반영).\n */\nexport const TRUST_TIER: Record<string, TrustTier> = Object.fromEntries(\n  EXTERNAL_ASSETS.map((a) => [a.id, a.tier]),\n);\n\n/** 자산의 검증 tier. 미분류(catalog 외 id)는 보수적으로 experimental. */\nexport function assetTrustTier(assetId: string): TrustTier {\n  return TRUST_TIER[assetId] ?? \"experimental\";\n}\n\n/**\n * v26.47.0 — User override of preset/option condition (Phase C full, SPEC §3.1).\n * - `forceInclude`: condition 무관 강제 포함 (사용자가 명시 추가)\n * - `forceExclude`: condition 무관 강제 제외 (사용자가 추천 ✓ 풀음)\n *\n * 우선순위: `forceExclude` > `forceInclude` > `condition`.\n */\nexport interface UserOverride {\n  forceInclude: ReadonlyArray<string>;\n  forceExclude: ReadonlyArray<string>;\n}\n\n/**\n * 조건 평가 — 주어진 spec(tracks + options + userOverride)에서 자산이 설치 대상인지 판정.\n */\nexport function shouldInstallAsset(\n  asset: ExternalAsset,\n  ctx: {\n    tracks: ReadonlyArray<Track>;\n    options: OptionFlags;\n    userOverride?: UserOverride;\n  },\n): boolean {\n  // v26.47.0 — userOverride 우선순위: forceExclude > forceInclude > condition.\n  if (ctx.userOverride?.forceExclude.includes(asset.id)) return false;\n  if (ctx.userOverride?.forceInclude.includes(asset.id)) return true;\n  // v26.71.1 — experimental(T3) opt-in only (PRD v26-71 R6/AC4). condition 매치만으론 미설치.\n  //   --with <id> (forceInclude) 또는 interactive 체크(→forceInclude) 시에만 설치 — 선택권 유지(강제 차단 아님).\n  //   v26.71.0 은 recommendedExternalAssets(pre-check)에만 제외 적용 → 비대화형/미체크 install 경로 누락 버그 fix.\n  //   게이트는 명시 분류(TRUST_TIER[id])만 본다 — assetTrustTier()의 unknown→experimental default 는\n  //   DISPLAY(경고 배지)용이며 설치 게이트엔 미적용. 실 자산 전부 매핑은 \"no-missing\" 테스트가 강제(AC1).\n  if (TRUST_TIER[asset.id] === \"experimental\") return false;\n  return matchesCondition(asset, ctx);\n}\n\n/**\n * Track/option condition 매치만 평가 (tier·override 무관). shouldInstallAsset 의 조건절 +\n * experimentalOptInCandidates 의 \"조건은 맞지만 T3\" 판정에서 공유 (SSOT).\n */\nfunction matchesCondition(\n  asset: ExternalAsset,\n  ctx: { tracks: ReadonlyArray<Track>; options: OptionFlags },\n): boolean {\n  const cond = asset.condition;\n  switch (cond.kind) {\n    case \"any-track\":\n      return ctx.tracks.some((t) => cond.tracks.includes(t));\n    case \"has-dev-track\":\n      // SSOT — track-match.ts hasDevTrack(): csr-*|ssr-*|data|full|tooling (= DEV_TRACKS).\n      // EXECUTIVE_STYLE_TRACKS (executive + project-management + growth-marketing) 는 제외.\n      return hasDevTrack(ctx.tracks);\n    case \"option\":\n      return ctx.options[cond.flag] === true;\n    case \"opt-in\":\n      // v26.81.0 (ADR-022) — 순수 opt-in: condition 으론 절대 설치 안 됨.\n      //   forceInclude(wizard 체크 / --with <id>)가 shouldInstallAsset 상위에서 처리.\n      return false;\n  }\n}\n\n/**\n * v26.81.0 (ADR-022) — spec 에서 특정 자산의 선택 여부 판정.