Skip to content
Open
Show file tree
Hide file tree
Changes from all commits
Commits
File filter

Filter by extension

Filter by extension

Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
6 changes: 6 additions & 0 deletions .server-changes/query-boundary-and-retry-cap.md
Original file line number Diff line number Diff line change
@@ -0,0 +1,6 @@
---
area: webapp
type: improvement
---

Queries stay read-only, and the agent now stops after a few failed queries in a row and answers with what it found instead of spending the whole reply retrying.
1 change: 1 addition & 0 deletions apps/webapp/app/services/queryService.server.ts
Original file line number Diff line number Diff line change
Expand Up @@ -397,6 +397,7 @@ export async function executeQuery<TOut extends z.ZodSchema>(
...getDefaultClickhouseSettings(),
...queryCacheSettings,
...baseOptions.clickhouseSettings, // Allow caller overrides if needed
readonly: "1", // Not overridable: every query through here is read-only.
},
querySettings: {
maxRows: env.QUERY_CLICKHOUSE_MAX_RETURNED_ROWS,
Expand Down
168 changes: 168 additions & 0 deletions apps/webapp/test/queryRouteReadOnly.test.ts
Original file line number Diff line number Diff line change
@@ -0,0 +1,168 @@
import { generateJWT } from "@trigger.dev/core/v3/jwt";
import { beforeEach, describe, expect, it, vi } from "vitest";

/**
* The query API is read-only, and the grammar is what enforces it. A parser test alone would
* stay green if the route ever compiled agent SQL somewhere else, so these drive the real route
* with a real signed environment JWT and stub only the ClickHouse client. A write must be
* refused before anything reaches ClickHouse.
*/

const ENVIRONMENT_ID = "env_1234";
const API_KEY = "tr_dev_abcdefghijklmnop";

const environment = {
id: ENVIRONMENT_ID,
type: "DEVELOPMENT",
slug: "dev",
branchName: null,
apiKey: API_KEY,
organizationId: "org_1",
projectId: "proj_1",
archivedAt: null,
concurrencyLimitBurstFactor: { toNumber: () => 1 },
maximumConcurrencyLimit: 10,
project: { id: "proj_1", externalRef: "proj_ref", deletedAt: null },
organization: { id: "org_1" },
orgMember: null,
parentEnvironment: null,
};

const mocks = vi.hoisted(() => ({
runtimeEnvironmentFindFirst: vi.fn(),
queryWithStats: vi.fn(),
customerQueryCreate: vi.fn(),
}));

vi.mock("~/db.server", () => {
const client = {
runtimeEnvironment: {
findFirst: mocks.runtimeEnvironmentFindFirst,
findMany: async () => [],
},
revokedApiKey: { findMany: async () => [], findFirst: async () => null },
project: { findMany: async () => [] },
customerQuery: { findFirst: async () => null, create: mocks.customerQueryCreate },
};
return { prisma: client, $replica: client };
});
Comment on lines +37 to +48

Copy link
Copy Markdown
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

🟡 New tests rely on mocks, which the repository's testing rules forbid

The new test files replace real dependencies with stand-ins (vi.mock("~/db.server", ...) at apps/webapp/test/queryRouteReadOnly.test.ts:37-48, plus mocks of the environment config, ClickHouse factory, limiter and logger, and vi.fn() fakes in internal-packages/dashboard-agent/src/tool-query-retry-cap.test.ts:58-64), which the project's testing rules prohibit.
Impact: The checked-in tests conflict with the repository's stated testing approach, so they may need rewriting before they can be trusted as coverage.

Rule text and where it is violated

AGENTS.md states: "We use vitest exclusively. Never mock anything - use testcontainers instead." and ai/references/tests.md (referenced by .cursor/rules/writing-tests.mdc) repeats: "Do not mock anything. Do not use mocks in tests. Do not use spies in tests."

apps/webapp/test/queryRouteReadOnly.test.ts:31-83 builds a hoisted mock registry and mocks ~/db.server, ~/env.server, ~/services/clickhouse/clickhouseFactoryInstance.server, ~/services/platform.v3.server, ~/services/queryConcurrencyLimiter.server, ~/services/logger.server, and two service modules. internal-packages/dashboard-agent/src/tool-query-retry-cap.test.ts uses vi.fn() spies for postQuery.

