agent-builder

name

agent-builder

description

Build, edit, test and deploy a Prisme.ai Agent Factory agent whose definition is git-versioned inside a workspace. The prompt lives in AGENT.md; model/profile/capabilities/app+mcp/tests live in agent.yml. Use when the user wants to scaffold a new agent, change an agent's prompt or capabilities, list/search the platform's available capabilities, install and wire an App+MCP connector as an agent tool, manage the test battery (list/add/edit tests, run the suite), choose or change the deployment organization, or deploy the agent (create/update + publish + push the workspace) to sandbox or prod. Triggers include "crée un agent", "modifie le prompt de l'agent", "ajoute une capacité", "quelles capacités sont dispo", "installe l'app+mcp X", "ajoute/lance les tests", "déploie dans l'organisation X", "change d'organisation", "déploie l'agent", "/agent-builder ...".

Agent Builder

Assist the user in building and maintaining a single Agent Factory agent whose entire definition is stored in a git-versioned workspace folder, so it can be diffed, reviewed and rolled back like code.

This skill is the action-oriented sibling of agent-workspace. Where agent-workspace bootstraps a one-shot RAG+evaluation workspace, agent-builder gives the user durable CRUD over one agent: prompt, capabilities, App+MCP tools, tests, and multi-environment deployment.

Source of truth (git-versioned)

Everything the agent IS lives in two files inside the workspace folder:

File	Holds
AGENT.md	The agent's system prompt / instructions (plain Markdown, clean diffs).
agent.yml	The manifest: name, model, temperature, profile, visibility, capabilities[], appMcp[], tests[], and deployedIds (written by the skill).

deploy is declarative: the skill reads AGENT.md + agent.yml and makes the remote agent match them exactly. The user edits files; the skill reconciles.

Workspace layout

Agent workspaces live under agents/<slug>/ in the prismeai-workspaces repo — a top-level sibling of workspaces/ and pages/ dedicated to /agent-builder agents (so Agent Factory agents are not mixed in with the connector/app workspaces). Detect the repo root before doing anything (it is the directory that contains workspaces/, pages/, agents/ and docs/); if the CWD is the parent "Prisme.ai Projects" dir, the repo root is prismeai-workspaces/.

Migrated 2026-06-26: agent workspaces moved out of workspaces/<slug>/ into the new top-level agents/<slug>/. Discover existing agents under agents/; when an older agent is still found under workspaces/<slug>/, treat that as legacy and offer to move it to agents/<slug>/.

agents/<slug>/
├── AGENT.md                 # prompt (source of truth)
├── agent.yml                # manifest (source of truth)
├── index.yml                # workspace config + secret schema
├── security.yml
├── imports/
│   ├── Agents.yml           # Agents app (used by runTests for sendMessage)
│   ├── Custom Code.yml      # evalTest() — robust answer extraction + assertion
│   └── <connector>.yml      # one per installed App+MCP connector
└── automations/             # FLAT — every file at this level (see slug=path rule below)
    ├── deploy.yml           # slug deploy · name "Deploy this agent" — run AS THE USER
    ├── runTests.yml         # slug runTests · name "Run the test battery"
    ├── _afCall.yml          # slug _afCall · name "utils/…" · agent-factory webhook (asUser | org-key)
    └── whoAmI.yml           # slug whoAmI · name "utils/…" (legacy) — deploy now returns org

⚠️ push_workspace keys each automation by its FILE PATH relative to automations/, and SILENTLY DROPS any file whose slug: ≠ that path. So automations/utils/_afCall.yml (slug _afCall) is never imported — the deploy then 404s with "Automation not found: _afCall", and a webhook at /webhooks/getContext 404s "no matching trigger". Keep helper files flat at automations/<slug>.yml (slug = filename). The studio "utils" folder is purely cosmetic via the name: prefix (name: utils/Agent Factory API call) — it groups the UI without touching the slug. A nested file is only OK if its slug includes the folder (e.g. automations/v1/status.yml with slug: v1/status). Verified on cd76 2026-06-26.

Scaffold templates for every file live in templates/ next to this SKILL.md.

