v0.5.0-prerelease — Phase 4: Agent-Loop + Skill-Korpus

Agent-loop CLI primitives + skill-corpus (30 markdown files, 2854 lines)
encoding Samuel's vault anti-slop doctrine, UI-UX-Pro-Max corpus (MIT fork),
Huashu directions, and Open-design's 5 OKLCh presets as an integrated
discoverability surface.

Critical architectural insight: Claude Code's reasoning IS the agent.
src/agent/poll-loop.ts provides one-shot pollOnce and postEvent
primitives; the while-loop lives declaratively in
skills/wisp-design/SKILL.md. Restartable across Claude sessions because
the bridge holds the cursor, not the agent.

What landed

Contracts (src/contracts/agent.ts, 433 lines)

• PollOnceOptions, PollOnceResult, PostEventOptions, PostEventResult
• VariantGenerationRequest / Response with 5 variant axes
  (hierarchy / layout / typography / color / density)
• BrandSpec (9-section, name+oneLiner required, rest optional)
• DesignTokens, SyncSource, SkillsIndexResult
• Constants: POLL_LOOP_DEFAULT_TIMEOUT_MS=270000,
  ALLOWED_VARIANT_ANCHORS (13), ALLOWED_VARIANT_COUNT ({1,3,5,8}),
  VARIANT_RATIONALE_MAX_LEN=180

Agent modules (src/agent/, 4 files / 1144 lines)

• _helpers.ts (187) - readPortLock with PID-liveness probe, parseFlags,
  typed PortLockMissingError / PortLockStaleError
• poll-loop.ts (366) - pollOnce + postEvent (Node 20+ fetch +
  AbortController with 5s headroom), routeEvent exhaustive switch,
  runPollOnce + runPostEvent CLI runners
• skills-index.ts (309) - indexSkills recursive walk + namespace counting,
  searchSkills BM25-lite with sqrt-length-norm, runSkills dispatcher
  (index / search / --help)
• sync.ts (282) - syncFromVault with minimal glob compiler, SHA-256
  idempotent skip, YAML attribution frontmatter, runSync CLI

CLI lazy-load (src/index.ts + tsup.config.ts)

String-variable indirect import pattern bypasses TS module resolution at
compile time. tsup multi-entry now emits dist/agent/poll-loop.js,
dist/agent/skills-index.js, dist/agent/sync.js so the dispatcher's
runtime await import(...) resolves successfully.

Skill corpus (skills/, 30 files / 2854 markdown lines)

• wisp-design/SKILL.md - auto-trigger skill with loop pseudocode
• reference/ - live.md (KERN variant-generation prompt) + 5 themed
  sub-prompts (polish, bolder, quieter, colorize, layout)
• policy/anti-slop.md - 7 Vault hard-bans (em-dash UI, gradient text,
  default glassmorphism, hero-metric template, side-stripe, purple-blue
  gradient, generic AI illustrations) + Huashu 5-dim self-critique
• methodology/narrative-questions.md - 4 Pre-Code questions
  (Role / Distance / Temperature / Capacity)
• methodology/junior-designer-flow.md - 4-phase Stub > Checkpoint > Fill > Verify
• methodology/brand-asset-5-10-2-8.md - 5min / 10min / 2h / 8w protocol
  (opt-in via --brand flag per Open Decision #4)
• data/anchors/ - 13 variant anchors (5 vault apps: Linear, Stripe,
  Anthropic, Aceternity, Apple + 5 Open-design presets: editorial,
  modern-minimal, tech-utility, brutalist, soft-warm + 3 vault-adjacents:
  restrained-cool, committed-indigo, drenched-warm)
• data/directions/ - 00-INDEX with 25-direction list + README with MIT
  attribution to huashu-design
• data/corpus/ - INDEX + README (UI-UX-Pro-Max MIT attribution) +
  30-row sample CSV demonstrating Quadruple shape

Open Decisions confirmed

• #4 Brand-Asset 5-10-2-8: opt-in via --brand flag, default off
• #5 UI-UX-Pro-Max: MIT-fork with attribution (Phase 4 ships INDEX +
  sample; full 14-CSV fork deferred to Phase 7)
• #6 Vault-sync: explicit wisp-design sync command only, no watcher

Doctor (now 9/9 OK)

Added checkSkillsLayout + checkSkillManifest as WARN (not FAIL) so Phase 0
checkouts don't trip the gate.

Tests (tests/agent/ + tests/skills/, 77 new in 5 files)

• poll-loop.test.ts (20) - bridge integration on port 31370, pollOnce/
  postEvent happy + error paths, routeEvent exhaustive, CLI runner I/O
• skills-index.test.ts (19) - enumeration, search relevance, CLI dispatch
• sync.test.ts (15) - glob matcher, idempotent SHA skip, attribution
• skill-corpus.test.ts (15) - file existence, frontmatter, axis coverage,
  hard-ban presence, anchor count, attribution presence
• anti-slop-canary.test.ts (8) - mission-critical canary locks the 7
  hard-bans + all 5 critique dimensions against silent removal

Cumulative: 347/347 tests green (77 agent/skills + 99 source + 94 browser

• 77 bridge + 2 smoke + extras).

Quality gates

typecheck:  clean
test:       347/347 in ~4s (24 files)
build:      index.js 11.92 kB + agent/poll-loop.js 21.40 kB +
            agent/skills-index.js 16.27 kB + agent/sync.js 17.66 kB +
            live.js 31.30 kB
doctor:     9/9 OK

Phase-4 deferred items (Phase 7 launch prep)

• Full UI-UX-Pro-Max 14-CSV fork with AgentDB HNSW indexing
• Full Huashu 20+ directions content fork (reconcile synthesized list with
  upstream)
• Real AgentDB pattern-store wiring (currently in-process BM25-lite, marked
  agentDbController: "phase-4-stub" per Phase-4 contract)

Next

Phase 5 - Verification-Gate (USP). Anti-slop linter that enforces the 7
Vault hard-bans on generated CSS; axe-core a11y delta with AA-blocking
policy; multi-viewport screenshot (375 / 768 / 1280 / 1920 + light + dark);
console-clean scan; tab-order smoke; reduced-motion diff. Verification-Gate
runs PARALLEL to LLM generate, not sequentially. Total p95 budget <= 3s.
This is the central wisp-design differentiator - no other live-edit tool
gates accept on a11y + screenshot + anti-slop.