Skip to content

AGENTS.md

Latest commit

 

History

History
401 lines (296 loc) · 35 KB

AGENTS.md

File metadata and controls

401 lines (296 loc) · 35 KB

AGENTS.md

Cross-tool agent configuration. Read by: Claude Code, Codex, Copilot, Cursor, Windsurf, Amp, Devin, Continue. Tool-specific files (CLAUDE.md, .cursorrules, copilot-instructions.md) layer on top for advanced features.

Git Workflow — DIRECT TO MAIN (read this first)

Every surface commits straight to main. No pull requests. No feature branches. (F.089) This is a single shared-tree, multi-agent repo — PRs and feat/* branches create bottlenecks and branch sprawl, not safety.

  • main is NOT branch-protected. git push origin main works from every environment — desktop, hardware, and mobile/cloud Linux agents. Use it.
  • Mobile / cloud Claude Code & Codex agents: do NOT use your default remote-agent PR flow. Commit and git push origin main directly. Branching + opening a PR is the wrong default here.
  • Explicit paths only: git add <file> — NEVER git add -A or . (you will snatch peer in-flight files in the shared tree).
  • On Windows desktop, prefer .\scripts\safe-commit.ps1 (uses git commit --only to avoid the multi-agent index race).
  • Railway auto-deploys on push to main — validate locally before pushing.
  • Fallback only: if your environment genuinely cannot push direct and forces a PR, it is reaped (squash-merged + branch deleted) by HoloCI's pr-reaper on its schedule — do not wait on it, and do not open PRs by choice.

CI is HoloCI, not GitHub Actions. Validation runs on the vast.ai fleet via the mcp-orchestrator queue and reports through free GitHub commit statuses (~/.ai-ecosystem/scripts/holo-ci/; triggered by the pre-push hook, with reconcile.mjs as the scheduled floor and pr-reaper.mjs as the stray-PR safety net). GitHub Actions is billing-locked and unused — do not add .github/workflows/*.

What This Project Does

HoloScript turns descriptions into working interfaces. Describe a dashboard, a robot arm, or a VR room — it runs on your screen, in a headset, or on a hologram. Write .hs, .hsplus, or .holo — the runtime handles the rest. If a compiler exists for your target platform, it optimizes automatically. If not, runtime interpretation keeps it working.

MCP server at mcp.holoscript.net — discover tools via POST /mcp with tools/list.

First Surface: Skills

Before raw grep, raw curl, or long-form code spelunking, check the skill surface. Skills are the front door for HoloScript work because they carry the current workflows, dispatch rules, gotchas, and validation habits.

  • Claude Code: invoke the matching skill directly.
  • Codex, Copilot, Cursor, Gemini-Antigravity, Windsurf, Devin: read the relevant SKILL.md directly. The ai-ecosystem inventory is C:/Users/josep/.ai-ecosystem/SKILL_MAP.md.
  • If no skill matches, use MCP discovery, Absorb Graph RAG, NORTH_STAR.md, and local commands.

Skipping a matching skill is a workflow bug. Common examples: use Absorb or /codebase for codebase intelligence, /compile for compile-target work, /frontend for UI audits, /critic for hard review, and /room for HoloMesh coordination.

Mechanism-design target: skills should be laid out like a payoff matrix where the right skill is the dominant move. See C:/Users/josep/.ai-ecosystem/docs/handbooks/skill-routing-game-theory.md.

Tracking source map: before trusting board exports, root JSON files, IDE session stores, backup clones, or worktree folders as current state, classify them with C:/Users/josep/.ai-ecosystem/docs/handbooks/agent-tracking-source-map.md. Live board, git status, and the knowledge store win.

File Formats

Extension Purpose When to use
.hs Data pipelines, simple scenes, configuration Structured data that compiles to any target
.hsplus Behaviors, agents, economics, IoT, physics When you need traits like @grabbable @physics @spatial_audio
.holo Full compositions, cross-platform scenes AI-generated scenes, multi-object layouts, dashboards
.ts/.tsx Tooling, infrastructure, tests TypeScript for the platform itself (not user content)

Build and Test

pnpm install                              # Install dependencies
pnpm build                                # Build (core first, then rest — order matters)
pnpm --filter @holoscript/core build      # Build specific package
pnpm --filter @holoscript/net-service run build  # Authoritative holoscript.net service build
pnpm test                                 # Run all tests (vitest)
pnpm --filter @holoscript/core test       # Test specific package
pnpm lint                                 # ESLint
pnpm format                               # Prettier
pnpm bench                                # Benchmarks
pnpm run health:deps                      # Bounded pnpm audit; emits JSON pass/fail/cached/skip

Build order matters: @holoscript/core must build before any downstream package.

