feat(ai-agent): answer questions and decline unsupported edits in text

context/architecture-context.md

@@ -33,15 +33,19 @@
 - **One turn = one Trigger.dev run** (`design-agent`). Routes validate input and access, trigger the run, and
   return `202`; they never call the model.
 - **A turn** reads the Liveblocks room read-only and summarises it (`lib/ai/canvas-summary.ts`), then analyzes the
-  message (updates the requirements brief and decides ask / plan / generate), drafts or revises a plan, or
-  generates the diagram for the approved plan, validates it (`lib/ai/graph-validation.ts`), makes at most one repair
+  message (updates the requirements brief and decides ask / plan / generate / reply / unsupported), drafts or
+  revises a plan, answers in text, or generates the diagram for the approved plan, validates it (`lib/ai/graph-validation.ts`), makes at most one repair
   call, and writes it with `mutateFlow`.
-- **Code enforces the rules** the model is asked to follow (question rounds, skip, no generating without a plan) and
+- **Code enforces the rules** the model is asked to follow (question rounds, skip, no generating without a plan, text-only
+  answers only for typed messages) and
   fills in everything render-related: node and edge types, sizes, fills from `NODE_ROLES`, dashed async edges,
   connection points, and layout. Existing components are referenced by `ex-N` refs; generated duplicates fold into
   existing nodes.
+- **Add-only.** The agent never removes, merges, renames, or moves existing components. `reply` answers questions and
+  off-topic messages in text; `unsupported` (edit requests) and a plan with nothing new to add return a fixed message
+  written in code. Both settle as a `TEXT` message and leave the session phase and brief unchanged.
 - **Turns settle on read.** The turns route stores a `PENDING` reply with the run id; reading the session retrieves
-  the run, validates its output with zod, and stores `QUESTIONS`, `PLAN`, or `RESULT`. The task has no database
+  the run, validates its output with zod, and stores `QUESTIONS`, `PLAN`, `RESULT`, or `TEXT`. The task has no database
   access and never writes blobs.
 - **Reliability:** per-call timeouts (analyze 90s, plan 150s, generate 120s), task `maxDuration` 300s, a single task
   attempt, one retry per call on a schema mismatch, list limits trimmed in code instead of `maxItems`, and friendly

context/progress-tracker.md

@@ -1560,3 +1560,20 @@ Update this file whenever the current phase, active feature, or implementation s
 - `lib/prisma.ts` Accelerate URL check fixed (`prisma+postgres://`).
 - Validation: `pnpm typecheck` and eslint on changed files passed; sanitizer checked on empty and partial drafts.
   Not yet verified: an "add components" AI turn in production after the Trigger.dev deploy.
+
+### AI agent: text replies and unsupported edits (2026-09-16)
+
+- Problem seen in production: "simplify the graph" and "what is my name?" both ended as a red error ("The plan didn't
+  include any new components…"), because the analyze step could only ask, plan, or generate and the agent is add-only.
+- `TURN_DECISIONS` gains `reply` (questions about the diagram or plan, greetings, off-topic) and `unsupported`
+  (remove / merge / rename / move / restyle / simplify existing components, even mixed with additions).
+  `enforceDecision` turns both into `plan` for `answers` and `skip` turns.
+- New `reply` result action: the answer (or `UNSUPPORTED_EDIT_MESSAGE`, written in code) is stored as a `TEXT`
+  message with no phase or brief change, so a proposed plan stays generatable. A plan with nothing to add now returns
+  a `reply` instead of aborting, so it no longer shows as an error; the passthrough stays for runs on older task
+  versions. These turns make no extra model call.
+- Validation: `pnpm typecheck` and eslint passed; decision rules checked offline; live analyze on
+  `gemini-3.5-flash-lite` with a 10-component CI/CD canvas classified all five test messages as expected
+  (simplify → unsupported, name → reply, "why a separate CD Deployer" → reply, add Slack notifier → plan,
+  remove + add → unsupported).
+- Still to do: editing existing components (a separate feature) and layout quality.

