@dbx-tools/appkit-mastra 0.1.26 → 0.1.27

package/README.md

@@ -199,7 +199,7 @@ once regardless of how many spaces are wired:
   Genie statement. Use only when the agent needs to read values
   to reason about them; otherwise embed `[data:<statement_id>]`.
 - `prepare_chart({statement_id, title?, description?})` - mint a
-  short `chartId`, kick off a background chart-planner task,
+  `chartId` (v4 UUID), kick off a background chart-planner task,
   cache the resolved Echarts spec under the id for one hour.
   Returns `{chartId}` synchronously so the agent can embed
   `[chart:<chartId>]` in prose without blocking.
@@ -254,8 +254,9 @@ Genie writer event flow:
 - Charts ride out-of-band: `prepare_chart` mints a `chartId`
   synchronously, the planner runs in the background and writes
   the spec to the chart cache (1h TTL). The host UI long-polls
-  `${MastraClientConfig.chartsPathTemplate}` by id and renders
-  inline at the matching `[chart:<chartId>]` marker.
+  `${MastraClientConfig.embedPathTemplate}` (`/embed/chart/:id`)
+  by id and renders inline at the matching `[chart:<chartId>]`
+  marker.
 - After a hard reload, the live `started` / `status` / `sql`
   events are gone (they only ride the writer, not persisted
   tool results); the central agent's text reply (with embedded
@@ -285,7 +286,7 @@ off the chart-planner in the background. The tool returns just
 `{ chartId }` so the model can embed `[chart:<chartId>]` in
 prose immediately without blocking on chart generation.
 
-1. Mint a short `chartId` (8 hex chars).
+1. Mint a `chartId` (v4 UUID via `crypto.randomUUID()`).
 2. Cache an empty `{ chartId }` placeholder so the first
    `fetchChart` call always sees an entry (no spurious 404 race).
 3. Fire-and-forget: resolve the dataset (trivial for
@@ -297,26 +298,27 @@ prose immediately without blocking on chart generation.
 4. Return `{ chartId }` to the LLM immediately.
 
 The host UI resolves `[chart:<chartId>]` markers by hitting
-the plugin's `GET /charts/:chartId` route, which long-polls
+the plugin's generic `GET /embed/chart/:id` route, which long-polls
 until the entry settles or the server-side timeout elapses.
 
 #### Inline placement contract
 
-The model embeds `[[chart:<chartId>]]` on its own line in its
-markdown reply at the position where the chart should appear:
+The model embeds `[chart:<chartId>]` on its own line in its
+markdown reply at the position where the chart should appear
+(the `<chartId>` is the v4 UUID the tool returned):
 
 ```markdown
 ## Audit Score
 
 Audit Score is stable at ~94%, hovering between 93.5 and 95.0.
 
-[[chart:a3f9c1d2]]
+[chart:a3f9c1d2-7b4e-4c1a-9f2d-1e6b8c0a5d31]
 
 ## Service Time
 
 Service time is the outlier at 162.5s, up from a target of 150s.
 
-[[chart:b7e2d4f1]]
+[chart:b7e2d4f1-3a9c-4e58-8d10-2f7a4b6c9e02]
 ```
 
 The chat client splits the assistant text on these markers and
@@ -330,7 +332,7 @@ frame so the layout doesn't jump as charts resolve.
 #### Trade-offs
 
 - Chart entries live in the cache with a 1h TTL. After expiry
-  the `/charts/:chartId` route returns 404. The model can call
+  the `/embed/chart/:id` route returns 404. The model can call
   `render_data` again on the next turn if the user wants the
   chart back.
 - The chart-planner is a separate model call per dataset (fast

package/dist/src/agents.js

@@ -188,7 +188,7 @@ export async function buildAgents(opts) {
     const style = resolveStyleInstructions(config);
     // Default-on protection against the model copying turn-scoped
     // chartIds from prior assistant tool results into the new
-    // turn's `[[chart:<id>]]` markers. Opt out per-plugin via
+    // turn's `[chart:<id>]` markers. Opt out per-plugin via
     // `config.stripStaleCharts: false`.
     const inputProcessors = config.stripStaleCharts === false ? [] : [stripStaleChartsProcessor];
     const agents = {};