Package Structure

packages/
  core/               # Parser, AST, traits, compilers (ALL live here)
  mcp-server/         # MCP tools (Streamable HTTP transport)
  cli/                # CLI: holoscript / hs binary
  runtime/            # Direct interpretation (no compiler needed)
  studio/             # Next.js creation environment
  engine/             # Rendering, physics, animation, ECS
  framework/          # Agent orchestration, board, economy
  r3f-renderer/       # React Three Fiber components
  lsp/                # Language Server Protocol
  connectors/         # GitHub, Railway, Docker connectors
  plugins/            # Domain plugins (banking, neuroscience, film, etc. — count via `ls packages/plugins/`)
  snn-webgpu/         # GPU spiking neural networks
  ...                 # More — run `ls packages/` for full list

services/
  export-api/         # Export/rendering API
  holoscript-net/     # Production web service
  llm-service/        # LLM proxy service

Code Conventions

Native-first — read this before the hygiene rules below. These conventions govern platform-tooling code only (the CLI, parser, compilers, adapters, and tests written in TypeScript). They are NOT how you author HoloScript content or behavior. A trait, a render surface, an agent brain, or a compile target is authored as declarative data a tool consumes — not hand-written imperative TypeScript — and its correctness is enforced by a structural gate, never asserted (the two principles: behavior-as-data; gate-enforced/derived correctness). Authoring those the pretrained TS/React way produces code that passes a generic lint but is structurally wrong here: invisible to the compiler, unverifiable by provenance, or uncompilable to non-TS targets. Use the File Format table above (.hs/.hsplus/.holo vs .ts/.tsx) to decide what you're writing, and before authoring any HoloScript content read docs/handbooks/holoscript-native-authoring-vs-pretrained.md — the surface→authoring map is the table at its top. The rules below are the floor for tooling, not the shape of HoloScript.

TypeScript-tooling hygiene (platform/tooling code only):

  • TypeScript strict mode: strict: true, target ES2020, ESNext modules, bundler resolution
  • No any: Use unknown. This is enforced.
  • Test framework: vitest (never Jest)
  • Build tool: tsup
  • Package manager: pnpm with workspaces
  • Node version: >= 18.0.0
  • JSX: hand-written .tsx is permitted for tooling/CLI/parser/adapter code ONLY — and there it MUST use the .tsx extension. Perceivable render/UI surfaces are NEVER hand-written .tsx: author them as .holo and let a compiler @generate the .tsx (enforced LIVE by scripts/holo-ci/check-render-surface-native.mjs — a non-generated .tsx under a render root is SURFACE-GREW, exit 1, commit blocked, in both pre-commit and HoloCI).
  • Types: dist/index.d.ts is hand-crafted by scripts/generate-types.mjs (not tsc). New exports require updating BOTH src/index.ts AND the mainDTS template in generate-types.mjs.

Compilers

All compilers live in @holoscript/core. Count via find packages/core/src -name "*Compiler.ts" -not -name "CompilerBase*" -not -name "*.test.*".

Each compiler extends CompilerBase and requires RBAC authorization:

// Required mock for ALL compiler tests
vi.mock('../../security/rbac', () => ({
  checkPermission: vi.fn().mockResolvedValue(true),
}));

const result = await compiler.compile(source, 'test-token');

Traits

All traits live in @holoscript/core/src/traits/. Count via find packages/core/src/traits -name "*.ts" -not -name "*.test.*". Category files in traits/constants/ (118 files — verify via ls).