Non-negotiables

• Deployment goes through the AgentBuilderSync app — a single central dashboard (one per environment), NOT a per-agent SPA or playground run. The agent workspace only needs its definition in config.value.agent (compiled from AGENT.md + agent.yml at push time) plus the label agent-builder to be discoverable. The app lists the labeled workspaces, reads each definition, lets the user pick a model from the active org's allowed models, and deploys the selected agent as the user (so it lands in the active session org, human as owner). The skill does NOT create agents via call_api, does NOT pick an org, and does NOT ship a deploy SPA per agent. Give the user the app URL for the env (see "AgentBuilderSync app" below). (Test sending still uses the Agents app inside runTests; never fetch Agent Factory directly.)
• AGENT.md and agent.yml are the source of truth. Never mutate the remote agent in a way that drifts from them — always update the files, then deploy.
• Deploy needs no Governance org key — it runs as the user. The agentFactoryApiKey secret (a iak_... org key) is optional/legacy, used only by the whoAmI org-key helper; leave it unset unless you keep using that helper. Apps the user installs use their own credentials (x-prismeai-api-key), unrelated to this.
• Connectors are sandbox-first. Default every push / deploy / id to sandbox unless the user explicitly asks for prod.
• The deployment organization = the org active in the studio when the user deploys via the app (session.org.slug), and the user becomes the owner (owner_id = user.id). To deploy elsewhere, the user switches org in the studio and re-deploys — the skill never lists or selects orgs. The user must be a member of the target org with rights to manage agents (else the deploy 403/401s). See Operation 7 (and Operation 6b for the underlying mechanism).
• Confirm the spec with the user before creating files (scaffold) or before the first remote deploy. Subsequent edits + redeploys need no re-confirmation unless the user is changing environment.
• Run validate_automation on the changed automation (deploy.yml, _afCall.yml, runTests.yml, …) after any change, and on the folder after scaffolding a new workspace.
• A workspace's React app (if any) lives in the sibling top-level pages/<slug>/ folder (alongside workspaces//agents/, named after the workspace), never inside the agent folder. Agent workspaces are DSUL-pure and have none by default. push_workspace targets the agent's DSUL folder (agents/<slug>/) only and never touches pages/<slug>/ — see the project-wide rule.

AgentBuilderSync app (how agents are deployed)

Deployment is centralized in one app per environment — the agent-builder-sync workspace + its SPA. It discovers every workspace labeled agent-builder, reads each one's config.value.agent, and deploys the chosen agent as the user into the active org. So an agent workspace is just a definition holder: config.value.agent (compiled from AGENT.md + agent.yml) + the agent-builder label. No per-agent deploy SPA.

App coordinates (recall and hand the URL to the user when deploying):

Env	Workspace id	App URL
sandbox	65xmBBG	https://sandbox.prisme.ai/apps/agent-builder-sync
prod	JDNlu-Q	https://prisme.ai/apps/agent-builder-sync

The app's own DSUL/SPA lives at agents/agent-builder-sync/ + pages/agent-builder-sync/ in the prismeai-workspaces repo (built/deployed via /workspace-page-implement). Its deploy webhook refetches the target workspace's config.value.agent, calls agent-factory as the user (_afCall asUser: true), and persists per-(workspace, org) state in global.deploys. The agent's org = the user's active studio org, owner = the user.

MCP tools = capability references (env/org-agnostic) — NEVER hardcode URLs

In config.value.agent.tools, store an App+MCP tool as a capability reference: { type: mcp, capability: <catalog name> } — NOT a full { server, auth, … } with URLs. At deploy, the deploy webhook looks capability up in the ACTIVE org's capabilities catalog (/workspaces/slug:capabilities/webhooks/v1/servers?type=mcp&search=<name>, exact name match) and builds the real tool: server (= the entry's config_schema.properties. server.default) + the OAuth auth block ({type: oauth, connect_url, status_url, disconnect_url, scopes} from the entry's auth). Why this matters:

• Env-agnostic: the catalog is per-env, so the same definition resolves to sandbox URLs on sandbox and prod URLs on prod — no host swap when pushing a definition across envs.
• Per-org: the catalog is per-org, so cd76 can register its own connector under the same capability name (sharepoint_next) pointing at slug:sharepoint-cd76, and the deploy picks it up automatically for that org.
• OAuth always wired: the auth block comes from the catalog → the OAuth "Connect" popup always appears (agent-factory _route-tool gates it on the tool's auth block; a tool without auth silently runs unauthenticated and never shows a Connect button).
• If the capability isn't in the target org's catalog, deploy returns CapabilityNotFound with the list of available capabilities. Pre-req: the connector must be registered in each target env/org's capabilities catalog (POST …/v1/servers, or the connector SPA's CatalogPublish) — deploying the connector workspace alone does NOT add it to the catalog.