package/dist/src/chart.d.ts

@@ -147,7 +147,7 @@ export declare function fetchChart(chartId: string, options?: FetchChartOptions)
  * an empty placeholder, and kicks off the chart-planner in the
  * background. Returns just the `chartId`; the host UI resolves
  * `[chart:<chartId>]` markers by hitting the plugin's
- * `/charts/:chartId` route.
+ * `/embed/chart/:id` route.
  *
  * For Genie statement results, prefer the Genie agent's
  * `prepare_chart` tool, which accepts a `statement_id` and

package/dist/src/chart.js

@@ -199,9 +199,7 @@ const chartPlanSchema = z.object({
  * sees a planner request, only the resolved {@link Chart}.
  */
 export const chartPlannerRequestSchema = z.object({
-    title: z
-        .string()
-        .describe(stringUtils.toDescription(`
+    title: z.string().describe(stringUtils.toDescription(`
         Concise title shown above the chart (e.g. "Top 10 SKUs by Revenue").
       `)),
     description: z
@@ -530,7 +528,7 @@ function planToEchartsOption(plan, fallbackTitle) {
  * an empty placeholder, and kicks off the chart-planner in the
  * background. Returns just the `chartId`; the host UI resolves
  * `[chart:<chartId>]` markers by hitting the plugin's
- * `/charts/:chartId` route.
+ * `/embed/chart/:id` route.
  *
  * For Genie statement results, prefer the Genie agent's
  * `prepare_chart` tool, which accepts a `statement_id` and

package/dist/src/config.d.ts

@@ -191,7 +191,7 @@ export interface MastraPluginConfig extends BasePluginConfig {
      * processor that strips `chartId` fields from prior assistant
      * tool-invocation results before they reach the model. This
      * prevents the model from reusing turn-scoped chartIds it sees
-     * in memory recall (which would leave `[[chart:<id>]]` markers
+     * in memory recall (which would leave `[chart:<id>]` markers
      * pointing at writer events that no longer exist).
      *
      * Set to `false` to opt out - useful if a non-default agent

package/dist/src/genie.js

@@ -614,9 +614,13 @@ function buildPrepareChartTool(opts) {
             `
         To display the chart in your reply, embed
         \`[chart:<chartId>]\` on its own line at the position you
-        want it to appear. The tool returns immediately - do NOT
-        wait or call it again to "check progress"; the chart
-        resolves asynchronously on the host UI's side.
+        want it to appear, using the EXACT \`chartId\` string this
+        call returned. Never construct a chart id yourself (it is
+        not the \`statement_id\` or any variation of it) - only a
+        value returned by this tool resolves to a real chart. The
+        tool returns immediately - do NOT wait or call it again to
+        "check progress"; the chart resolves asynchronously on the
+        host UI's side.
       `,
             `
         Use this only when the data has a story a chart conveys
@@ -728,6 +732,17 @@ export const GENIE_INSTRUCTIONS = stringUtils.toDescription([
               chart should appear. Use a chart when the data has a
               story a visual conveys better than a table (trends,
               rankings, distributions, parts-of-a-whole).
+
+              NEVER invent or hand-build a \`<chartId>\`. A valid
+              \`<chartId>\` is the opaque token a \`prepare_chart\`
+              call returned to you in THIS turn - nothing else. It
+              is NOT a \`statement_id\`, and it is NOT a
+              \`statement_id\` prefix with a label appended (e.g.
+              \`01f1...-region-fill\`). If you have not called
+              \`prepare_chart\` and received an id back, do not write
+              a \`[chart:...]\` marker at all - use \`[data:...]\`
+              instead. A fabricated chart id renders nothing and
+              wastes a request.
             `,
                     ],
                 },