Categories span far beyond spatial:

Domain Example categories
Spatial/XR core-vr-interaction, physics-expansion, locomotion-movement, xr-platform, spatial-algorithms
Rendering rendering, post-processing, global-illumination, visual-effects, volumetric-webgpu, vfx-audio
Characters humanoid-avatar, facial-expression, character-pipeline, character-materials, npc-roles, creatures-mythical
Environment atmosphere-sky, terrain-ocean, weather-phenomena, weather-particles, environmental-biome, nature-life
AI/ML intelligence-behavior, ml-inference, ml-tensor, networking-ai, iot-autonomous-agents
Data/Services data-pipeline, data-storage, database-persistence, api-gateway, search, file-storage
Business payment, social-commerce, enterprise-multitenancy, workflow-bpm, compliance-governance, audit-trail
Industry robotics-industrial, healthcare-medical, scientific-computing, construction-building, maritime-naval
Security auth-identity, security-crypto, safety-boundaries, feature-flags
DevOps devops-ci, testing-qa, observability, analytics-observability, containers-storage
Creative music-performance, narrative-storytelling, procedural-generation, magic-fantasy, cooking-food
Communication communication, notification-alerting, signs-communication, media-content

Adding a new trait:

  1. First run /stub-audit — many trait names already exist with a correct seam but a placeholder body (Pattern B stub). The native move is to wire+build the existing name, not author a parallel duplicate that leaves the original advertised-but-dead.
  2. Author the runtime behavior native — the trait's public seam is a plain TraitHandler<TConfig> object literal (onAttach/onUpdate/onEvent/onDetach), NOT a class with methods. Per-instance state lives on node.__<name>State (created in onAttach, deleted in onDetach) — never class fields/module vars, or every node sharing the handler shares state. Traits never call each other: emit an event via context.emit(...). Multi-phase behavior is the @state_machine decorator, not if/else in onUpdate. Behavioral traits are authored in .hsplus. (Full shape + file:line evidence: docs/handbooks/holoscript-native-authoring-vs-pretrained.md § Traits.)
  3. Define constant in packages/core/src/traits/constants/
  4. Add visual preset in packages/core/src/traits/visual/presets/
  5. Register in the category index
  6. Add R3F handler if it has rendering (R3FCompiler.ts)
  7. Add tests

MCP Server

Production: https://mcp.holoscript.net Local: npx tsx packages/mcp-server/src/index.ts Tool discovery: POST /mcp → tools/list (tool count changes with deploys — never hardcode) Health: GET /health → tools field for current count

Tool categories: parsing, traits, generation, codebase intelligence, compilation, IDE, browser control, networking, self-improvement. Discover specific tools via MCP protocol, not from docs.

Codebase Intelligence

Cache at ~/.holoscript/graph-cache.json (24h TTL). Always check freshness first:

  1. holo_graph_status — is cache fresh?
  2. holo_absorb_repo — scan (fast from cache, ~3-10s fresh)
  3. holo_query_codebase — architectural Q&A
  4. holo_impact_analysis — blast radius for changes

Never call holo_absorb_repo with force: true unless holo_graph_status reports stale.

Testing

Vitest Mock Rules

  • Use vi.hoisted() for mock variables
  • Use function(){} (not arrow functions) for mock constructors
  • GPU tests use Dawn WebGPU with mock fallback (see core/src/physics/__tests__/gpu-setup.ts)

Pre-Commit

  • Run pnpm test before committing
  • Run pnpm lint for style issues
  • All tests must pass

Numbers Policy

Never hardcode ecosystem counts (tools, compilers, traits, tests, plugins, packages) in docs, configs, or code comments. They change with every deploy.

  • SSOT: docs/NUMBERS.md — verification commands for every metric
  • In docs: reference the command or link to NUMBERS.md
  • In code comments: say "verify via find *Compiler.ts" not "44 compilers"
  • MCP tools: discover via tools/list, verify via GET /health → tools field
  • Compilers: find packages/core/src -name "*Compiler.ts" -not -name "CompilerBase*" -not -name "*.test.*"
  • Traits: find packages/core/src/traits -name "*.ts" -not -name "*.test.*"

Git Rules

  • NEVER use git add -A or git add . — stage files explicitly: git add path/to/file.ts
  • Commit message format: conventional commits (feat:, fix:, test:, docs:, refactor:, chore:)
  • Large batches (10+ files): split into sectioned commits by topic
  • Docs must use lowercase filenames

Scratch Output

  • Put one-off command, test, lint, tsc, vitest, API-board, and local investigation output under /.scratch/<YYYY-MM-DD>-<agent-or-task>/.
  • Do not add new root-level ignore patterns for transient logs, dumps, board exports, or ad hoc scripts. Redirect the command output or fix the script to write into /.scratch/.
  • Treat /.scratch/ as ignored, disposable local output. Do not use it for source files, durable fixtures, or evidence that must survive another checkout.
  • Promote durable evidence into the right tracked home, such as docs/, research/, .bench-logs/, package fixtures, or service-specific artifact directories.

