Roster

The agents that work in this repository: tier, model, whether they block, and what each owns.

Every role is defined in this repository under .agents/agents/, and the roster is ADR 0258 (roster v2). The sdlc plugin is retired here (ADR 0121 D3). Each agent file is the source of truth for its own checks; this page is the map.

Tiers and models

An agent declares a tier, never a model. .agents/models.yaml maps each tier to a model per assistant, so a model change is one edit (ADR 0193).

Tier	Claude	OpenCode	Used for
lead	Opus	—	The main session and the incident commander
guardian	Opus	—	Blocking gates
reviewer	Sonnet	—	Advisory reviews; escalate to Opus on a disputed finding
executor	Sonnet	Kimi K3	Builders that write code, tests or specs
steward	Haiku	Kimi K3	Mechanical sweeps

Fable is used only for ADR and constitution reversals and deep audits. OpenCode receives only steward and executor agents (ADR 0121 D10), and an agent marked assistants: [claude] never reaches it: secret-manager, compiler-engineer, data-engineer, infra-engineer and spec-writer are Claude-only because each touches change risk R3 or a generated contract. If a cheaper tier produces a wrong result, escalate one tier for that task rather than editing the agent.

Who blocks

1. Where enforcement and prose disagree, the enforcement is the fact. If a charter and the check disagree, the charter is the bug.
2. Never report a stage complete on the strength of a gate that could not have observed it. Say what was and was not run.
3. No agent both writes a change and holds the blocking verdict on it. test-engineer writes tests; the Quality Gate blocks on a red suite.

A human approves every release and every production promotion. No agent promotes to production.

Roster

Lead

Agent	Model	Owns
the main session	Opus	Classify (lane, change risk), brief, delegate, adjudicate analyze, keep Linear in state, merge via pr:merge
incident-commander	Opus	Pipeline, release and leaked-secret incidents; post-mortems and runbooks

Guardians (block)

Agent	Blocks on
architecture-guardian	Placement, tier direction, module layout, the not-built-yet list, a decision with no ADR
security-tenancy-reviewer	Auth, IAM, tenancy and RLS, secrets, webhooks, SSRF, personal data in logs, events and errors
release-operations	CI, the version pull request, publishing, promotion readiness, mobile and desktop trains

Executors (write)

Agent	Owns
package-builder	One package change; the default executor for spec tasks in packages
compiler-engineer	stackra gen and the CLI in packages/developer; the only reader of the fenced Encore reference
service-builder	One service change: modules, actions, queries, listeners, workflows
data-engineer	Models, @stackra/database, migrations and their safety rules, RLS policies, seeds
frontend-builder	The console and web experience packages: panel resources, pages, widgets
desktop-pwa-builder	PWA and Electron targets, @stackra/platform, peripherals, client storage, offline
mobile-builder	apps/mobile and EAS profiles
infra-engineer	Pulumi components, runway.yaml, edge, reusable workflows; previews only
test-engineer	Tests at every tier, isolation tests, Playwright with axe, Maestro
spec-writer	spec-kit documents for one Linear issue; lists open questions, decides none
secret-manager	pnpm secrets:import from the secrets inbox; never sees a value

Reviewers (advise)

Agent	Blocks?	Owns
code-reviewer	No; its AI-REVIEW verdict gates pr:merge	Correct, simple, reused, right altitude
api-contract-reviewer	No; escalates to the guardian	Action and event contracts, envelope, ids, MCP exposure, versions
accessibility-reviewer	No; axe in the end-to-end tier fails critical and serious	WCAG 2.2 AA, web and native, LTR and RTL
performance-reviewer	No	The quality budgets and query cost
design-brand-reviewer	No	Reuse first, tokens only, brand rules

Stewards (sweep)

Agent	Owns
code-standards-steward	Naming, kind folders, imports, barrels, package entries and exports, errors
dependency-steward	Catalogs, advisories, licences, mise pins, Renovate, vendored skills
docs-adr-steward	ADRs, their index and navigation
docs-changesets-steward	Changesets on the fixed version line
i18n-steward	Keys, RTL, locale data; reports only

Review chain

Every builder’s change goes to code-reviewer, then the domain reviewer its charter names, then the guardian its change risk triggers: architecture-guardian at R2 (a public contract, a new member or dependency), security-tenancy-reviewer at R3 (ADR 0120). The Quality Gate runs, and the lead merges only after an independent AI-REVIEW: PASS.

Routing

.agents/routing.yaml maps a task type to an executor, model and agent; pnpm delegate reads it. A route that sends an agent to OpenCode when the agent is not projected there fails pnpm agents:check.

The work	Goes to
Adds a member or dependency, crosses a tier, or reverses a decision	architecture-guardian before anyone builds
A package change	package-builder
The generator, the CLI, anything that emits Encore	compiler-engineer
Backend behaviour in a service	service-builder
A model, table, migration or RLS policy	data-engineer, then security-tenancy-reviewer
A web screen or panel resource	frontend-builder, then design-brand-reviewer, accessibility-reviewer, performance-reviewer
A PWA, desktop, device or offline change	desktop-pwa-builder
A mobile screen or EAS profile	mobile-builder
Pulumi, runway.yaml, workflows	infra-engineer, then release-operations
Tests, isolation tests, flaky tests	test-engineer
A spec, plan or task list	spec-writer, then /speckit.analyze and the lead
Changes an action, event or MCP contract	api-contract-reviewer
Touches auth, secrets, tenancy, payments, files or webhooks	security-tenancy-reviewer (gate)
Versions, advisories, licences, skills	dependency-steward
A secret value to set	secret-manager
CI, release, promotion	release-operations

Not created

• delivery-orchestrator, docs-governance and spec-planner fold into the lead (ADR 0121 D9).
• spec-executor is not ported: the builders implement spec tasks (ADR 0258).
• package-steward is folded into code-standards-steward (ADR 0193).
• sql-reviewer and privacy-officer are deferred to the first tenant-owned table that holds personal data; until then security-tenancy-reviewer and the migration lint cover them.
• A read-only auditor is the built-in Explore agent.
• A design agency (director, senior designer, design engineer) was judged too heavy; one non-blocking design-brand-reviewer plus the design skills replaces it.

Related: Working with agents, Skills, Spec lifecycle.