package/dist/src/intercept.d.ts

@@ -0,0 +1,48 @@
+/**
+ * Server-side Server-Sent-Events frame interceptor for the native
+ * Mastra agent `/stream` endpoint.
+ *
+ * The Mastra stream (`@mastra/express`) emits one `data: <json>\n\n`
+ * SSE frame per chunk, where the decoded JSON carries a
+ * discriminating `type` (`text-delta`, `reasoning-delta`,
+ * `tool-call`, `step-finish`, ...) and is terminated by a
+ * `data: [DONE]` sentinel. {@link installStreamEventInterceptor} wraps
+ * an Express response and runs a caller-supplied
+ * {@link StreamFrameInterceptor} over each decoded frame so the caller
+ * can keep, rewrite, or drop it; every non-`data:` byte (SSE
+ * keep-alive comments, the `[DONE]` sentinel, blank padding) passes
+ * through verbatim.
+ *
+ * The wrap is inert unless the response turns out to be a streamed
+ * `200 text/event-stream` body, so non-streaming JSON routes and
+ * error responses are never touched.
+ */
+import type express from "express";
+/**
+ * Per-frame interceptor invoked with the `JSON.parse`d object of every
+ * `data: {json}` frame on the Mastra `/stream` response. The return
+ * value decides the frame's fate:
+ *
+ * - `true`: forward the frame's original bytes unchanged (no
+ *   re-serialization).
+ * - `false`: drop the frame entirely.
+ * - `{ replace }`: re-emit the frame with the original `data:` prefix
+ *   and trailing whitespace preserved and only the JSON body replaced
+ *   by `JSON.stringify(replace)`.
+ */
+export type StreamFrameInterceptor = (chunk: unknown) => {
+    replace: unknown;
+} | true | false;
+/**
+ * Wrap `res.write` / `res.end` so each SSE frame on a streamed Mastra
+ * response is run through `interceptor` (keep / rewrite / drop).
+ *
+ * Interception only engages once the response is confirmed to be a
+ * `200 text/event-stream` body (decided lazily on the first write,
+ * after the handler has set status + content-type); any other
+ * response streams through unchanged. A {@link StringDecoder}
+ * reassembles multi-byte UTF-8 sequences that straddle chunk
+ * boundaries, and a running buffer holds the trailing partial frame
+ * until its `\n\n` arrives.
+ */
+export declare function installStreamEventInterceptor(res: express.Response, interceptor: StreamFrameInterceptor): void;