Migration note: data-analyst-cd76 still carries a legacy per-agent deploy SPA + automations (the POC). New agents should NOT get one — they only need config.value.agent + the label. Slimming the POC is a separate cleanup step.

The manifest in detail (agent.yml)

name: <Display name>
slug: <agent-slug>
model: gpt-4o-mini            # llm-gateway model id — ASK the user, do not assume
temperature: 0.2
profile: agent_light          # simple|workflow|agent_light|agent_full|orchestrator
visibility: restricted        # publish visibility: restricted|public|private
# --- Per-agent budget overrides (optional; omit to inherit the profile defaults) ---
token_budget: null            # per-task token budget; agent_light default = 20000
tool_call_budget: null        # max tool calls per task; agent_light default = 15
max_turns: null               # max ReAct turns per task; agent_light default = 10
capabilities: []              # tools other than App+MCP (see Tool types below)
appMcp: []                    # installed App+MCP connectors wired as mcp tools
tests: []                     # inline test battery
# knownOrgs is legacy — the skill no longer selects orgs (deploy follows the active
# playground session). Kept only as a human note of orgs you target; not used by the skill.
deployedIds:                  # written by the skill, do not hand-edit
  sandbox: { workspaceId: null, agentId: null, org: null }  # org = { slug } (read back from deploy)
  prod:    { workspaceId: null, agentId: null, org: null }

Profiles gate capability tiers (see docs/agent-factory/overview.md): simple (no tools) < workflow < agent_light (ReAct loop, tools) < agent_full (planning, reflection, long-term memory, HITL) < orchestrator (+ delegation). An agent that uses tools/MCP needs agent_light or higher.

Native system tools (todo_write, knowledge_search, planning_create_plan, memory_*, human_*, agent_delegate, …) are injected automatically by the profile. Never list them in capabilities.

Per-agent budget (token_budget / tool_call_budget / max_turns)

Each profile caps every task by a token budget, a tool-call budget, and a max turns (agent-factory _init-defaults reads the agent record first, then falls back to config.profiles[<profile>].limits). The agent_light defaults are token_budget: 20000, tool_calls: 15, max_turns: 10 — low for tool-heavy agents (big MCP tool schemas + verbose tool results blow 20000 fast). The symptom is a truncated reply: "Réponse incomplète — le traitement a été interrompu … limite de budget." and usage.tool_calls/ tokens well under what the flow needed.

Override per agent (don't touch the shared agent_light profile — it's a protected one-product core workspace and affects every agent): set token_budget / tool_call_budget / max_turns in agent.yml. _init-defaults uses the agent's value when present.

⚠️ Deploy gotcha — budgets do NOT survive a deploy. The deploy automation sends only name/instructions/model/temperature/profile/tools, so a deploy leaves token_budget/tool_call_budget/max_turns unset → the agent reverts to the profile default (e.g. 20000). Set them with a direct PATCH to the agent record, in the same org/user context as the deploy — i.e. as the user, against the agent that now lives in the active org:

PATCH {{apiUrl}}/workspaces/slug:agent-factory/webhooks/v1/agents/<agentId>
       body: { token_budget, tool_call_budget, max_turns }

Run it the same way the deploy ran (from the playground / as the user — e.g. an _afCall asUser:true PATCH), since the agent is owner-scoped to that user in that org. Re-do this after EVERY deploy. Read them back via GET .../v1/agents/<agentId>.

Tool types

capabilities[] and appMcp[].tool entries map to Agent Factory tools:

type	Shape passed to addTool / tools[]
function	{ type, name, description, extra: { url, parameters } } (parameters = JSON Schema)
mcp	{ type, name, description, server } (server = MCP endpoint URL)
file_search	{ type, name, description } (RAG over a Storage vector store)
skill	{ type, name, description, ... } (a Prisme.ai automation as a tool)
guardrail	{ type, name, description } (inline safety check, not LLM-chosen)

On deploy, the skill builds the agent's tools array = capabilities[] + one mcp tool per appMcp[] entry (with server resolved per environment).