lib/ai/agent-schema.ts

@@ -80,7 +80,11 @@ export const clarifyQuestionSchema = z.object({
 
 export type ClarifyQuestion = z.infer<typeof clarifyQuestionSchema>;
 
-export const TURN_DECISIONS = ["ask", "plan", "generate"] as const;
+/**
+ * - `reply`: a question or remark that needs no design change; answered in text
+ * - `unsupported`: an edit the agent cannot make yet (remove, merge, rename, rearrange)
+ */
+export const TURN_DECISIONS = ["ask", "plan", "generate", "reply", "unsupported"] as const;
 
 export type TurnDecision = (typeof TURN_DECISIONS)[number];
 
@@ -91,11 +95,16 @@ export const turnAnalysisSchema = z.object({
     .enum(TURN_DECISIONS)
     .describe(
       "'ask' to gather missing requirements, 'plan' to propose or revise a plan, " +
-        "'generate' only when a plan exists and the user approved it.",
+        "'generate' only when a plan exists and the user approved it, " +
+        "'reply' to answer a question or remark that needs no design change, " +
+        "'unsupported' when the user wants existing components removed, merged, renamed, or rearranged.",
     ),
   reply: z
     .string()
-    .describe("One or two plain sentences to the user. No markdown, and do not restate the questions."),
+    .describe(
+      "Plain text to the user, no markdown. One or two sentences, and do not restate the questions; " +
+        "for 'reply', the full answer in at most four sentences.",
+    ),
   questions: z
     .array(clarifyQuestionSchema)
     .describe(
@@ -171,6 +180,15 @@ export type DesignPlanDraft = z.infer<typeof designPlanDraftSchema>;
 export const PLAN_NOTHING_TO_ADD_MESSAGE =
   "The plan didn't include any new components to add. Say which components you want added or changed.";
 
+/** Shown for edits the agent cannot make yet; written in code so it never promises more than it can do. */
+export const UNSUPPORTED_EDIT_MESSAGE =
+  "I can add new components and connections to this diagram, but I can't remove, merge, rename, or rearrange " +
+  "existing ones yet. You can make those edits directly on the canvas, or tell me what to add.";
+
+/** Stands in for an empty answer to a `reply` turn. */
+export const REPLY_FALLBACK_MESSAGE =
+  "I'm here to help design this system. Tell me what you'd like to add, or ask about the diagram.";
+
 /** Output of one design-agent run. The server validates it before storing it. */
 export const designAgentResultSchema = z.discriminatedUnion("action", [
   z.object({
@@ -193,6 +211,11 @@ export const designAgentResultSchema = z.discriminatedUnion("action", [
     decisions: z.array(planDecisionSchema),
     brief: designBriefSchema,
   }),
+  // A text answer: no plan, no canvas change, and the session's brief and phase stay as they were.
+  z.object({
+    action: z.literal("reply"),
+    reply: z.string().min(1),
+  }),
 ]);
 
 export type DesignAgentResult = z.infer<typeof designAgentResultSchema>;

lib/ai/design-agent-engine.ts

@@ -62,11 +62,16 @@ function toMessages(payload: DesignAgentPayload): ModelMessage[] {
 /**
  * Applies the rules the model is asked to follow but cannot be trusted to:
  * no questions past the round limit or after a skip, no generating without a
- * plan, no "ask" without questions.
+ * plan, no "ask" without questions, and no text-only answer to a turn that
+ * answers questions or skips them.
  */
 export function enforceDecision(analysis: TurnAnalysis, payload: DesignAgentPayload): TurnAnalysis {
   let { decision } = analysis;
 