package/dist/src/intercept.js

@@ -0,0 +1,167 @@
+/**
+ * Server-side Server-Sent-Events frame interceptor for the native
+ * Mastra agent `/stream` endpoint.
+ *
+ * The Mastra stream (`@mastra/express`) emits one `data: <json>\n\n`
+ * SSE frame per chunk, where the decoded JSON carries a
+ * discriminating `type` (`text-delta`, `reasoning-delta`,
+ * `tool-call`, `step-finish`, ...) and is terminated by a
+ * `data: [DONE]` sentinel. {@link installStreamEventInterceptor} wraps
+ * an Express response and runs a caller-supplied
+ * {@link StreamFrameInterceptor} over each decoded frame so the caller
+ * can keep, rewrite, or drop it; every non-`data:` byte (SSE
+ * keep-alive comments, the `[DONE]` sentinel, blank padding) passes
+ * through verbatim.
+ *
+ * The wrap is inert unless the response turns out to be a streamed
+ * `200 text/event-stream` body, so non-streaming JSON routes and
+ * error responses are never touched.
+ */
+import { StringDecoder } from "node:string_decoder";
+/** SSE event separator used by the Mastra stream wire format. */
+const FRAME_DELIMITER = "\n\n";
+/**
+ * Matches a whole SSE frame whose `data:` payload is a JSON object,
+ * capturing it in three parts: the `data:` field prefix (including
+ * whatever inter-token whitespace the producer used), the JSON object
+ * body, and any trailing whitespace. Anchored end-to-end so it only
+ * matches a single self-contained object frame; anything else (SSE
+ * comments, the `[DONE]` sentinel, text/array payloads) fails the
+ * match and is forwarded verbatim without a parse attempt. On rewrite
+ * the captured prefix/suffix are re-emitted as-is so only the body
+ * changes.
+ */
+const DATA_OBJECT_RE = /^(data:\s*)(\{[\s\S]*\})(\s*)$/;
+/**
+ * True when `res` is a `200 text/event-stream` body - the only
+ * response shape we intercept. Content-type based so it tracks the
+ * Mastra `/stream` route without hard-coding the path, and naturally
+ * skips the JSON routes (history, models, charts, statements) and any
+ * non-200 error response.
+ */
+function isEventStream(res) {
+    if (res.statusCode !== 200)
+        return false;
+    const contentType = String(res.getHeader("content-type") ?? "").toLowerCase();
+    return contentType.includes("text/event-stream");
+}
+/**
+ * Run `interceptor` over a single SSE frame. Forwards anything that
+ * isn't a parseable `data: {json}` object frame verbatim (comments,
+ * the `[DONE]` sentinel, blank padding, unparseable / non-object
+ * payloads) so the interceptor only ever sees real chunk objects.
+ * Returns the frame string to emit, or `undefined` to drop the frame.
+ */
+function interceptFrame(frame, interceptor) {
+    const match = DATA_OBJECT_RE.exec(frame);
+    if (!match)
+        return frame;
+    const [, prefix, body, suffix] = match;
+    let parsed;
+    try {
+        parsed = JSON.parse(body);
+    }
+    catch {
+        // Forward unparseable payloads untouched - feeding a partial /
+        // malformed frame to the interceptor risks corrupting bytes the
+        // client could still have handled.
+        return frame;
+    }
+    const result = interceptor(parsed);
+    if (result === true)
+        return frame;
+    if (result === false)
+        return undefined;
+    // Preserve the exact captured prefix/suffix so only the JSON body
+    // changes; the framing the client parses stays byte-identical.
+    return `${prefix}${JSON.stringify(result.replace)}${suffix}`;
+}
+/**
+ * Wrap `res.write` / `res.end` so each SSE frame on a streamed Mastra
+ * response is run through `interceptor` (keep / rewrite / drop).
+ *
+ * Interception only engages once the response is confirmed to be a
+ * `200 text/event-stream` body (decided lazily on the first write,
+ * after the handler has set status + content-type); any other
+ * response streams through unchanged. A {@link StringDecoder}
+ * reassembles multi-byte UTF-8 sequences that straddle chunk
+ * boundaries, and a running buffer holds the trailing partial frame
+ * until its `\n\n` arrives.
+ */
+export function installStreamEventInterceptor(res, interceptor) {
+    const origWrite = res.write.bind(res);
+    const origEnd = res.end.bind(res);
+    const decoder = new StringDecoder("utf8");
+    let engaged;
+    let buffer = "";
+    // Append decoded text, emit every whole `\n\n`-delimited frame the
+    // interceptor keeps, and retain any trailing partial frame in
+    // `buffer`. On `final`, the residual buffer is flushed as the last
+    // frame so a stream that doesn't end on a delimiter isn't dropped.
+    const intercept = (text, final) => {
+        buffer += text;
+        const out = [];
+        let idx;
+        while ((idx = buffer.indexOf(FRAME_DELIMITER)) !== -1) {
+            const frame = buffer.slice(0, idx);
+            buffer = buffer.slice(idx + FRAME_DELIMITER.length);
+            const kept = interceptFrame(frame, interceptor);
+            if (kept !== undefined) {
+                out.push(kept, FRAME_DELIMITER);
+            }
+        }
+        if (final && buffer.length > 0) {
+            const kept = interceptFrame(buffer, interceptor);
+            if (kept !== undefined)
+                out.push(kept);
+            buffer = "";
+        }
+        return out.length === 0 ? "" : out.join("");
+    };
+    const toText = (chunk) => {
+        if (typeof chunk === "string")
+            return chunk;
+        if (Buffer.isBuffer(chunk))
+            return decoder.write(chunk);
+        if (chunk instanceof Uint8Array) {
+            return decoder.write(Buffer.from(chunk.buffer, chunk.byteOffset, chunk.byteLength));
+        }
+        return "";
+    };
+    res.write = function patchedWrite(chunk, encodingOrCb, cb) {
+        if (engaged === undefined)
+            engaged = isEventStream(res);
+        if (!engaged) {
+            return origWrite(chunk, encodingOrCb, cb);
+        }
+        const callback = typeof encodingOrCb === "function" ? encodingOrCb : cb;
+        const out = intercept(toText(chunk), false);
+        if (out.length === 0) {
+            // The chunk only advanced a partial frame; nothing to forward
+            // yet. Honor the write callback so backpressure-aware writers
+            // don't stall waiting on an ack that never comes.
+            if (callback)
+                queueMicrotask(() => callback());
+            return true;
+        }
+        return origWrite(out, callback);
+    };
+    res.end = function patchedEnd(chunk, encodingOrCb, cb) {
+        if (engaged === undefined)
+            engaged = isEventStream(res);
+        if (!engaged) {
+            return origEnd(chunk, encodingOrCb, cb);
+        }
+        const callback = typeof chunk === "function"
+            ? chunk
+            : typeof encodingOrCb === "function"
+                ? encodingOrCb
+                : cb;
+        const text = (typeof chunk !== "function" && chunk !== undefined ? toText(chunk) : "") +
+            decoder.end();
+        const out = intercept(text, true);
+        if (out.length > 0)
+            origWrite(out);
+        return origEnd(callback);
+    };
+}