\n * 내부 자산(tauri-desktop)의 manifest/transform 게이팅이 이전\n * `spec.options.withTauri` 등 boolean 자리를 대체해 호출. (wizard 체크/--with 는\n * forceInclude 로 들어오므로 shouldInstallAsset 가 그대로 판정)\n */\nexport function isAssetSelected(\n  assetId: string,\n  ctx: {\n    tracks: ReadonlyArray<Track>;\n    options: OptionFlags;\n    userOverride?: UserOverride;\n  },\n): boolean {\n  const asset = EXTERNAL_ASSETS.find((a) => a.id === assetId);\n  return asset ? shouldInstallAsset(asset, ctx) : false;\n}\n\n/**\n * v26.71.1 — track/option condition 은 매치하지만 T3(experimental)라서 default 설치에서 제외된 자산.\n * forceInclude(--with / interactive 체크)된 것은 이미 설치되므로 제외. 비대화형 install 의\n * discoverability 힌트용 (Transparent Defaults — 숨김 0건. --with 로 opt-in 가능함을 사용자에게 알림).\n */\nexport function experimentalOptInCandidates(ctx: {\n  tracks: ReadonlyArray<Track>;\n  options: OptionFlags;\n  userOverride?: UserOverride;\n}): ReadonlyArray<ExternalAsset> {\n  return EXTERNAL_ASSETS.filter(\n    (a) =>\n      TRUST_TIER[a.id] === \"experimental\" &&\n      !ctx.userOverride?.forceInclude.includes(a.id) &&\n      matchesCondition(a, ctx),\n  );\n}\n\n/**\n * spec에 적용 가능한 자산 필터.\n */\nexport function filterApplicableAssets(\n  assets: ReadonlyArray<ExternalAsset>,\n  ctx: {\n    tracks: ReadonlyArray<Track>;\n    options: OptionFlags;\n    userOverride?: UserOverride;\n  },\n): ReadonlyArray<ExternalAsset> {\n  return assets.filter((a) => shouldInstallAsset(a, ctx));\n}\n","import type { Track } from \"./types.js\";\n\n/** Match a track against a glob-style pattern (bash-compatible: `csr-*`, `csr-*|full`). */\nexport function matchTrack(track: Track, pattern: string): boolean {\n  return pattern.split(\"|\").some((p) => globToRegex(p.trim()).test(track));\n}\n\nexport function anyTrack(tracks: ReadonlyArray<Track>, pattern: string): boolean {\n  return tracks.some((t) => matchTrack(t, pattern));\n}\n\nexport function hasDevTrack(tracks: ReadonlyArray<Track>): boolean {\n  // Dev tracks = anything that is NOT executive AND NOT tooling-only-meta.\n  // Mirrors setup-harness.sh has_dev_track: csr-*, ssr-*, data, full, tooling.\n  return anyTrack(tracks, \"csr-*|ssr-*|data|full|tooling\");\n}\n\nexport function hasUiTrack(tracks: ReadonlyArray<Track>): boolean {\n  return anyTrack(tracks, \"csr-*|ssr-*|full\");\n}\n\n/** Convert a bash glob (only `*` is supported) to a strict-anchored RegExp. */\nfunction globToRegex(pattern: string): RegExp {\n  const escaped = pattern.replace(/[.+?^${}()|[\\]\\\\]/g, \"\\\\$&\").replace(/\\*/g, \".*\");\n  return new RegExp(`^${escaped}$`);\n}\n","/**\n * Category-based installer (v26.43.0) — 분야별 대표 스킬 큐레이션.\n *\n * 9 카테고리: Frontend / Backend / Data / Business / Dev Tools / Understanding / Visual & Media / Workflow / ECC Suite.\n * 각 카테고리 안에서 자산은 (출처, ExternalAsset.id) 로 식별. Step 2 UI 그룹화 기준.\n *\n * Source labels = 정확한 GitHub org/user (Phase A의 SSOT).\n * `[3rd-party]` 같은 generic 라벨 금지 (SPEC R6).