yzxoi/synergy-oryn/add-agent

Add a Built-in Agent

Establish the Boundary

1. Confirm the request requires a repository-built agent. Route configurable agents to 60-agents.jsonc, plugin agents to the plugin SDK, and external coding agents to external-agent/.
2. Read Cortex, Workflows, and Execution boundaries when delegation, review, or permissions are involved.
3. Inspect agent.ts, builtin-context.ts, the target builtin-*.ts, and two neighboring agent factories. Treat these files as authoritative; do not copy an old inventory from documentation.
4. Load integrate-llm when host code will invoke the agent for classification, extraction, generation, review, or delegated work. Decide explicitly whether the call is sessionless, continues an existing session, or launches a Cortex child.

Implement

1. Choose the owning catalog:

• primary orchestrator: builtin-primary.ts
• classic subagent: builtin-legacy-subagents.ts
• coding-harness subagent: a prompt factory registered in builtin-max-subagents.ts
• hidden utility or model-only agent: builtin-internal.ts

1. Add or update the prompt under agent/prompt/. Match the nearest flat prompt or base.txt plus builder.ts pattern. Keep product policy in code/config contracts and keep the prompt focused on the agent's role and completion criteria.
2. Define the agent with the current Agent.Info or createSubagent(ctx, definition) contract. Select the narrowest existing SubagentPermissionProfile; add a new profile only when no current profile expresses the required capability boundary.
3. Keep tool exposure separate from authorization. Native task-callable subagents retain the common search_tools and expand_tools permissions so they can activate deferred tools already allowed by their profile; profile-specific denies must still keep those tools unavailable after expansion.
4. Set visibleTo, delegationGroups, and hidden deliberately. Primary agents may target only agents exposed through their catalog. BlueprintLoop and Light Loop reviewers remain host-selected rather than direct primary targets, while their Cortex tasks are visible in the execution session's Subagent Dock.
5. Register a new max-subagent factory in FACTORIES; register other new catalogs through Agent.create() only if a genuinely new catalog is required.
6. Update generated agent-table behavior or tests if the new agent changes routing-visible metadata. Do not maintain a second hand-written agent list in prompts or docs.
7. Host-owned identities used for workflow authorization must be reserved before config aliases, plugin and external-agent discovery can merge them. Test identity/prompt/profile replacement as well as tool permissions; a final permission clamp cannot repair a renamed identity.
8. Keep agent registration separate from invocation. A hidden model-only agent does not by itself justify a new local LLM.stream() wrapper or a manually created child session.

Prompt Quality

1. Give the agent a role that is a real-world professional identity plus a concrete domain (for example, "a social-content growth strategist focused on short-video platforms"). Reject product, brand, or codename as the role, reject pipeline-step labels ("verb + agent"), and reject generic wrappers such as "smart assistant".
2. Keep prompts of end-user-deployed agents runtime-isolated: they may not reference code-side schema, field, or class names or pipeline-position vocabulary such as upstream/downstream — write constraints as agent-facing behavior ("You receive the already-confirmed context", not "PushMessageDraft.risk_level comes from upstream"). Internal subagents that operate on this repository may keep precise code vocabulary (class and pipeline names) where it is the accurate domain language.
3. Use a layered structure: role and mission, scope and non-goals, tool authority and tool-use policy, output contract, safety and failure handling, and verification and self-check.

Verify

1. Add a behavioral or catalog test before implementation when behavior changes. Assert visibility, permission, model-role, or routing invariants rather than source text.
2. Run the narrow agent/session tests from packages/synergy.
3. Run bun run typecheck and bun run quality:quick from the repository root.
4. Exercise the affected primary catalog in an isolated development instance when prompt routing or delegation changed; use the develop-synergy skill.
5. Update AGENTS.md only for a durable repository rule or built-in-agent boundary, and update canonical architecture docs only when the system contract changed.

Handoff

Report the agent class, visibility/delegation boundary, permission profile, model role, prompt location, tests run, and any deliberate catalog exclusions.