package/dist/src/plugin.d.ts

@@ -110,8 +110,8 @@ export declare class MastraPlugin extends Plugin<MastraPluginConfig> {
     clientConfig(): Record<string, unknown>;
     injectRoutes(router: IAppRouter): void;
     /**
-     * Implementation backing the `/statements/:statementId` route.
-     * Runs inside the AppKit user-context proxy so
+     * Implementation backing the `data` embed resolver
+     * (`GET /embed/data/:id`). Runs inside the AppKit user-context proxy so
      * `getExecutionContext()` returns the OBO-scoped workspace
      * client, then reuses the same `fetchStatementData` pipeline
      * the `get_statement` tool runs so the LLM and the UI see the

package/dist/src/plugin.js

@@ -27,7 +27,7 @@
  *   AI SDK transport URL is `/api/mastra/route/chat/<agentId>`.
  */
 import { genie, getExecutionContext, lakebase, Plugin, toPlugin, } from "@databricks/appkit";
-import { appkitUtils, logUtils } from "@dbx-tools/shared";
+import { appkitUtils, commonUtils, logUtils } from "@dbx-tools/shared";
 import { chatRoute } from "@mastra/ai-sdk";
 import { Mastra } from "@mastra/core/mastra";
 import express from "express";
@@ -37,6 +37,7 @@ import { historyRoute } from "./history.js";
 import { createMemoryBuilder, needsLakebase } from "./memory.js";
 import { buildObservability } from "./observability.js";
 import { attachRoutePatchMiddleware, MastraServer } from "./server.js";
+import { installStreamEventInterceptor, } from "./intercept.js";
 import { clearServingEndpointsCache, listServingEndpoints, resolveServingConfig, } from "./serving.js";
 import { fetchStatementData, isStatementNotFoundError, STATEMENT_ROW_CAP, } from "./statement.js";
 const GENIE_MANIFEST = appkitUtils.data(genie).plugin.manifest;
@@ -173,8 +174,7 @@ export class MastraPlugin extends Plugin {
             modelsPath: `${basePath}/models`,
             historyPath: `${basePath}/route/history`,
             historyPathTemplate: `${basePath}/route/history/:agentId`,
-            chartsPathTemplate: `${basePath}/charts/:chartId`,
-            statementsPathTemplate: `${basePath}/statements/:statementId`,
+            embedPathTemplate: `${basePath}/embed/:type/:id`,
             defaultAgent: this.built?.defaultAgentId ?? FALLBACK_AGENT_ID,
             agents: Object.keys(this.built?.agents ?? {}),
         };
@@ -192,112 +192,95 @@ export class MastraPlugin extends Plugin {
                 .then((endpoints) => res.json({ endpoints }))
                 .catch(next);
         });
-        // `GET /charts/:chartId` long-polls the chart cache.
-        //
-        // Charts kicked off by the Genie agent's `prepare_chart` tool
-        // (or the system `render_data` tool) resolve asynchronously
-        // and land in the AppKit cache under a chartId. The host UI
-        // resolves `[chart:<chartId>]` markers by hitting this route,
-        // which blocks until the entry settles (`result` or `error`
-        // set) or the server-side budget elapses (then returns the
-        // last seen still-processing entry so the client can poll
-        // again).
+        // `GET /embed/:type/:id` is the single resolver for every embed
+        // marker the agent emits in prose (`[chart:<id>]`,
+        // `[data:<id>]`, ...). `:type` selects a resolver from the
+        // registry below; `:id` is that resolver's lookup key. The
+        // grammar (see `marker.ts`) is type-agnostic on purpose - new
+        // embed kinds are added by registering a resolver here, with no
+        // client or grammar change.
         //
         // Status codes:
-        //   - 200 with JSON body when the entry is found (any state).
-        //   - 404 when the chartId doesn't exist or its 1h TTL elapsed.
+        //   - 200 with the resolver's JSON body when the id resolves.
+        //   - 404 when `:type` isn't registered (unsupported embed
+        //     type) OR a registered resolver can't find `:id` (unknown
+        //     / expired - e.g. a chart past its 1h TTL or a fabricated
+        //     id the model never minted).
+        //   - 400 when `:id` is empty.
         //
-        // Query params:
-        //   - `timeoutMs` (optional, default 60000): how long to long-poll
-        //     before returning a still-processing entry.
+        // Per-type query knobs and behavior:
+        //   - `chart`: long-polls the chart cache until the entry
+        //     settles (`result` / `error`) or the budget elapses (then
+        //     returns the still-processing entry to poll again).
+        //     `?timeoutMs=<n>` (default 60s, capped 5min) tunes it.