(Note that many existing webapp tests use vi.mock too, so this is a widespread pre-existing deviation rather than something unique to this PR.)

Open in Devin Review

Was this helpful? React with 👍 or 👎 to provide feedback.

vi.mock("~/env.server", () => ({
env: {
SESSION_SECRET: "test-session-secret",
QUERY_CLICKHOUSE_MAX_EXECUTION_TIME: "30",
QUERY_CLICKHOUSE_MAX_MEMORY_USAGE: 1000000,
QUERY_CLICKHOUSE_MAX_AST_ELEMENTS: 50000,
QUERY_CLICKHOUSE_MAX_EXPANDED_AST_ELEMENTS: 500000,
QUERY_CLICKHOUSE_MAX_BYTES_BEFORE_EXTERNAL_GROUP_BY: 1000000,
QUERY_CLICKHOUSE_MAX_RETURNED_ROWS: 1000,
},
}));
vi.mock("~/services/clickhouse/clickhouseFactoryInstance.server", () => ({
clickhouseFactory: {
getClickhouseForOrganization: async () => ({
reader: { queryWithStats: mocks.queryWithStats },
}),
},
}));
vi.mock("~/services/platform.v3.server", () => ({ getLimit: async () => 30 }));
vi.mock("~/services/queryConcurrencyLimiter.server", () => ({
queryConcurrencyLimiter: {
acquire: async () => ({ success: true }),
release: async () => {},
},
DEFAULT_ORG_CONCURRENCY_LIMIT: 10,
GLOBAL_CONCURRENCY_LIMIT: 100,
}));
vi.mock("~/services/logger.server", () => ({
logger: { debug: vi.fn(), error: vi.fn(), warn: vi.fn(), info: vi.fn() },
}));
vi.mock("~/v3/services/worker/workerGroupTokenService.server", () => ({
WorkerGroupTokenService: class {},
}));
vi.mock("~/v3/services/common.server", () => ({ ServiceValidationError: class extends Error {} }));
vi.mock("@internal/run-engine", () => ({ EngineServiceValidationError: class extends Error {} }));
Comment on lines +31 to +83

Copy link
Copy Markdown
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

🔒 Security & Privacy | 🟠 Major | 🏗️ Heavy lift

Use real integration dependencies for this route-level suite.

This test replaces the Prisma and ClickHouse boundaries with mocks. It cannot detect integration failures in authentication, query dispatch, or ClickHouse settings serialization.

Use Testcontainers for the database and ClickHouse dependencies. If the intent is only grammar validation, extract that behavior into a pure unit test.

As per coding guidelines, “Use Vitest exclusively and never mock dependencies; use Testcontainers for integration dependencies.”

Source: Coding guidelines


import { action } from "~/routes/api.v1.query";
import { executeQuery } from "~/services/queryService.server";

/** The claims the env-JWT exchange mints (api.v1.projects.$projectRef.$env.jwt.ts). */
function mintEnvJwt(scopes: string[]) {
return generateJWT({
secretKey: API_KEY,
payload: {
sub: ENVIRONMENT_ID,
pub: true,
scopes,
act: { sub: "usr_1", client: "dashboard-agent" },
},
expirationTime: "1h",
});
}

async function runQuery(query: string): Promise<{ status: number; body: any }> {
const jwt = await mintEnvJwt(["read:query"]);
const response = await action({
request: new Request("https://api.trigger.dev/api/v1/query", {
method: "POST",
headers: { Authorization: `Bearer ${jwt}`, "Content-Type": "application/json" },
body: JSON.stringify({ query }),
}),
params: {},
context: {},
} as any);
return { status: response.status, body: await response.json() };
}

describe("the query API route", () => {
beforeEach(() => {
vi.clearAllMocks();
mocks.runtimeEnvironmentFindFirst.mockResolvedValue(environment);
mocks.customerQueryCreate.mockResolvedValue({ id: "cq_1" });
mocks.queryWithStats.mockReturnValue(async () => [null, { rows: [], stats: {} }]);
});