Security

StdlibPolicy

All I/O from traits/compositions is gated by StdlibPolicy:

  • allowFileRead, allowFileWrite, allowFileDelete
  • allowProcessExec, allowNetFetch
  • allowMediaDecode, allowDepthInference, allowGpuCompute

Traits MUST NOT perform direct filesystem access. All I/O goes through stdlib BehaviorTree actions with into: convention for blackboard key prefix.

Execution Sandbox

  • vm.createContext (not vm2)
  • Ed25519 cryptographic signatures
  • Per-tool RBAC authorization on MCP server

Simulation Stack (Scientific Computing)

The engine includes PDE-based simulation solvers for thermal, structural, and hydraulic analysis. As of 2026-04-11, the solver math is verified but the product is not accessible to non-developer users.

What EXISTS (validated, tested, committed)

Layer Status Details
Solvers Working, benchmarked Thermal (FDM, Euler+Jacobi), Structural (linear tet FEM, PCG), Hydraulic (Hardy-Cross, Darcy-Weisbach)
Coupling Working 6 coupling chains via CouplingManager (sequential operator splitting)
Materials Cited, T-dependent 15+ materials with Incropera/NIST/ASM citations, uncertainty bounds, piecewise-linear T-dependent lookup for 5 core materials
Units Enforced 30+ branded physical quantity types, 40+ unit conversions, DimensionalMismatchError at runtime
V&V Partial Analytical benchmarks for all 3 domains (steady-state, patch test, Darcy-Weisbach), convergence studies, Richardson extrapolation, GCI
Export Working VTK (StructuredPoints, UnstructuredGrid, PolyData), CSV, JSON metadata
Provenance Working Immutable SimulationRun records, determinism verification, run comparison
Documentation Complete Full mathematical formulations with equations, discretization, limitations, literature refs
Reporting Working V&V report generator (markdown + LaTeX)

Key paths: packages/engine/src/simulation/ (solvers, units, export, provenance, verification)

What DOES NOT EXIST (gaps to "scientists use this")

Gap What's missing Why it matters Effort
Geometry import No CAD import (STEP/IGES), no shape drawing in Studio Scientists need to define their problem geometry Large
Mesh generation No automatic tet mesher — users must provide node/element arrays Nobody hand-writes mesh data Large (integrate TetGen or similar)
Studio UI No simulation tab in Studio — can't configure BCs, materials, loads visually Non-developers can't use the solvers Large
Post-processing in Studio No in-browser visualization of results (color maps, probes, plots) Scientists shouldn't need ParaView for basic inspection Medium
Solver scope Linear only (no plasticity, no turbulence, no dynamics, no contact) Many real-world problems are nonlinear Large per physics model
Mesh quality Uniform grids (thermal), no adaptive refinement, no mesh quality metrics Complex geometries need adaptive meshes Large
V&V depth 3 benchmarks per domain. Credible V&V needs 50+ (NAFEMS, ASME PTC, experimental) Thin validation won't satisfy reviewers Medium (ongoing)
HPC/parallel All solvers are single-threaded JS — millions of DOFs won't run Real scientific problems are large Large (WebGPU/WASM workers)
Code-to-code comparison No comparison against FEniCS, OpenFOAM, EPANET Reviewers expect cross-validation Medium

Bottom line for agents: The engine's simulation math is verified and the trust infrastructure (V&V, provenance, units, export) is solid. But there is no user-facing product yet — no geometry, no meshing, no UI. A TypeScript developer can use the solver APIs directly. A lab researcher cannot. Do NOT represent the simulation stack as "ready for scientists" in docs, pitches, or Moltbook posts without this caveat.