+        //   - `data`: one OBO-scoped Statement Execution fetch.
+        //     `?limit=<n>` caps rows (clamped to STATEMENT_ROW_CAP).
         //
-        // Wired with `req.signal` so a closed connection unblocks the
-        // poll immediately and the request thread frees up. Cache
-        // failures bubble through `next(err)` to Express's default
-        // error handler.
-        router.get("/charts/:chartId", (req, res, next) => {
-            const chartId = req.params["chartId"];
-            if (!chartId) {
-                res.status(400).json({ error: "chartId is required" });
+        // Built once (this handler is registered once) and keyed by the
+        // raw `:type` token. Each resolver gets the request (for query
+        // parsing + OBO scoping) and an `AbortSignal` bridged off the
+        // connection `close` event so a long-poll unblocks the instant
+        // the client disconnects. `undefined` from a resolver maps to a
+        // clean 404; thrown errors bubble through `next(err)`.
+        const embedResolvers = {
+            chart: (req, id, signal) => {
+                const timeoutMs = parseTimeoutMs(req.query["timeoutMs"]);
+                return fetchChart(id, {
+                    ...(timeoutMs !== undefined ? { timeoutMs } : {}),
+                    signal,
+                });
+            },
+            data: (req, id, signal) => {
+                const limit = parseStatementLimit(req.query["limit"]);
+                return this.userScopedSelf(req).fetchStatement(id, {
+                    ...(limit !== undefined ? { limit } : {}),
+                    signal,
+                });
+            },
+        };
+        router.get("/embed/:type/:id", (req, res, next) => {
+            const type = req.params["type"] ?? "";
+            const id = req.params["id"];
+            const resolve = embedResolvers[type];
+            if (!resolve) {
+                res.status(404).json({ error: `unsupported embed type: ${type}` });
                 return;
             }
-            const timeoutMs = parseTimeoutMs(req.query["timeoutMs"]);
-            // Express's `req` predates `AbortSignal`. Bridge the
-            // `close` event onto an `AbortController` so a closed
-            // connection unblocks the long-poll immediately and frees
-            // the request thread. The `close` listener also fires on
-            // a normal completion, but at that point we no longer
-            // care about the abort and the listener is GC'd with the
-            // request.
-            const controller = new AbortController();
-            req.on("close", () => controller.abort());
-            fetchChart(chartId, {
-                ...(timeoutMs !== undefined ? { timeoutMs } : {}),
-                signal: controller.signal,
-            })
-                .then((entry) => {
-                if (!entry) {
-                    res.status(404).json({ error: "chart not found" });
-                    return;
-                }
-                res.json(entry);
-            })
-                .catch(next);
-        });
-        // `GET /statements/:statementId` resolves a Genie / Statement
-        // Execution result for inline rendering as a table.
-        //
-        // The agent embeds `[data:<statement_id>]` markers in prose
-        // (see `GENIE_INSTRUCTIONS`) whenever it wants the host UI to
-        // display a result set without the model re-spelling every
-        // row in markdown. This route is what those markers resolve
-        // against - one OBO-scoped fetch returns the columns + rows
-        // so the client renders an AppKit Table inline.
-        //
-        // Status codes:
-        //   - 200 with `StatementData` body when the statement is found.
-        //   - 404 when the workspace can't find / read the statement.
-        //
-        // Query params:
-        //   - `limit` (optional): row cap. Clamped to
-        //     {@link STATEMENT_ROW_CAP} so a runaway result set
-        //     can't hose the response.
-        //
-        // OBO-scoped via {@link userScopedSelf} so the caller's
-        // workspace permissions apply; SDK errors propagate through
-        // `next(err)` to Express's default error handler.
-        router.get("/statements/:statementId", (req, res, next) => {
-            const statementId = req.params["statementId"];
-            if (!statementId) {
-                res.status(400).json({ error: "statementId is required" });
+            if (!id) {
+                res.status(400).json({ error: "id is required" });