\n */\n\nexport const CATEGORIES = [\n  \"frontend\",\n  \"backend\",\n  \"data\",\n  \"business\",\n  \"dev-tools\",\n  \"understanding\",\n  \"visual-media\",\n  \"workflow\",\n  \"ecc-suite\",\n] as const;\nexport type Category = (typeof CATEGORIES)[number];\n\nexport const CATEGORY_TITLES: Record<Category, string> = {\n  frontend: \"🎨 Frontend (UI · Design)\",\n  backend: \"🗄️  Backend (API · DB · Deploy)\",\n  data: \"📊 Data\",\n  business: \"💼 Business (Documents)\",\n  \"dev-tools\": \"🛡️  Dev Tools (Security · Quality)\",\n  // v26.78.0 — 에이전트 인지 증강: 환경에 대한 지각(영상·웹·코드) + 기억(memory).\n  understanding: \"🧠 Understanding (Perception · Memory)\",\n  // v26.85.0 — 코드-퍼스트 비주얼/미디어 제작: 슬라이드·다이어그램·모션·동영상.\n  \"visual-media\": \"🎬 Visual & Media (Slides · Diagrams · Motion · Video)\",\n  workflow: \"🔄 Workflow (Development Cycle)\",\n  \"ecc-suite\": \"📦 ECC Suite\",\n};\n\n/** Source labels — GitHub org/user. Maps the canonical owner shown to the user. */\nexport const SOURCE_LABELS = {\n  anthropics: \"anthropics\",\n  \"vercel-labs\": \"vercel-labs\",\n  \"shadcn-ui\": \"shadcn/ui\",\n  vercel: \"vercel\",\n  netlify: \"netlify\",\n  supabase: \"supabase\",\n  railwayapp: \"railwayapp\",\n  trailofbits: \"trailofbits\",\n  obra: \"obra\",\n  addyosmani: \"addyosmani\",\n  pbakaus: \"pbakaus\",\n  \"K-Dense-AI\": \"K-Dense-AI\",\n  wshobson: \"wshobson\",\n  \"testdino-hq\": \"testdino-hq\",\n  yonatangross: \"yonatangross\",\n  alirezarezvani: \"alirezarezvani\",\n  \"affaan-m\": \"affaan-m\",\n  \"get-shit-done-cc\": \"get-shit-done-cc\",\n  \"fission-ai\": \"Fission-AI\",\n  \"bmad-code-org\": \"bmad-code-org\",\n  // v26.78.0 — Understanding 카테고리 신규 출처\n  bradautomates: \"bradautomates\",\n  Lum1104: \"Lum1104\",\n  rohitg00: \"rohitg00\",\n  // v26.85.0 — Visual & Media 카테고리 출처 (Docker 실설치 검증 PASS).\n  zarazhangrui: \"zarazhangrui\",\n  greensock: \"greensock\",\n  softaworks: \"softaworks\",\n  \"remotion-dev\": \"remotion-dev\",\n  // v26.86.0 — Visual & Media 프레젠테이션 4종 (Issue #176, Docker 실설치 4/4 PASS).\n  hugohe3: \"hugohe3\",\n  bytedance: \"bytedance\",\n  ConardLi: \"ConardLi\",\n  ryanbbrown: \"ryanbbrown\",\n  // v26.91.0 — Business 카테고리 마케팅 번들 (coreyhaines31/marketingskills 35k★, MIT).\n  coreyhaines31: \"coreyhaines31\",\n  uzys: \"this project\",\n} as const;\nexport type Source = keyof typeof SOURCE_LABELS;\n"],"mappings":";;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;AACA,OAAO,UAAU;AACjB,SAAS,qBAAqB;AAF9B;AAAA;AAAA;AAAA;AAAA;;;ACAA;;;ACAA;AAGO,SAAS,WAAW,OAAc,SAA0B;AACjE,SAAO,QAAQ,MAAM,GAAG,EAAE,KAAK,CAAC,MAAM,YAAY,EAAE,KAAK,CAAC,EAAE,KAAK,KAAK,CAAC;AACzE;AAEO,SAAS,SAAS,QAA8B,SAA0B;AAC/E,SAAO,OAAO,KAAK,CAAC,MAAM,WAAW,GAAG,OAAO,CAAC;AAClD;AAEO,SAAS,YAAY,QAAuC;AAGjE,SAAO,SAAS,QAAQ,+BAA+B;AACzD;AAEO,SAAS,WAAW,QAAuC;AAChE,SAAO,SAAS,QAAQ,kBAAkB;AAC5C;AAGA,SAAS,YAAY,SAAyB;AAC5C,QAAM,UAAU,QAAQ,QAAQ,sBAAsB,MAAM,EAAE,QAAQ,OAAO,IAAI;AACjF,SAAO,IAAI,OAAO,IAAI,OAAO,GAAG;AAClC;;;AD2EA,IAAM,mBAA4B;AAAA,EAChC;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AACF;AAGA,IAAM,sBAA+B;AAAA,EACnC;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AACF;AAEA,IAAM,iBAA0B,CAAC,eAAe,eAAe,YAAY,cAAc,MAAM;AAwBxF,IAAM,aAAmC;AAAA,EAC9C;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AACF;AAOO,IAAM,qBAA2C,CAAC,GAAG,YAAY,oBAAoB;AAQrF,IAAM,kBAAgD;AAAA;AAAA,EAE3D;AAAA,IACE,IAAI;AAAA,IACJ,MAAM;AAAA;AAAA,IACN,aAAa;AAAA,IACb,UAAU;AAAA,IACV,QAAQ;AAAA,IACR,WAAW,EAAE,MAAM,aAAa,QAAQ,CAAC,QAAQ,MAAM,EAAE;AAAA,IACzD,QAAQ,EAAE,MAAM,SAAS,QAAQ,sCAAsC,OAAO,SAAS;AAAA;AAAA,EAEzF;AAAA,EACA;AAAA,IACE,IAAI;AAAA,IACJ,MAAM;AAAA;AAAA,IACN,aAAa;AAAA,IACb,UAAU;AAAA,IACV,QAAQ;AAAA,IACR,WAAW,EAAE,MAAM,aAAa,QAAQ,CAAC,QAAQ,MAAM,EAAE;AAAA,IACzD,QAAQ,EAAE,MAAM,SAAS,QAAQ,sCAAsC,OAAO,OAAO;AAAA,EACvF;AAAA,EACA;AAAA,IACE,IAAI;AAAA,IACJ,MAAM;AAAA;AAAA,IACN,aAAa;AAAA,IACb,UAAU;AAAA,IACV,QAAQ;AAAA,IACR,WAAW,EAAE,MAAM,aAAa,QAAQ,CAAC,QAAQ,MAAM,EAAE;AAAA,IACzD,QAAQ;AAAA,MACN,MAAM;AAAA,MACN,QAAQ;AAAA,MACR,OAAO;AAAA,IACT;AAAA,EACF;AAAA,EACA;AAAA,IACE,IAAI;AAAA,IACJ,MAAM;AAAA;AAAA,IACN,aAAa;AAAA,IACb,UAAU;AAAA,IACV,QAAQ;AAAA,IACR,WAAW,EAAE,MAAM,aAAa,QAAQ,CAAC,QAAQ,MAAM,EAAE;AAAA,IACzD,QAAQ;AAAA,MACN,MAAM;AAAA,MACN,QAAQ;AAAA,MACR,OAAO;AAAA,IACT;AAAA,EACF;AAAA,EACA;AAAA,IACE,IAAI;AAAA,IACJ,MAAM;AAAA;AAAA,IACN,aAAa;AAAA,IACb,UAAU;AAAA,IACV,QAAQ;AAAA,IACR,WAAW,EAAE,MAAM,aAAa,QAAQ,CAAC,QAAQ,MAAM,EAAE;AAAA,IACzD,QAAQ;AAAA,MACN,MAAM;AAAA,MACN,aAAa;AAAA,MACb,UAAU;AAAA,IACZ;AAAA,EACF;AAAA;AAAA,EAGA;AAAA,IACE,IAAI;AAAA,IACJ,MAAM;AAAA;AAAA,IACN,aAAa;AAAA,IACb,UAAU;AAAA,IACV,QAAQ;AAAA,IACR,WAAW,EAAE,MAAM,SAAS;AAAA,IAC5B,QAAQ,EAAE,MAAM,YAAY,KAAK,gBAAgB;AAAA,EACnD;AAAA;AAAA;AAAA;AAAA,EAKA;AAAA,IACE,IAAI;AAAA,IACJ,MAAM;AAAA;AAAA,IACN,aACE;AAAA,IACF,UAAU;AAAA,IACV,QAAQ;AAAA,IACR,WAAW,EAAE,MAAM,gBAAgB;AAAA,IACnC,QAAQ,EAAE,MAAM,YAAY,KAAK,uBAAuB;AAAA,EAC1D;AAAA,EACA;AAAA,IACE,IAAI;AAAA,IACJ,MAAM;AAAA;AAAA,IACN,aACE;AAAA,IACF,UAAU;AAAA,IACV,QAAQ;AAAA,IACR,WAAW,EAAE,MAAM,gBAAgB;AAAA,IACnC,QAAQ,EAAE,MAAM,YAAY,KAAK,mBAAmB;AAAA,EACtD;AAAA,EACA;AAAA,IACE,IAAI;AAAA,IACJ,MAAM;AAAA;AAAA,IACN,aACE;AAAA,IACF,UAAU;AAAA,IACV,QAAQ;AAAA,IACR,WAAW,EAAE,MAAM,gBAAgB;AAAA,IACnC,QAAQ,EAAE,MAAM,YAAY,KAAK,0BAA0B;AAAA,EAC7D;AAAA,EACA;AAAA,IACE,IAAI;AAAA,IACJ,MAAM;AAAA;AAAA,IACN,aACE;AAAA,IACF,UAAU;AAAA,IACV,QAAQ;AAAA,IACR,WAAW,EAAE,MAAM,gBAAgB;AAAA,IACnC,QAAQ,EAAE,MAAM,YAAY,KAAK,qBAAqB;AAAA,EACxD;AAAA,EACA;AAAA,IACE,IAAI;AAAA,IACJ,MAAM;AAAA;AAAA,IACN,aACE;AAAA,IACF,UAAU;AAAA,IACV,QAAQ;AAAA,IACR,WAAW,EAAE,MAAM,gBAAgB;AAAA,IACnC,QAAQ,EAAE,MAAM,YAAY,KAAK,qBAAqB;AAAA,EACxD;AAAA,EACA;AAAA,IACE,IAAI;AAAA,IACJ,MAAM;AAAA;AAAA,IACN,aACE;AAAA,IACF,UAAU;AAAA,IACV,QAAQ;AAAA,IACR,WAAW,EAAE,MAAM,gBAAgB;AAAA,IACnC,QAAQ,EAAE,MAAM,YAAY,KAAK,oBAAoB;AAAA,EACvD;AAAA;AAAA;AAAA;AAAA;AAAA,EAKA;AAAA,IACE,IAAI;AAAA,IACJ,MAAM;AAAA;AAAA,IACN,aACE;AAAA,IACF,UAAU;AAAA,IACV,QAAQ;AAAA,