Boundaries

  • ALWAYS: Validate HoloScript files after editing
  • ALWAYS: Run impact analysis on packages/core changes
  • ALWAYS: Check VR frame budget (11.1ms) for render path code
  • ASK FIRST: Modify generate-types.mjs or dist/index.d.ts
  • ASK FIRST: Changes touching 10+ files across packages
  • NEVER: Put ML classifiers in VR render loop
  • NEVER: Delete tests to bypass failures
  • NEVER: Commit secrets or API keys

Cursor & Peer Protocol

If you are a Cursor agent (or any peer agent operating in an asynchronous team environment), follow the HoloMesh Peer Protocol:

  • Ask tools before the founder: Use HoloScript MCP, Absorb (GraphRAG), and orchestrator tools for factual and codebase questions. Use mcp_servers.absorb in ~/.cursor/mcp.json if configured.
  • Board queue: Authoritative open work is GET /api/holomesh/team/{id}/board. A committed board.json (e.g. in the founder’s ~/.ai-ecosystem repo) is only a git snapshot — refresh from the API before treating it as the live queue.
  • Handoffs & Blockers: If blocked, post a handoff message to HoloMesh instead of waiting on the human founder.
  • Marathon loop: Refer to the user's ~/.cursor/skills/room-autonomous/SKILL.md for single-session deep focus.
  • For full details, read docs/TEAM_PEER_PROTOCOL.md and docs/cursor/CURSOR_MCP_ABSORB.md in the .ai-ecosystem repository.

Ecosystem Skills (Read Before Working)

Skills are concentrated knowledge files — the best single-file summary of each subsystem. Read the relevant skill file before working in any domain. They contain architecture, API endpoints, decision trees, and conventions that no other doc has.

Inventory: SKILL_MAP.md in the ai-ecosystem repo (~/.ai-ecosystem/SKILL_MAP.md locally) is the cross-surface canonical list — every skill across user-level, ai-ecosystem plugin, and this HoloScript plugin, with descriptions and absolute paths. Generated by ai-ecosystem/scripts/generate-skill-map.mjs. The workflow tables below are the decision aid (which skill for which task) — SKILL_MAP.md is the authoritative inventory (what exists where).

Audit 2026-04-26 (revised): prior version used ls ~/.claude/skills/ as registration check, which misses plugin-installed skills. Canonical source-of-truth: the available-skills block in the session-start system-reminder for Claude Code, plus SKILL_MAP.md for cross-surface use. Table below grouped by workflow per W.GOLD.341 follow-up — investigate every caveat before propagating, name plugin substitutes explicitly.

Decision & Coordination

Workflow trigger Skill Notes
Founder decision / about to bandaid/workaround/demote/stall for Joseph / 17-paper editorial /founder ✅ Authority order, vision pillars, refusal list
Board ops (read, claim, done, send DMs, switch mode, knowledge sync) /room ✅ HARD RULE — see ⛔ ANTIPATTERN below. Never fabricate task_* IDs
"This task needs expertise we don't have on the team" /find-collaborator ✅ searches mesh, flags gap, posts ready DMs
New agent online (new IDE/model/seat) /onboard ✅ wizard, .env, hooks, first heartbeat

⛔ ANTIPATTERN: raw-curl-for-room-ops (RECURRENT)

Multiple agents across multiple sessions have closed their work with the personal note "I should have used /room instead of raw curl." That's the signal this needs to be a hard rule, not a personal followup.

When you reach for curl -X (POST|PATCH|GET) ".../api/holomesh/team/.../board..." or .../knowledge or .../message or .../presence or .../suggestions — STOP and use /room instead. The skill exists; it handles:

  • Bearer auth resolution — per-surface x402 seat resolves to your real agentId on the server, not the agent_founder fallback that obscures attribution
  • Response shape parsing — the recurrent gotcha is arrow-character encoding crashing inline-Python parsers; the skill wraps the parse
  • Endpoint-shape drift — bare /done 404s; canonical path is /board/done and /board/<TASK_ID> PATCH with action:done. The skill knows; raw curl forces you to rediscover
  • Truncation warnings — task descriptions silently slice at 2000 chars; the skill surfaces the warnings[] array on the 201 response
  • Token cost — ~5k tokens per multi-step session, because the skill runs in forked context and only the result returns to the main window

Why agents drop to curl anyway: build-mode flows feel faster with direct curl visibility. That feeling is wrong. The curl path costs more total tokens because of the parse-then-debug-then-reparse cycle (verified across multiple session retros: every "I should have used /room" note follows a curl-and-debug detour).

The hard rule: if the URL contains /api/holomesh/team/, /knowledge, /message, /presence, or /suggestions, you are in /room territory. Use the skill. One-off read probes for endpoint debugging are the only legitimate exception — and they should be a single call, not the start of a multi-step session.

