Skip to content

t1k:web:core:agent-domain-routing

FieldValue
Modulecore
Version1.26.1
Effortlow
Tools—

Keywords: agent domain, agent for, agent lookup, agent routing, discover agent, route to agent, spawn agent, web agent, which agent

/t1k:web:core:agent-domain-routing

Web rows for core rules/orchestration-rules.md § “Task-Type → Agent Routing”. Role keys: .claude/t1k-routing-web.json (kit-wide) and .claude/modules/core/t1k-routing-web-core.json (module overlay). For SKILL discovery (not agent discovery), see t1k-web-core-skill-domain-routing instead: this table routes to spawn AGENTS only.

Modeled on theonekit-unity’s t1k-unity-base-agent-domain-routing (moved out of its rule on 2026-08-24 as a context-budget lever; this table is consulted only after an agent has already decided to spawn a sub-agent for web work). Web ships 6 domain agents across 5 non-overlapping concerns (the visual director and the visual reviewer share one concern, split by phase (brief before the work, review against it after)), so there is no DOTS-style tie-break to preserve; the companion rule agent-domain-routing-web.md stays a thin pointer to this skill.

Task patternAgentTier
Full-stack: React/Next.js (App Router, RSC, SSR/ISR), Node.js/Python/Go backend (NestJS/FastAPI/Express APIs), Postgres/MongoDB/Prisma/Drizzle, OAuth/JWT/Better Auth, Docker/Cloudflare Workers infrat1k-web-core-developersonnet
Exact-spec mechanical edits: a rename, a diff to apply across named files, a measurement to transcribe, boilerplate from a templatet1k-web-core-developersonnet
UI/UX implementation: design systems (color/typography/spacing), Tailwind/shadcn component work, WCAG 2.1 AA accessibility, responsive/dark-mode, a visual lane with a brief of recordt1k-web-ui-developersonnet
Motion & rendering: GSAP/ScrollTrigger/SplitText, Lenis/smooth scroll, canvas/WebGL/GLSL shaders (activates t1k-web-3d-shader), View Transitions, scroll-driven CSS, motion performance budgetst1k-web-frontend-motion-developersonnet
Testing: Playwright E2E, Vitest unit/integration, k6 load tests, visual regression, axe-core accessibility audits, Lighthouse suites, REST/GraphQL endpoint validationt1k-web-testing-testerhaiku
Visual judgment before implementation: a measurable design brief, creative art direction (“which of these should the hero take”)t1k-web-core-visual-directorfable
Visual review of a built page: a checklist pass/fix list against a brief of record (one per round, see § Director loop), screenshot comparison, contrast computed from tokenst1k-web-core-visual-reviewersonnet
Write-shaped migration lane: multi-service data-layer migration, readers/writers swap, backfill/dual-write cutover — a lane that changes state, not just filest1k-web-core-developersonnet
Independent code review of a lane (correctness + tests + security): one spawn PER CONCERN — t1k-review-correctness / t1k-review-tests / t1k-review-security — synthesised by t1k-review-synthesizer; never one t1k-review-tests for all threet1k-review-correctness + t1k-review-tests + t1k-review-security (→ t1k-review-synthesizer)sonnet

Tier here is a routing key, not a cost hint. The fable row is never cheap-routed; it is the one to reach for when the lane needs creative judgment (a design brief, art direction). The sonnet rows implement against briefs and measured constraints that are already made; an implementation lane whose decisions are still open goes to the fable row first. Sending a no-brief judgment lane to an implementation row produces plausible work against guessed constraints; sending a settled punch-list to the fable row burns premium on transcription. The one sonnet exception is the visual reviewer: it is a sonnet pin with mrHints.requires: ["vision"], so any routing stays inside the vision-capable pipe (see the reviewer body’s § Constraints).

History. Before 2026-09-13 the implementation lanes were sonnet-pinned, so transparent model routing sent every one of them to a budget model; an orchestrator hitting a quality wall forked its own premium session and did ~20 lanes inline on a single page. The two implementation rows were briefly pinned opus (2026-09-13 to 2026-09-16, telemetry showed ~80% of their briefs were punch-list transcriptions) before returning to sonnet; the “do not fork your own session” constraint in the companion rule is what keeps that escape closed now. The mirror of that constraint is why the tiers stay split: what earns a separate lane for settled work is skill context (the ui/motion rows carry their own skills), and an exact-spec edit list does not need them. Escalating settled punch-list work to the fable director pays premium for a judgment the brief already made.

The orchestrator that wrote the lane briefs owns the whole visual loop; the implementers are leaves and cannot spawn the director themselves.

  1. One mode-1 brief per page/round from t1k-web-core-visual-director (fable), reviewed in mode 2 by t1k-web-core-visual-reviewer (sonnet); never one per lane.
  2. Spawn all lanes for that round against it.
  3. ONE mode-2 review per round, after the lanes land: one spawn covering every lane’s output, not one review per lane as each finishes. This is where the premium saving lives.
  4. Relay the per-lane fix items back to the lane that owns each.
  5. Cap: 2 review rounds per page. Still failing after the second, stop and re-issue the brief: a third round means the brief was wrong, not the lanes.

Backstop when a cheap lane does not hold up

Section titled “Backstop when a cheap lane does not hold up”

A cheap lane that returns non-compiling work is re-briefed once with the compile/test command as acceptance; a consumer that keeps hitting it pins the agent via modelRouter.excludeAgents, never by forking.

.claude/t1k-routing-web.json holds only implementer and the errorRecovery chain; the ui-developer, tester and motion-developer rows live in the inline routingOverlay of the ui, testing and frontend modules’ module.json (present only when that module is installed). Read those when resolving by role key rather than by task pattern. /t1k:find-skill also resolves this table by keyword.