// Pins the seam the two refusals assert against: a read really does reach ClickHouse here,
// so `not.toHaveBeenCalled()` below means refused, not unreachable.
it("runs a read against ClickHouse", async () => {
const result = await runQuery("SELECT count() FROM runs");

expect(result.status).toBe(200);
expect(mocks.queryWithStats).toHaveBeenCalled();
});

it("refuses a write smuggled in as a second statement", async () => {
const result = await runQuery("SELECT 1 FROM runs; DROP TABLE runs");

expect(result.status).toBe(400);
expect(mocks.queryWithStats).not.toHaveBeenCalled();
});

it("refuses a mutating statement", async () => {
const result = await runQuery("INSERT INTO runs (task_identifier) VALUES ('x')");

expect(result.status).toBe(400);
expect(mocks.queryWithStats).not.toHaveBeenCalled();
});
});

describe("the query service", () => {
beforeEach(() => {
vi.clearAllMocks();
mocks.queryWithStats.mockReturnValue(async () => [null, { rows: [], stats: {} }]);
});

it("keeps ClickHouse read-only when a caller overrides the settings", async () => {
await executeQuery({
name: "test-query",
query: "SELECT count() FROM runs",
scope: "environment",
organizationId: "org_1",
projectId: "proj_1",
environmentId: ENVIRONMENT_ID,
clickhouseSettings: { readonly: "0" },
} as any);

expect(mocks.queryWithStats).toHaveBeenCalled();
expect(mocks.queryWithStats.mock.calls[0][0].settings.readonly).toBe("1");
});
});
23 changes: 22 additions & 1 deletion internal-packages/dashboard-agent/src/tool-api.ts
Original file line number Diff line number Diff line change
Expand Up @@ -177,6 +177,9 @@ export function withLiveState(metrics: unknown, queueType: "task" | "custom", li
};
}

/** Failed `run_query` calls in a row before the tool tells the model to stop and answer. */
export const MAX_CONSECUTIVE_QUERY_FAILURES = 3;

export function buildApiTools(args: {
ctx: DashboardAgentToolContext;
client: DashboardAgentApiClient;
Expand All @@ -186,6 +189,12 @@ export function buildApiTools(args: {
const { userActorToken, projectRef, environmentName, environmentBranch } = ctx;
const { origin, hasAuth, envApiGet, postQuery, validateChartQuery } = client;

// A failed query hands the model the database error to fix, and it usually does. When it
// doesn't, the only other limit is the turn's 10 steps, so one broken query can eat the
// whole turn and leave the user with no answer at all. This tool set is built per turn,
// so the counter caps consecutive failures within one turn.
let consecutiveQueryFailures = 0;
Comment on lines +192 to +196

Copy link
Copy Markdown
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

🔍 Counter lifetime depends on the tools hook being invoked once per turn

The counter's per-turn scope rests on buildTurnTools being called once per turn: dashboard-agent.ts:381 wires it into the agent's tools hook, and agent-runtime.ts:388 only reads locals.get(dashboardAgentToolsKey) (set solely by tests/evals) before building a fresh set — so nothing memoizes the tool set. If the SDK ever invokes the tools hook per step rather than per turn, each run_query call would get a fresh counter and the cap would never fire; conversely if the set were cached per chat, stale failure counts would leak into later turns. Worth confirming against the chat SDK's hook contract. Note watch-actions.ts:780 builds its own independent tool set, so watch-driven queries have a separate counter.

Open in Devin Review

Was this helpful? React with 👍 or 👎 to provide feedback.

Comment on lines +192 to +196

@devin-ai-integration devin-ai-integration Bot Aug 10, 2026

Copy link
Copy Markdown
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

🔍 Chart-query validation failures bypass the same cap

render_view validates chart queries through the same postQuery seam (internal-packages/dashboard-agent/src/tool-api.ts:407) but does not touch consecutiveQueryFailures, so a model that loops on a broken chart query can still burn the whole turn — the failure mode the cap was added to prevent, just through a different tool. Also note the counter lives on the closure, so it only caps a turn if buildApiTools really is invoked once per turn (internal-packages/dashboard-agent/src/dashboard-agent.ts:381); the eval path injects a fixed tool set via dashboardAgentToolsKey and would share one counter across turns.