Knowledge & Synthesis

Workflow trigger Skill Notes
Contradictions across knowledge / gaps in collective understanding /oracle ✅ curates cross-domain connections
Deep research (8-phase uAA2++) /research ✅ was ai-workspace; same protocol
TODO + git + MCP /health + knowledge gaps in one shot /scan ✅ registered to user-level 2026-04-26 from .claude/skills/scan/
End-of-session retrospective /retrospective ✅ productivity patterns, recurring failures
Documentation voice / staleness / version consistency /documenter ✅ agent-first writing
Memory file consolidation anthropic-skills:consolidate-memory ✅ merge duplicates, fix stale facts

Code & Shipping

Workflow trigger Skill Notes
Honest critique (code, architecture, docs, pitch, demos, plans, papers) /critic ✅ registered to user-level 2026-04-26 from .claude/skills/critic/; no silver linings
Code-only PR review (security/performance/correctness) engineering:code-review ✅ or simplify for inline fixes
Structured debugging engineering:debug ✅ reproduce → isolate → diagnose → fix
Pre-deployment checklist engineering:deploy-checklist ✅ CI, migrations, flags, rollback triggers
Production incident response engineering:incident-response ✅ triage + comms + postmortem
Railway deployment orchestration /deploy ✅ health checks, rollback, post-deploy verify
Frontend ops (React/Vue/Angular review, a11y, bundle) /frontend-ops ✅ WCAG 2.1, vitest/jest
Building / shipping HoloScript-specific patterns engineering:standup (commits/PRs) ✅ + .claude/skills/dev/SKILL.md (Read tool) no /dev plugin yet

Codebase Intelligence

Workflow trigger Skill Notes
Find callers / blast radius / "where does X live" holo_query_codebase, holo_impact_analysis MCP tools mcp.holoscript.net UP — verified 2026-04-26
Natural-language Q&A over codebase holo_ask_codebase MCP GraphRAG + HoloEmbed/structural (keyless; OpenAI opt-in only, not required — F.106)
Semantic search by meaning holo_semantic_search MCP top-k retrieval
Higher-level architectural reasoning engineering:system-design + engineering:tech-debt ✅ flat code-spelunking is the anti-pattern

Papers (Gated Research)

Workflow trigger Skill Notes
Threat model / red-team / Paper 21 /security-audit ✅ adversarial-first, attack PoC + measured defense
Lean 4 formalization / Papers 22+23 /lean4 ✅ Mathlib, kernel type-check
ML metrics / F1 / ablation / Papers 17-20 /ml-experiments ✅ benchmark protocol, reviewer-survivable
SNN / neuromorphic no plugin — read .claude/skills/neuro/SKILL.md directly for ML metrics defer to /ml-experiments

External Presence

Workflow trigger Skill Notes
Moltbook engagement (replies, posts, karma) /moltbook ✅ F.005: engagement only, no auto-crosspost
External-facing copy (X, Reddit, HN, LinkedIn, Discord, PH, GitHub README) /marketer ✅ documentarian voice, shows what was built

Platform Slices

Workflow trigger Skill Notes
HoloMesh network admin (infrastructure, governance) /network ✅ CEO-level mesh authority
Ecosystem dashboard / cross-service admin /admin ✅
Shared ecosystem helpers /lib ✅
Workspace / multi-agent build orchestration /builder-manager ✅
VR/AR worlds, HoloLand platform /hololand ✅ autonomous world management
Consolidated HoloScript platform skill none — slice via above; for architecture read NORTH_STAR.md + this AGENTS.md

Full Protocol

Workflow trigger Skill
Full 7-phase uAA2++ end-to-end /∞ ✅
Domain-specific knowledge intake /intake (master) or /intake-<domain> ✅

For Claude Code agents: Invoke any ✅ skill via the Skill tool — forks context, returns condensed result. For ungrouped or "no plugin" rows, read the file directly via the Read tool.

For all other agents (Copilot, Cursor, Gemini, Codex, Windsurf, Devin): Most plugin skills aren't accessible. Read the underlying skill file directly — it's the best single-file domain summary.

Skill registration mechanism (W.110): Repo-local skills under <repo>/.claude/skills/ are NOT auto-discovered by the Skill tool. Register via cp -r <repo>/.claude/skills/<name>/ ~/.claude/skills/<name>/. Effect immediate (no restart). Plugin migration follow-up: task_1777238974785_qdd5.