IACR,WAAW,EAAE,MAAM,gBAAgB;AAAA,IACnC,QAAQ,EAAE,MAAM,YAAY,KAAK,sBAAsB;AAAA,EACzD;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA,EAUA;AAAA,IACE,IAAI;AAAA,IACJ,MAAM;AAAA;AAAA,IACN,aACE;AAAA,IACF,UAAU;AAAA,IACV,QAAQ;AAAA,IACR,WAAW,EAAE,MAAM,SAAS;AAAA,IAC5B,QAAQ,EAAE,MAAM,YAAY,KAAK,iBAAiB;AAAA,EACpD;AAAA;AAAA,EAGA;AAAA,IACE,IAAI;AAAA,IACJ,MAAM;AAAA;AAAA,IACN,aAAa;AAAA,IACb,UAAU;AAAA,IACV,QAAQ;AAAA,IACR,WAAW,EAAE,MAAM,SAAS;AAAA,IAC5B,QAAQ;AAAA,MACN,MAAM;AAAA,MACN,aAAa;AAAA,MACb,UAAU;AAAA,IACZ;AAAA,EACF;AAAA,EACA;AAAA,IACE,IAAI;AAAA,IACJ,MAAM;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA,IAKN,aAAa;AAAA,IACb,UAAU;AAAA,IACV,QAAQ;AAAA,IACR,WAAW,EAAE,MAAM,SAAS;AAAA,IAC5B,QAAQ;AAAA,MACN,MAAM;AAAA,MACN,aAAa;AAAA,MACb,UAAU;AAAA,IACZ;AAAA,EACF;AAAA,EACA;AAAA;AAAA;AAAA;AAAA,IAIE,IAAI;AAAA,IACJ,MAAM;AAAA;AAAA,IACN,aAAa;AAAA,IACb,UAAU;AAAA,IACV,QAAQ;AAAA,IACR,WAAW,EAAE,MAAM,SAAS;AAAA,IAC5B,QAAQ;AAAA,MACN,MAAM;AAAA,MACN,aAAa;AAAA,MACb,UAAU;AAAA,IACZ;AAAA,EACF;AAAA,EACA;AAAA;AAAA,IAEE,IAAI;AAAA,IACJ,MAAM;AAAA;AAAA,IACN,aAAa;AAAA,IACb,UAAU;AAAA,IACV,QAAQ;AAAA,IACR,WAAW,EAAE,MAAM,SAAS;AAAA,IAC5B,QAAQ,EAAE,MAAM,OAAO,KAAK,wBAAwB,SAAS,QAAQ;AAAA,EACvE;AAAA,EACA;AAAA;AAAA;AAAA,IAGE,IAAI;AAAA,IACJ,MAAM;AAAA;AAAA,IACN,aAAa;AAAA,IACb,UAAU;AAAA,IACV,QAAQ;AAAA,IACR,WAAW,EAAE,MAAM,SAAS;AAAA,IAC5B,QAAQ;AAAA,MACN,MAAM;AAAA,MACN,KAAK;AAAA,MACL,SAAS;AAAA,MACT,MAAM,CAAC,WAAW,eAAe,KAAK,WAAW,eAAe,OAAO;AAAA,IACzE;AAAA,EACF;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA,EAOA;AAAA,IACE,IAAI;AAAA,IACJ,MAAM;AAAA;AAAA,IACN,aAAa;AAAA,IACb,UAAU;AAAA,IACV,QAAQ;AAAA,IACR,WAAW,EAAE,MAAM,aAAa,QAAQ,eAAe;AAAA,IACvD,QAAQ;AAAA,MACN,MAAM;AAAA,MACN,aAAa;AAAA,MACb,UAAU;AAAA,IACZ;AAAA,EACF;AAAA;AAAA,EAGA;AAAA,IACE,IAAI;AAAA,IACJ,MAAM;AAAA;AAAA,IACN,aAAa;AAAA,IACb,UAAU;AAAA,IACV,QAAQ;AAAA,IACR,WAAW,EAAE,MAAM,aAAa,QAAQ,CAAC,gBAAgB,MAAM,EAAE;AAAA,IACjE,QAAQ,EAAE,MAAM,OAAO,KAAK,UAAU,SAAS,UAAU;AAAA,EAC3D;AAAA,EACA;AAAA,IACE,IAAI;AAAA,IACJ,MAAM;AAAA;AAAA,IACN,aAAa;AAAA,IACb,UAAU;AAAA,IACV,QAAQ;AAAA,IACR,WAAW,EAAE,MAAM,aAAa,QAAQ,CAAC,gBAAgB,MAAM,EAAE;AAAA,IACjE,QAAQ,EAAE,MAAM,OAAO,KAAK,eAAe,SAAS,SAAS;AAAA,EAC/D;AAAA,EACA;AAAA,IACE,IAAI;AAAA,IACJ,MAAM;AAAA;AAAA,IACN,aAAa;AAAA,IACb,UAAU;AAAA,IACV,QAAQ;AAAA,IACR,WAAW,EAAE,MAAM,aAAa,QAAQ,CAAC,gBAAgB,MAAM,EAAE;AAAA,IACjE,QAAQ,EAAE,MAAM,OAAO,KAAK,YAAY,SAAS,UAAU;AAAA,EAC7D;AAAA;AAAA,EAGA;AAAA,IACE,IAAI;AAAA,IACJ,MAAM;AAAA;AAAA,IACN,aACE;AAAA,IACF,UAAU;AAAA,IACV,QAAQ;AAAA,IACR,WAAW,EAAE,MAAM,aAAa,QAAQ,iBAAiB;AAAA;AAAA,IAEzD,QAAQ,EAAE,MAAM,SAAS,QAAQ,sBAAsB,OAAO,aAAa;AAAA,EAC7E;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA,EAMA;AAAA,IACE,IAAI;AAAA,IACJ,MAAM;AAAA;AAAA,IACN,aACE;AAAA,IACF,UAAU;AAAA,IACV,QAAQ;AAAA,IACR,WAAW,EAAE,MAAM,gBAAgB;AAAA,IACnC,QAAQ;AAAA,MACN,MAAM;AAAA,MACN,aAAa;AAAA,MACb,UAAU;AAAA,IACZ;AAAA,EACF;AAAA;AAAA,EAGA;AAAA,IACE,IAAI;AAAA,IACJ,MAAM;AAAA;AAAA,IACN,aAAa;AAAA,IACb,UAAU;AAAA,IACV,QAAQ;AAAA,IACR,WAAW,EAAE,MAAM,gBAAgB;AAAA;AAAA,IAEnC,QAAQ;AAAA,MACN,MAAM;AAAA,MACN,QAAQ;AAAA,MACR,OAAO;AAAA,IACT;AAAA,EACF;AAAA,EACA;AAAA,IACE,IAAI;AAAA,IACJ,MAAM;AAAA;AAAA,IACN,aAAa;AAAA,IACb,UAAU;AAAA,IACV,QAAQ;AAAA,IACR,WAAW,EAAE,MAAM,gBAAgB;AAAA,IACnC,QAAQ,EAAE,MAAM,SAAS,QAAQ,sBAAsB,OAAO,cAAc;AAAA,EAC9E;AAAA,EACA;AAAA,IACE,IAAI;AAAA,IACJ,MAAM;AAAA;AAAA,IACN,aACE;AAAA;AAAA,IAEF,UAAU;AAAA,IACV,QAAQ;AAAA,IACR,WAAW,EAAE,MAAM,gBAAgB;AAAA,IACnC,QAAQ,EAAE,MAAM,OAAO,KAAK,iBAAiB,SAAS,SAAS;AAAA,EACjE;AAAA;AAAA,EAEA;AAAA,IACE,IAAI;AAAA,IACJ,MAAM;AAAA;AAAA,IACN,aACE;AAAA,IACF,UAAU;AAAA,IACV,QAAQ;AAAA,IACR,WAAW,EAAE,MAAM,SAAS;AAAA,IAC5B,QAAQ;AAAA,MACN,MAAM;AAAA,MACN,aAAa;AAAA,MACb,UAAU;AAAA,IACZ;AAAA,EACF;AAAA,EACA;AAAA,IACE,IAAI;AAAA,IACJ,MAAM;AAAA;AAAA,IACN,aACE;AAAA,IACF,UAAU;AAAA,IACV,QAAQ;AAAA,IACR,WAAW,EAAE,MAAM,SAAS;AAAA,IAC5B,QAAQ;AAAA,MACN,MAAM;AAAA,MACN,aAAa;AAAA,MACb,UAAU;AAAA,IACZ;AAAA,EACF;AAAA,EACA;AAAA,IACE,IAAI;AAAA,IACJ,MAAM;AAAA;AAAA,IACN,aACE;AAAA,IACF,UAAU;AAAA,IACV,QAAQ;AAAA,IACR,WAAW,EAAE,MAAM,SAAS;AAAA,IAC5B,QAAQ;AAAA,MACN,MAAM;AAAA,MACN,aAAa;AAAA,MACb,UAAU;AAAA,IACZ;AAAA,EACF;AAAA;AAAA;AAAA,EAIA;AAAA,IACE,IAAI;AAAA,IACJ,MAAM;AAAA;AAAA,IACN,aACE;AAAA,IACF,UAAU;AAAA,IACV,QAAQ;AAAA,IACR,WAAW,EAAE,MAAM,SAAS;AAAA,IAC5B,QAAQ;AAAA,MACN,MAAM;AAAA,MACN,aAAa;AAAA,MACb,UAAU;AAAA,IACZ;AAAA,EACF;AAAA,EACA;AAAA,IACE,IAAI;AAAA,IACJ,MAAM;AAAA;AAAA,IACN,aAAa;AAAA,IACb,UAAU;AAAA,IACV,QAAQ;AAAA;AAAA,IAER,WAAW,EAAE,MAAM,SAAS;AAAA,IAC5B,QAAQ,EAAE,MAAM,SAAS,QAAQ,4BAA4B,OAAO,aAAa;AAAA,EACnF;AAAA,EACA;AAAA,IACE,IAAI;AAAA,IACJ,MAAM;AAAA;AAAA,IACN,aACE;AAAA,IACF,UAAU;AAAA,IACV,QAAQ;AAAA,IACR,WAAW,EAAE,MAAM,SAAS;AAAA,IAC5B,QAAQ,EAAE,MAAM,SAAS,QAAQ,4BAA4B,OAAO,mBAAmB;AAAA,EACzF;AAAA,EACA;AAAA,IACE,IAAI;AAAA,IACJ,MAAM;AAAA;AAAA,IACN,aACE;AAAA,IACF,UAAU;AAAA,IACV,QAAQ;AAAA,IACR,WAAW,EAAE,MAAM,SAAS;AAAA,IAC5B,QAAQ;AAAA,MACN,MAAM;AAAA,MACN,aAAa;AAAA,MACb,UAAU;AAAA,IACZ;AAAA,EACF;AAAA,EACA;AAAA;AAAA;AAAA;AAAA,IAIE,IAAI;AAAA,IACJ,MAAM;AAAA;AAAA,IACN,aACE;AAAA,IACF,UAAU;AAAA,IACV,QAAQ;AAAA,IACR,WAAW,EAAE,MAAM,SAAS;AAAA,IAC5B,QAAQ,EAAE,MAAM,SAAS,QAAQ,uBAAuB,OAAO,0BAA0B;AAAA,EAC3F;AAAA;AAAA,EAEA;AAAA,IACE,IAAI;AAAA,IACJ,MAAM;AAAA;AAAA,IACN,aACE;AAAA,IACF,UAAU;AAAA,IACV,QAAQ;AAAA,IACR,WAAW,EAAE,MAAM,SAAS;AAAA,IAC5B,QAAQ,EAAE,MAAM,SAAS,QAAQ,sBAAsB,OAAO,aAAa;AAAA,EAC7E;AAAA,EACA;AAAA;AAAA,IAEE,IAAI;AAAA,IACJ,MAAM;AAAA;AAAA,IACN,aACE;AAAA,IACF,UAAU;AAAA,IACV,QAAQ;AAAA,IACR,WAAW,EAAE,MAAM,SAAS;AAAA,IAC5B,QAAQ,EAAE,MAAM,SAAS,QAAQ,uBAAuB,OAAO,iBAAiB;AAAA,EAClF;AAAA,EACA;AAAA,IACE,IAAI;AAAA,IACJ,MAAM;AAAA;AAAA,IACN,aACE;AAAA,IACF,UAAU;AAAA,IACV,QAAQ;AAAA,IACR,WAAW,EAAE,MAAM,SAAS;AAAA,IAC5B,QAAQ,EAAE,MAAM,SAAS,QAAQ,0BAA0B,OAAO,yBAAyB;AAAA,EAC7F;AAAA,EACA;AAAA,IACE,IAAI;AAAA,IACJ,MAAM;AAAA;AAAA,IACN,aACE;AAAA,IACF,UAAU;AAAA,IACV,QAAQ;AAAA,IACR,WAAW,EAAE,MAAM,SAAS;AAAA,IAC5B,QAAQ,EAAE,MAAM,SAAS,QAAQ,6BAA6B,OAAO,WAAW;AAAA,EAClF;AAAA,EACA;AAAA,IACE,IAAI;AAAA,IACJ,MAAM;AAAA;AAAA,IACN,aACE;AAAA,IACF,UAAU;AAAA,IACV,QAAQ;AAAA,IACR,WAAW,EAAE,MAAM,gBAAgB;AAAA,IACnC,QAAQ;AAAA,MACN,MAAM;AAAA,MACN,QAAQ;AAAA,MACR,OAAO;AAAA,IACT;AAAA,EACF;AAAA;AAAA,EAGA;AAAA,IACE,IAAI;AAAA,IACJ,MAAM;AAAA;AAAA,IACN,aACE;AAAA,IACF,UAAU;AAAA,IACV,QAAQ;AAAA,IACR,WAAW,EAAE,MAAM,aAAa,QAAQ,CAAC,gBAAgB,MAAM,EAAE;AAAA,IACjE,QAAQ;AAAA,MACN,MAAM;AAAA,MACN,aAAa;AAAA,MACb,UAAU;AAAA,IACZ;AAAA,EACF;AAAA,EACA;AAAA,IACE,IAAI;AAAA,IACJ,MAAM;AAAA;AAAA,IACN,aACE;AAAA,IACF,UAAU;AAAA,IACV,QAAQ;AAAA,IACR,WAAW,EAAE,MAAM,aAAa,QAAQ,CAAC,gBAAgB,MAAM,EAAE;AAAA,IACjE,QAAQ;AAAA,MACN,MAAM;AAAA,MACN,aAAa;AAAA,MACb,UAAU;AAAA,IACZ;AAAA,EACF;AAAA;AAAA;AAAA;AAAA,EAKA;AAAA,IACE,IAAI;AAAA,IACJ,MAAM;AAAA;AAAA,IACN,aACE;AAAA,IACF,UAAU;AAAA,IACV,QAAQ;AAAA,IACR,WAAW,EAAE,MAAM,aAAa,QAAQ,oBAAoB;AAAA,IAC5D,QAAQ;AAAA,MACN,MAAM;AAAA,MACN,QAAQ;AAAA;AAAA;AAAA,MAGR,OAAO;AAAA,IACT;AAAA,EACF;AAAA,EACA;AAAA,IACE,IAAI;AAAA,IACJ,MAAM;AAAA;AAAA,IACN,aAAa;AAAA,IACb,UAAU;AAAA,IACV,QAAQ;AAAA,IACR,WAAW,EAAE,MAAM,aAAa,QAAQ,oBAAoB;AAAA;AAAA,IAE5D,QAAQ,EAAE,MAAM,SAAS,QAAQ,aAAa,OAAO,SAAS;AAAA,EAChE;AAAA,EACA;AAAA,IACE,IAAI;AAAA,IACJ,MAAM;AAAA;AAAA,IACN,aACE;AAAA,IACF,UAAU;AAAA,IACV,QAAQ;AAAA,IACR,WAAW,EAAE,MAAM,aAAa,QAAQ,oBAAoB;AAAA,IAC5D,QAAQ;AAAA,MACN,MAAM;AAAA,MACN,QAAQ;AAAA,MACR,OAAO;AAAA,IACT;AAAA,EACF;AAAA;AAAA,EAEA;AAAA,IACE,IAAI;AAAA,IACJ,MAAM;AAAA;AAAA,IACN,aAAa;AAAA,IACb,UAAU;AAAA,IACV,QAAQ;AAAA,IACR,WAAW,EAAE,MAAM,aAAa,QAAQ,CAAC,aAAa,MAAM,EAAE;AAAA,IAC9D,QAAQ;AAAA,MACN,MAAM;AAAA,MACN,aAAa;AAAA,MACb,UAAU;AAAA,IACZ;AAAA,EACF;AAAA;AAAA;AAAA;AAAA,EAIA;AAAA,IACE,IAAI;AAAA,IACJ,MAAM;AAAA;AAAA,IACN,aAAa;AAAA,IACb,UAAU;AAAA,IACV,QAAQ;AAAA,IACR,WAAW,EAAE,MAAM,aAAa,QAAQ,CAAC,aAAa,MAAM,EAAE;AAAA,IAC9D,QAAQ;AAAA,MACN,MAAM;AAAA,MACN,aAAa;AAAA,MACb,UAAU;AAAA,IACZ;AAAA,EACF;AAAA,EACA;AAAA,IACE,IAAI;AAAA,IACJ,MAAM;AAAA;AAAA,IACN,aAAa;AAAA,IACb,UAAU;AAAA,IACV,QAAQ;AAAA;AAAA,IAER,WAAW,EAAE,MAAM,aAAa,QAAQ,CAAC,aAAa,QAAQ,kBAAkB,EAAE;AAAA,IAClF,QAAQ;AAAA,MACN,MAAM;AAAA,MACN,aAAa;AAAA,MACb,UAAU;AAAA,IACZ;AAAA,EACF;AAAA,EACA;AAAA,IACE,IAAI;AAAA,IACJ,MAAM;AAAA;AAAA,IACN,aAAa;AAAA,IACb,UAAU;AAAA,IACV,QAAQ;AAAA,IACR,WAAW,EAAE,MAAM,aAAa,QAAQ,CAAC,aAAa,MAAM,EAAE;AAAA,IAC9D,QAAQ;AAAA,MACN,MAAM;AAAA,MACN,aAAa;AAAA,MACb,UAAU;AAAA,IACZ;AAAA,EACF;AAAA;AAAA;AAAA,EAIA;AAAA,IACE,IAAI;AAAA,IACJ,MAAM;AAAA;AAAA,IACN,aACE;AAAA,IACF,UAAU;AAAA,IACV,QAAQ;AAAA,IACR,WAAW,EAAE,MAAM,aAAa,QAAQ,CAAC,oBAAoB,EAAE;AAAA,IAC/D,QAAQ;AAAA,MACN,MAAM;AAAA,MACN,aAAa;AAAA,MACb,UAAU;AAAA,IACZ;AAAA,EACF;AAAA;AAAA;AAAA,EAGA;AAAA,IACE,IAAI;AAAA,IACJ,MAAM;AAAA;AAAA,IACN,aAAa;AAAA,IACb,UAAU;AAAA,IACV,QAAQ;AAAA,IACR,WAAW,EAAE,MAAM,aAAa,QAAQ,CAAC,GAAG,kBAAkB,EAAE;AAAA,IAChE,QAAQ;AAAA,MACN,MAAM;AAAA,MACN,aAAa;AAAA,MACb,UAAU;AAAA,IACZ;AAAA,EACF;AAAA;AAAA;AAAA,EAIA;AAAA,IACE,IAAI;AAAA,IACJ,MAAM;AAAA;AAAA,IACN,aACE;AAAA,IACF,UAAU;AAAA,IACV,QAAQ;AAAA,IACR,WAAW,EAAE,MAAM,aAAa,QAAQ,CAAC,kBAAkB,EAAE;AAAA,IAC7D,QAAQ;