Available models (per org)

An agent's model must be in the allowed-model list of the org the agent lives in, otherwise the agent fails at runtime (the org's quota_policy is typically hard_block). The allowed list is per-org and is the source of truth for what you may offer:

call_api { path: "/orgs/<orgSlug>", pick: ["slug","name","settings"], environment }
→ data.settings.llm.allowed_models   (+ default_completion_model / default_embedding_model)

• Whenever you ask the user to pick a model — at scaffold (Op 1) and at edit (Op 3 / changing the model) — first fetch the target org's allowed_models and present only those. Never propose a model that isn't in the list.
• Filter out embedding models for a chat/completion agent — drop entries whose id contains embed / embedding (e.g. text-embedding-ada-002, amazon.titan-embed-image-v1). Offer the completion models only.
• If the org has a default_completion_model, surface it as the recommended default.
• The relevant org is the one the agent lives in — deployedIds.<env>.org once deployed, otherwise the org the user says they'll switch to before running deploy. Ask the user for that org slug to fetch its allowed_models; don't enumerate orgs.
• On deploy, the chosen model is sent in the create/update body; an out-of-list model surfaces as a model/quota error, not as a generic failure.

---

Operations

The user invokes operations conversationally or as /agent-builder <verb> .... Pick the matching section. Always end by reporting what changed and, when relevant, the agent URL.

1. Scaffold a new agent

When the user asks to create a new agent:

1. Gather the spec: display name, slug (kebab-case), a first draft of the prompt, model (fetch the target org's allowed models and let the user pick — never assume; see "Available models" below), profile (default agent_light), and any initial capabilities/tests.
2. Present a concise plan (slug, model, profile, capabilities, tests, target env) and ask for confirmation.
3. Copy templates/ into agents/<slug>/, filling placeholders:
   • index.yml: name, slug.
   • AGENT.md: the prompt draft.
   • agent.yml: name, slug, model, temperature, profile, visibility, and any seed capabilities/tests.
   • keep imports/Agents.yml, imports/Custom Code.yml, and the automations (deploy.yml, _afCall.yml, runTests.yml, whoAmI.yml) as-is.
4. validate_automation on the automations folder.
5. Offer to deploy (Operation 7). Do not deploy without the user's go-ahead.

2. Edit the prompt

• Edit agents/<slug>/AGENT.md directly (it is plain Markdown).
• Summarize the change. Offer to deploy so it takes effect.

3. Manage capabilities (capability catalog)

The capability catalog is the source of truth for ready-to-add capabilities. It is the capabilities workspace (app MCPServers), exposed as a webhook API. Query it with call_api (the catalog is per-env; built-in entries are visible to everyone, org-custom entries are scoped to the caller's org — so to see an org's custom entries, query while that org is the active one).

List / search the catalog

call_api { path: "/workspaces/slug:capabilities/webhooks/v1/servers",
           query: { type, category, search, page, limit },
           pick: ["id","name","type","category","description"], environment }
→ { items: [...], total, page, limit }   # paginate: total may exceed limit (20)

Entry type is one of mcp, guardrail, file_search, function, skill, hook, memory. Present results grouped by type with name + one-line description. Filter with search/type/category for "list the X capabilities". Real examples seen: Sharepoint MCP, Powerpoint MCP, Brave Search, DataGalaxy, Excel, Outlook, ServiceNow (mcp); Send E-mail, Generate PPTX from template (function); the 6 guardrails (injection-detect, toxicity-check, pii-detect, hallucination-check, topic-guard, action-approval).

Native system tools (knowledge_search, planning, memory, delegation, …) are NOT in the catalog — they're injected by the profile. Don't add them; just mention them.

"Add the capability X" — read its config, then wire it

1. Resolve the entry id from the user's phrasing (list/search the catalog, match on name). Fetch full details:

   call_api { path: ".../v1/servers/<entry_id>",
              pick: ["id","name","type","config_schema","auth","tools","agent_tools","server"], environment }
   

2. config_schema (JSON Schema) tells you exactly how to configure it. Use its defaults; ask the user only for required fields that have no default. Per type:
   • mcp → { name, server (SSE/WS URL, usually has a default), headers?, scope? } → agent tool { type: mcp, name, server } (+ headers/scope if given).
   • function → { url, parameters (JSON Schema) } → tool { type: function, name, description, extra: { url, parameters } }.
   • file_search → knowledge-base config → tool { type: file_search, name, description }.
   • guardrail → tool { type: guardrail, name, description } (+ guardrail_type).
3. auth: if present (e.g. { type: oauth2, connect_url, status_url, disconnect_url, scopes }), the capability needs a per-user OAuth connection — adding the tool is not enough. Give the user the connect_url to connect their account (browser flow), and note they can check via status_url. (Pure config-key capabilities have no auth block — just fill config_schema.)
4. Add the resulting tool to agent.yml capabilities[] (or appMcp[] for an MCP with a connector you also install — Operation 4). Then deploy (declarative — the tool list is rebuilt from the manifest).