                 return;
             }
-            const limit = parseStatementLimit(req.query["limit"]);
+            // Express's `req` predates `AbortSignal`; bridge the `close`
+            // event onto an `AbortController` so a closed connection
+            // unblocks any long-poll immediately and frees the request
+            // thread. The listener is GC'd with the request on normal
+            // completion.
             const controller = new AbortController();
             req.on("close", () => controller.abort());
-            this.userScopedSelf(req)
-                .fetchStatement(statementId, {
-                ...(limit !== undefined ? { limit } : {}),
-                signal: controller.signal,
-            })
-                .then((data) => {
-                if (!data) {
-                    res.status(404).json({ error: "statement not found" });
+            resolve(req, id, controller.signal)
+                .then((entry) => {
+                if (entry === undefined) {
+                    res.status(404).json({ error: `${type} not found` });
                     return;
                 }
-                res.json(data);
+                res.json(entry);
             })
                 .catch(next);
         });
         router.use("", (req, res, next) => {
             if (!this.mastraApp)
                 return res.status(503).end();
+            // Run each Mastra `/stream` SSE frame through the interceptor
+            // (keep / rewrite / drop). Only engages on a
+            // `200 text/event-stream` body, so the JSON routes above and any
+            // error response are untouched.
+            installStreamEventInterceptor(res, interceptStreamFrame);
             return this.userScopedSelf(req).mastraApp(req, res, next);
         });
     }
     /**
-     * Implementation backing the `/statements/:statementId` route.
-     * Runs inside the AppKit user-context proxy so
+     * Implementation backing the `data` embed resolver
+     * (`GET /embed/data/:id`). Runs inside the AppKit user-context proxy so
      * `getExecutionContext()` returns the OBO-scoped workspace
      * client, then reuses the same `fetchStatementData` pipeline
      * the `get_statement` tool runs so the LLM and the UI see the
@@ -427,7 +410,7 @@ export class MastraPlugin extends Plugin {
 }
 /**
  * Parse the optional `?timeoutMs=<n>` query parameter from a
- * `GET /charts/:chartId` request. Accepts a positive integer up
+ * `GET /embed/chart/:id` request. Accepts a positive integer up
  * to 5 minutes (clamped) and rejects everything else as
  * `undefined` so {@link fetchChart} falls back to its default.
  * Express produces `string | string[] | undefined`; we normalize
@@ -444,7 +427,7 @@ function parseTimeoutMs(raw) {
 }
 /**
  * Parse the optional `?limit=<n>` query parameter from a
- * `GET /statements/:statementId` request. Accepts a non-negative
+ * `GET /embed/data/:id` request. Accepts a non-negative
  * integer and lets the route clamp to `STATEMENT_ROW_CAP`;
  * rejects anything else as `undefined` so the route falls back
  * to the server-side cap.
@@ -458,4 +441,37 @@ function parseStatementLimit(raw) {
         return undefined;
     return Math.floor(n);
 }
+/**
+ * {@link StreamFrameInterceptor} for Mastra `/stream` responses.
+ *
+ * Removes large, redundant payloads from terminal `step-finish`,
+ * `finish`, and `tool-result` frames before they reach the browser.
+ * These frames often repeat information already delivered incrementally
+ * via streamed text and tool events, including full tool outputs,
+ * accumulated responses, message history, SQL results, chart data, and
+ * other large result payloads.
+ *
+ * The interceptor deletes heavyweight payload properties
+ * (`output`, `messages`, `response`, and `result`) while preserving the
+ * event envelope and lifecycle metadata required by the client.
+ *
+ * Deletion is key based, so streams that do not contain these
+ * fields are passed through unchanged.
+ */
+const interceptStreamFrame = (chunk) => {
+    if (!commonUtils.isRecord(chunk))
+        return true;
+    if (!["step-finish", "finish", "tool-result"].find((type) => type === chunk.type))
+        return true;
+    const payload = chunk.payload;
+    if (!commonUtils.isRecord(payload))
+        return true;
+    const trimmedPayload = commonUtils.deleteKeys(payload, [
+        "output",
+        "messages",
+        "response",
+        "result",
+    ]);
+    return trimmedPayload ? { replace: chunk } : true;
+};
 export const mastra = toPlugin(MastraPlugin);