+  if ((decision === "reply" || decision === "unsupported") && payload.intent !== "message") {
+    decision = "plan";
+  }
+
   if (decision === "ask") {
     const mayAsk =
       payload.intent !== "skip" &&

lib/ai/prompts.ts

@@ -25,8 +25,15 @@ export const ANALYZE_SYSTEM_PROMPT = [
   `   - "plan": the brief is good enough to design, or the user asked to change an existing plan.`,
   `   - "generate": only when a plan already exists and the user's latest message approves it`,
   "     without asking for changes.",
+  `   - "reply": the message needs no change to the design — a question about the diagram, the plan,`,
+  "     or a choice in it, a general question, a greeting, or anything unrelated to the design.",
+  "     Answer it in the reply. You know nothing about the user beyond this conversation.",
+  `   - "unsupported": the user wants existing components or connections removed, merged, renamed,`,
+  "     moved, restyled, rearranged, or simplified. You can only add to a design, so choose this",
+  "     even when the request also asks for additions.",
   "   A detailed first message can go straight to a plan; a vague one should get questions.",
   "3. Write a short reply: one or two plain sentences, no markdown, no restating the questions.",
+  `   For "reply", give the full answer in at most four plain sentences.`,
   "",
   "The canvas may already hold a design; its contents are listed below. Treat the user's request as",
   "a change to that design unless they ask for a separate system: do not ask about anything the",

lib/ai/session-turns.ts

@@ -131,6 +131,7 @@ function describeFailedRun(runId: string, message: string | undefined): string {
   return describeRunFailure(message, {
     generic: GENERIC_RUN_ERROR,
     timeout: "The design agent took too long to respond. Try again.",
+    // Runs on task versions before the `reply` action aborted with this message.
     passthrough: [PLAN_NOTHING_TO_ADD_MESSAGE],
   });
 }
@@ -165,6 +166,14 @@ function outcomeFromResult(result: DesignAgentResult): TurnOutcome {
         phase: "COMPLETE",
         brief: result.brief,
       };
+    case "reply":
+      // No phase or brief: a text answer leaves the session where it was, so a proposed plan stays generatable.
+      return {
+        kind: "TEXT",
+        status: "COMPLETE",
+        content: result.reply,
+        payload: { reply: result.reply },
+      };
   }
 }
 

src/trigger/design-agent.ts

@@ -1,9 +1,11 @@
 import { mutateFlow } from "@liveblocks/react-flow/node";
-import { AbortTaskRunError, logger, metadata, task } from "@trigger.dev/sdk/v3";
+import { logger, metadata, task } from "@trigger.dev/sdk/v3";
 import type { LanguageModel } from "ai";
 
 import {
   PLAN_NOTHING_TO_ADD_MESSAGE,
+  REPLY_FALLBACK_MESSAGE,
+  UNSUPPORTED_EDIT_MESSAGE,
   type DesignAgentPayload,
   type DesignAgentResult,
   type DesignBrief,
@@ -156,6 +158,17 @@ export const designAgentTask = task({
       return { action: "ask", reply: analysis.reply, questions: analysis.questions, brief: analysis.brief };
     }
 
+    // Text-only turns end here: no plan call and no canvas change.
+    if (analysis.decision === "reply") {
+      setStage("done");
+      return { action: "reply", reply: analysis.reply.trim() || REPLY_FALLBACK_MESSAGE };
+    }
+
+    if (analysis.decision === "unsupported") {
+      setStage("done");
+      return { action: "reply", reply: UNSUPPORTED_EDIT_MESSAGE };
+    }
+
     if (analysis.decision === "generate" && payload.plan) {
       const counts = await drawPlan(model, payload.roomId, ctx.run.id, analysis.brief, payload.plan, canvas);
       setStage("done");
@@ -170,12 +183,14 @@ export const designAgentTask = task({
 
     setStage("planning");
     const plan = await draftPlan(model, analysis.brief, payload.plan, payload.input, canvas);
+    setStage("done");
     if (!plan) {
-      throw new AbortTaskRunError(PLAN_NOTHING_TO_ADD_MESSAGE);
+      // Not a failure: the request just needed nothing new. Answered in text, like a reply turn.
+      logger.log("Plan drafted with no components to add");
+      return { action: "reply", reply: PLAN_NOTHING_TO_ADD_MESSAGE };
     }
     logger.log("Plan drafted", { components: plan.components.length, decisions: plan.decisions.length });
 
-    setStage("done");
     return { action: "plan", reply: analysis.reply, plan, brief: analysis.brief };
   },
 });