Open in Devin Review

Was this helpful? React with 👍 or 👎 to provide feedback.


return {
list_projects: tool({
...listProjectsSchema,
Expand Down Expand Up @@ -341,7 +350,19 @@ export function buildApiTools(args: {
execute: async ({ query, period }) => {
const result = await postQuery(query, period);
if (isEnvUnavailable(result)) return envUnavailableError(result, "query");
if (!result.ok) return { error: result.error };
if (!result.ok) {
// Only SQL errors count toward the cap; transport errors are transient.
if (result.kind === "query") {
consecutiveQueryFailures++;
if (consecutiveQueryFailures >= MAX_CONSECUTIVE_QUERY_FAILURES) {
return {
error: `${result.error} That is ${consecutiveQueryFailures} queries in a row that failed. Stop querying and answer the user with what you already have.`,
};
}
}
return { error: result.error };
}
Comment thread
devin-ai-integration[bot] marked this conversation as resolved.
consecutiveQueryFailures = 0;
Comment thread
devin-ai-integration[bot] marked this conversation as resolved.
const cap = 200;
const rows = result.rows;
return { rows: rows.slice(0, cap), rowCount: rows.length, truncated: rows.length > cap };
Expand Down
90 changes: 90 additions & 0 deletions internal-packages/dashboard-agent/src/tool-query-retry-cap.test.ts
Original file line number Diff line number Diff line change
@@ -0,0 +1,90 @@
import { describe, expect, it, vi } from "vitest";
import { buildApiTools, MAX_CONSECUTIVE_QUERY_FAILURES } from "./tool-api";
import type { DashboardAgentApiClient } from "./tool-api-client";

/**
* A failed query hands the model the database error to fix. Without a cap, the only other
* limit is the turn's step budget, so a model that keeps rewriting the same broken query
* burns the whole turn and the user gets no answer. After three failures in a row the tool
* tells it to stop and answer.
*/

function queryTool(postQuery: DashboardAgentApiClient["postQuery"]) {
const client = {
origin: "https://api.example.com",
hasAuth: true,
envApiGet: async () => ({ ok: false as const, status: 500 }),
postQuery,
validateChartQuery: async () => null,
} as unknown as DashboardAgentApiClient;
const tools = buildApiTools({
ctx: { userActorToken: "uat", apiOrigin: client.origin },
client,
renderInvestigations: (() => []) as any,
});
return (query: string) => (tools.run_query as any).execute({ query }, {} as any);
}

const failure = {
ok: false as const,
kind: "query" as const,
error: "Unknown expression identifier 'createdAt'.",
};
const transportFailure = {
ok: false as const,
kind: "transport" as const,
error: "The environment is temporarily unavailable.",
};
const success = { ok: true as const, rows: [{ n: 1 }] };

describe("run_query's consecutive-failure cap", () => {
it("keeps handing back the plain error until the cap", async () => {
const run = queryTool(async () => failure);

for (let attempt = 1; attempt < MAX_CONSECUTIVE_QUERY_FAILURES; attempt++) {
const result = await run("SELECT createdAt FROM runs");
expect(result.error).toBe(failure.error);
}
});

it("tells the model to stop and answer at the cap", async () => {
const run = queryTool(async () => failure);

let result: { error: string } = { error: "" };
for (let attempt = 0; attempt < MAX_CONSECUTIVE_QUERY_FAILURES; attempt++) {
result = await run("SELECT createdAt FROM runs");
}

expect(result.error).toContain(failure.error);
expect(result.error).toContain("answer the user with what you already have");
});

it("counts consecutive failures only, so a good query clears the count", async () => {
const postQuery = vi
.fn()
.mockResolvedValueOnce(failure)
.mockResolvedValueOnce(failure)
.mockResolvedValueOnce(success)
.mockResolvedValue(failure);
const run = queryTool(postQuery as any);

await run("bad");
await run("bad");
await run("good");
const result = await run("bad");

expect(result.error).toBe(failure.error);
});

it("does not count transport errors toward the cap", async () => {
const run = queryTool(async () => transportFailure);

let result: { error: string } = { error: "" };
for (let attempt = 0; attempt < MAX_CONSECUTIVE_QUERY_FAILURES + 2; attempt++) {
result = await run("SELECT createdAt FROM runs");
}

expect(result.error).toBe(transportFailure.error);
expect(result.error).not.toContain("answer the user with what you already have");
});
});
Loading