Remove a capability

Delete its entry from agent.yml capabilities[]/appMcp[], then deploy (the agent's tools are replaced with the manifest set).

4. Install + configure an App+MCP connector

Goal: install a connector into the agent's workspace, configure its credentials, and wire it to the agent as an mcp tool.

1. list_apps (and/or get_app) to find the connector and read its config schema. For connector specifics, defer to the app-mcp-implement / app-mcp-test skills.
2. Ensure the workspace exists on the target env (deploy at least once, Op 7, so deployedIds.<env>.workspaceId is set).
3. install_app_instance into deployedIds.<env>.workspaceId with appSlug, a local slug, and the config (credentials/options). Mirror the resulting import locally as imports/<slug>.yml so it stays git-versioned.
4. Configure / verify credentials with update_app_instance_config if needed. Smoke the connector with the app-mcp-test skill when the user wants confidence.
5. Add an appMcp[] entry to agent.yml: appSlug, slug, config, and a tool block (type: mcp, name, description). Do not hardcode server — it embeds the per-env workspaceId.
6. deploy (Op 7). At deploy time the skill resolves each appMcp tool's server from the installed instance config (get_app_instance_config → mcpEndpoint) on the target env, so the URL is always correct for that environment. If that field comes back as an unresolved {{...}} template (connector-dependent), call the connector's own config/getConfig automation to get the concrete endpoint instead of wiring a literal template as the tool server.

5. Manage tests

Tests are the git-versioned source in agent.yml tests[]. Each test:

- name: <short label>
  input: <message sent to the agent>
  expected: <substring/regex (contains) or exact string (equals)>
  mode: contains    # contains (default, regex match) | equals (case-insensitive equality)

• List: read and print tests[].
• Add / edit / remove: edit agent.yml tests[].
• No remote sync needed — tests run live against the deployed agent (Operation 6).

6. Run the test battery

1. Ensure the agent is deployed on the target env (deployedIds.<env>.agentId set); if the manifest changed since last deploy, deploy first so tests hit current behavior.
2. execute_automation runTests on deployedIds.<env>.workspaceId with { agentId: <deployedIds.<env>.agentId>, tests: <agent.yml tests[]> }.
3. runTests sends each input via Agents.sendMessage, then calls the Custom Code function via the dispatcher Custom Code.run (function: evalTest, parameters: {…}) — NOT Custom Code.evalTest directly. The current Custom Code app exposes functions ONLY through .run; the per-function-as-automation form raises ObjectNotFoundError: Custom Code.evalTest at runtime (this is a hard error, not a reload lag — see below). evalTest extracts the answer from response.task.output.messages[<last>].parts[].text (plural messages array; older shape output.message.parts handled as fallback) and asserts per mode (in JS, so arbitrary LLM output can't crash the run).
4. Report the suite result: passed/total, and per-test name, answer, passed. For failures, show expected vs. actual and offer to refine the prompt or the test.

If runTests errors with Custom Code.evalTest "not found", it's almost always the call-style bug above (use Custom Code.run with function: evalTest), NOT a reload lag — a update_app_instance_config reload will NOT fix it. Only if you are already using Custom Code.run and still get "not found" right after the first push, force a reload (update_app_instance_config on the Custom Code instance with its full config).

6b. Deployment organization — no selection, it follows the playground session

The agent's org = the org active in the studio when deploy runs from the workspace playground (session.org.slug), and the user who clicks run becomes the owner (owner_id = user.id). This is enforced by the agent-factory _auth middleware, which reads org and owner straight from the calling session — verified in workspaces/agent-factory/automations/_auth.yml. The deploy automation authenticates as that user (_afCall asUser: true, no api key), so the runtime auto-mints a user bearer carrying the session, and the agent lands wherever the studio is pointed.

Consequences — the skill no longer chooses an org:

• There is no org-list step. Do not call /me to enumerate memberships, do not present orgs, do not ask the user to pick one, do not handle org iak_ keys for deploy.
• "Deploy to org X" = the user switches the studio to org X, then runs deploy from the playground. To move an agent to a different org, switch + re-run; the org is never passed as an argument (a body.orgSlug would be ignored anyway).
• The user must be a member of the active org with rights to manage agents, otherwise the webhook returns 401/403. That's the user's responsibility (they switch to an org they can manage); the skill just surfaces the error.
• After a deploy, read the landing org/owner from deploy's output (org, owner — it reads the agent back) and record deployedIds.<env>.org = { slug }. No separate whoAmI / call_api confirmation needed.