AAAA,MACN,MAAM;AAAA,MACN,aAAa;AAAA,MACb,UAAU;AAAA,IACZ;AAAA,EACF;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA,EAQA;AAAA,IACE,IAAI;AAAA,IACJ,MAAM;AAAA;AAAA,IACN,aAAa;AAAA,IACb,UAAU;AAAA,IACV,QAAQ;AAAA,IACR,WAAW,EAAE,MAAM,SAAS;AAAA,IAC5B,QAAQ;AAAA,MACN,MAAM;AAAA,MACN,aAAa;AAAA,MACb,UAAU;AAAA,IACZ;AAAA,EACF;AAAA;AAAA;AAAA;AAAA,EAIA;AAAA,IACE,IAAI;AAAA,IACJ,MAAM;AAAA;AAAA,IACN,aAAa;AAAA,IACb,UAAU;AAAA,IACV,QAAQ;AAAA,IACR,WAAW,EAAE,MAAM,aAAa,QAAQ,CAAC,kBAAkB,EAAE;AAAA,IAC7D,QAAQ;AAAA,MACN,MAAM;AAAA,MACN,aAAa;AAAA,MACb,UAAU;AAAA,IACZ;AAAA,EACF;AAAA;AAAA;AAAA;AAAA,EAKA;AAAA,IACE,IAAI;AAAA,IACJ,MAAM;AAAA;AAAA,IACN,aACE;AAAA,IACF,UAAU;AAAA,IACV,QAAQ;AAAA,IACR,WAAW,EAAE,MAAM,gBAAgB;AAAA,IACnC,QAAQ;AAAA,MACN,MAAM;AAAA,MACN,aAAa;AAAA,MACb,UAAU;AAAA,IACZ;AAAA,EACF;AAAA;AAAA,EAGA;AAAA;AAAA;AAAA;AAAA;AAAA,IAKE,IAAI;AAAA,IACJ,MAAM;AAAA;AAAA,IACN,aAAa;AAAA,IACb,UAAU;AAAA,IACV,QAAQ;AAAA,IACR,WAAW,EAAE,MAAM,SAAS;AAAA,IAC5B,QAAQ;AAAA,MACN,MAAM;AAAA,MACN,aAAa;AAAA,MACb,UAAU;AAAA,IACZ;AAAA,EACF;AAAA,EACA;AAAA,IACE,IAAI;AAAA,IACJ,MAAM;AAAA;AAAA,IACN,aAAa;AAAA,IACb,UAAU;AAAA,IACV,QAAQ;AAAA,IACR,WAAW,EAAE,MAAM,SAAS;AAAA;AAAA;AAAA;AAAA,IAI5B,QAAQ;AAAA,MACN,MAAM;AAAA,MACN,aAAa;AAAA,MACb,UAAU;AAAA,IACZ;AAAA,EACF;AAAA,EACA;AAAA,IACE,IAAI;AAAA,IACJ,MAAM;AAAA;AAAA,IACN,aACE;AAAA,IACF,UAAU;AAAA,IACV,QAAQ;AAAA,IACR,WAAW,EAAE,MAAM,UAAU,MAAM,YAAY;AAAA,IAC/C,QAAQ;AAAA,MACN,MAAM;AAAA,MACN,QAAQ;AAAA,MACR,MAAM,CAAC,WAAW,SAAS;AAAA,IAC7B;AAAA,EACF;AACF;AAOO,IAAM,uBAA8C;AAAA,EACzD;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA;AAAA,EAEA;AACF;AASO,IAAM,6BAAoD;AAAA,EAC/D,GAAG;AAAA;AAAA,EAEH;AACF;AAOO,IAAM,aAAwC,OAAO;AAAA,EAC1D,gBAAgB,IAAI,CAAC,MAAM,CAAC,EAAE,IAAI,EAAE,IAAI,CAAC;AAC3C;AAGO,SAAS,eAAe,SAA4B;AACzD,SAAO,WAAW,OAAO,KAAK;AAChC;AAiBO,SAAS,mBACd,OACA,KAKS;AAET,MAAI,IAAI,cAAc,aAAa,SAAS,MAAM,EAAE,EAAG,QAAO;AAC9D,MAAI,IAAI,cAAc,aAAa,SAAS,MAAM,EAAE,EAAG,QAAO;AAM9D,MAAI,WAAW,MAAM,EAAE,MAAM,eAAgB,QAAO;AACpD,SAAO,iBAAiB,OAAO,GAAG;AACpC;AAMA,SAAS,iBACP,OACA,KACS;AACT,QAAM,OAAO,MAAM;AACnB,UAAQ,KAAK,MAAM;AAAA,IACjB,KAAK;AACH,aAAO,IAAI,OAAO,KAAK,CAAC,MAAM,KAAK,OAAO,SAAS,CAAC,CAAC;AAAA,IACvD,KAAK;AAGH,aAAO,YAAY,IAAI,MAAM;AAAA,IAC/B,KAAK;AACH,aAAO,IAAI,QAAQ,KAAK,IAAI,MAAM;AAAA,IACpC,KAAK;AAGH,aAAO;AAAA,EACX;AACF;AAQO,SAAS,gBACd,SACA,KAKS;AACT,QAAM,QAAQ,gBAAgB,KAAK,CAAC,MAAM,EAAE,OAAO,OAAO;AAC1D,SAAO,QAAQ,mBAAmB,OAAO,GAAG,IAAI;AAClD;AAOO,SAAS,4BAA4B,KAIX;AAC/B,SAAO,gBAAgB;AAAA,IACrB,CAAC,MACC,WAAW,EAAE,EAAE,MAAM,kBACrB,CAAC,IAAI,cAAc,aAAa,SAAS,EAAE,EAAE,KAC7C,iBAAiB,GAAG,GAAG;AAAA,EAC3B;AACF;AAKO,SAAS,uBACd,QACA,KAK8B;AAC9B,SAAO,OAAO,OAAO,CAAC,MAAM,mBAAmB,GAAG,GAAG,CAAC;AACxD;;;AEnmCA;AAUO,IAAM,aAAa;AAAA,EACxB;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AACF;AAGO,IAAM,kBAA4C;AAAA,EACvD,UAAU;AAAA,EACV,SAAS;AAAA,EACT,MAAM;AAAA,EACN,UAAU;AAAA,EACV,aAAa;AAAA;AAAA,EAEb,eAAe;AAAA;AAAA,EAEf,gBAAgB;AAAA,EAChB,UAAU;AAAA,EACV,aAAa;AACf;","names":[]}