For how to actually run the deploy (where the user clicks, what payload to paste), see Operation 7.

7. Deploy the agent (via the AgentBuilderSync app)

The skill's job is to make the agent discoverable and current; the actual deploy is a click in the AgentBuilderSync app (the user picks the agent + model and deploys, as themselves, into their active org). The skill no longer runs deploy itself.

1. Confirm the environment (sandbox default; prod only on explicit request).
2. Compile the definition into config.value.agent in the workspace's index.yml, from AGENT.md (→ instructions) + agent.yml: { name, model (default), temperature, profile, visibility, tools, instructions }, where tools = capabilities[] + one resolved mcp tool per appMcp[] entry (resolve each mcp server via get_app_instance_config on the target env). This is what the app reads — re-compile it on every prompt/model/capability change.
3. Ensure the workspace carries the label agent-builder (so the app lists it) and exists remotely:
   • First time: create_workspace → capture id into deployedIds.<env>.workspaceId.
   • push_workspace (agents/<slug>/) to upload index.yml (with the fresh config.value.agent) + DSUL. push_workspace DOES update config.value, so the app sees the new definition immediately. No agentFactoryApiKey secret needed.
4. Hand the user the app URL for the env (see "AgentBuilderSync app" — e.g. sandbox https://sandbox.prisme.ai/apps/agent-builder-sync). Tell them: open it, switch the studio to the target org if needed, pick <agent name> in the left list, choose the model, and click Déployer. The app deploys as them (active org, human owner), shows a success popup with a link to the agent, and a "Voir l'agent" link if already deployed.
5. The app tracks the deployed agent id per (workspace, org) in its own global.deploys — you don't need to write deployedIds for the app to work. (You MAY still record the resulting agent id in agent.yml deployedIds.<env> for git history if the user wants.)
6. Budget overrides (token_budget/tool_call_budget/max_turns) are still NOT carried by the app's deploy — if agent.yml sets them, PATCH them onto the agent after it's deployed, in the same org/user context (see "Per-agent budget"): PATCH .../v1/agents/<agentId> { token_budget, tool_call_budget, max_turns }.

The underlying mechanism (deploy webhook → agent-factory as the user → org from session.org.slug, owner from user.id) is unchanged from Operation 6b; the app just wraps it with a dashboard. If AgentBuilderSync isn't deployed on the target env yet, deploy it first (agents/agent-builder-sync/ + pages/agent-builder-sync/ via /workspace-page-implement) and record its id + URL in the table above.

---

Reference

Agent Factory access (templates)

deploy and whoAmI call the agent-factory webhook directly via the private _afCall helper (the published Agents app's argument schema is rejected by the prod runtime). _afCall has two auth modes:

• asUser: true → no api key; the runtime auto-mints the triggering user's bearer, so agent-factory resolves org from session.org.slug and owner from user.id. Used by deploy (create POST /v1/agents, update PATCH /v1/agents/<id>, publish POST /v1/agents/<id>/publish, read back GET /v1/agents/<id>).
• default (org key) → x-prismeai-api-key: {{secret.agentFactoryApiKey}}, org fixed by the key. Used by the legacy whoAmI only.

The Agents app is still imported and used by runTests for Agents.sendMessage ({ agent_id, message: { message_id, role: user, parts: [{ type: text, text }] } } → { task: { output: { messages: [{ parts: [{ text }] }] } } }).

Multi-environment

deployedIds stores both the workspaceId and the agentId per environment, because each is distinct between sandbox and prod. The skill never reuses a sandbox id for prod.

Validation & guardrails

• validate_automation is authoritative — trust it over existing patterns.
• After scaffolding, guardrail-scan the new workspace for hardcoded agent/workspace ids and stray fetch: to Agent Factory: the only allowed agent-factory fetch is inside _afCall (templated, via {{global.apiUrl}} — never a hardcoded host or id).
• DSUL gotchas that bit these templates: string-literal compares use '{{x}} = "value"'; var-to-var compares use '{{a}} = {{b}}' (no quotes); matches is regex and condition-level only; lower() runs inside {% %}; endpoint automations read top-level args when called via execute_automation.

Response shape

When done, report: files changed; capabilities/tests touched; validation status; deploy target + agent id; test-suite result if run; and the agent URL to try it.