History

Make tool failures visible telemetry: span outcome attributes, derived error messages, e2e contract (#992)

* Tool failures become visible telemetry: outcome attributes, derived error messages, recording-tracer contracts Expected tool failures (ToolResult.fail) ride the Effect success channel, so the tracer recorded healthy spans while users hit upstream error walls — in a week of prod data, zero of 19.5k plugin.openapi.invoke spans carried http.status_code (annotated inside the inner OpenApi.invoke span, not the wrapper queries target) and 562 of 804 executor.tool.execute error spans had an empty status.message (tagged errors with no message field). - Stamp executor.tool.outcome / error_code / error_status on executor.tool.execute and mcp.tool.dispatch whenever a dispatch resolves to a ToolResult, plus tenant/subject attribution so queries no longer need a trace-id join against the outer request span. - Annotate http.status_code on the plugin.openapi.invoke wrapper span. - Derive messages for the message-less SDK tagged errors and stop formatInvocationCauseMessage rendering plain objects as [object Object]. - Add a recording Tracer to @executor-js/sdk/testing and telemetry contract tests that pin the span names, attributes, and error statuses production queries depend on. * Move product-observability plan note out of the repo * Pin the telemetry contract end-to-end instead of unit-level Replace the recording-tracer unit tests with one e2e scenario that drives the whole production topology — HTTP API → execution engine → sandbox → OpenAPI invoke → a real upstream returning 502 → OTLP export — and asserts the spans as the suite's motel store actually received them: outcome/error_code/error_status + tenant on executor.tool.execute, http.status_code on plugin.openapi.invoke, and ok/fail distinguishable. The unit tests verified span objects in process; the prod bug class this guards against (attribute stamped on a span the exporter never carries, failure riding the success channel) only shows up at the exported layer. Adds a Telemetry e2e service (motel span search with arrival polling), provided when the suite's motel boots (E2E_MOTEL_URL). * Document the telemetry surfaces for future work - prod-telemetry repo skill: the executor-cloud Axiom dataset layout (attributes.custom JSON map, span names and their attributes including the new outcome fields), working APL recipes for error digests and org attribution, prod-DB and PostHog access notes, and the deploy-canary procedure. - e2e/AGENTS.md: the Telemetry service — scenarios can assert on the spans the target actually exported, with the arrival-polling and fixture-tagging gotchas.

Rhys Sullivan committed 5 months ago

Commit 130d0e91 parent 42c0da414 files · +540 −17Browse files

Unified · Split

14 files
  1. .claude/skills/prod-telemetry/SKILL.md +114 −0
  2. e2e/AGENTS.md +25 −0
  3. e2e/cloud/telemetry-contract.test.ts +181 −0
  4. e2e/setup/cloud.globalsetup.ts +4 −0
  5. e2e/src/scenario.ts +7 −1
  6. e2e/src/services.ts +5 −0
  7. e2e/src/surfaces/telemetry.ts +75 −0
  8. packages/core/execution/src/tool-invoker.ts +5 −0
  9. packages/core/sdk/src/elicitation.ts +7 −1
  10. packages/core/sdk/src/errors.ts +52 −9
  11. packages/core/sdk/src/executor.ts +25 −4
  12. packages/core/sdk/src/index.ts +7 −1
  13. packages/core/sdk/src/tool-result.ts +28 −1
  14. packages/plugins/openapi/src/sdk/invoke.ts +5 −0
.claude/skills/prod-telemetry/SKILL.md+114−0
@@ -0,0 +1,114 @@
1+---
2+name: prod-telemetry
3+description: Query Executor's production telemetry — Axiom traces (executor-cloud dataset), prod Postgres via PlanetScale, PostHog product analytics — through the Executor MCP. Use when investigating prod errors, latency, usage, churn signals, or verifying a deploy's telemetry; includes the dataset field layout, working APL recipes, and the error-attribution join.
4+---
5+
6+# Production telemetry access
7+
8+All three stores are queryable through the Executor MCP's connected
9+integrations — no dashboards or credentials needed. Verify the connection
10+exists with `connections.list` if a call fails.
11+
12+## Axiom traces (`axiom_mcp`)
13+
14+Tool: `axiom_mcp.user.axiomMcpOAuth.querydataset` — the argument is `apl`
15+(NOT `query`). Dataset: `['executor-cloud']` (worker spans; browser spans
16+join the same traces via traceparent).
17+
18+**Field layout (the part you'd otherwise rediscover by failed queries):**
19+
20+- Custom span attributes live under the JSON map `['attributes.custom']`,
21+ NOT as top-level `attributes.*` columns. Read with
22+ `['attributes.custom']['mcp.tool.name']`. A nonexistent top-level field is
23+ a hard query error ("invalid field"), not an empty result.
24+- Span status: `['status.code']` (`"OK"`/`"ERROR"`), `['status.message']`.
25+- Exceptions: the `events` column carries `exception.type` /
26+ `exception.stacktrace` JSON.
27+- OTel basics are top-level: `name`, `trace_id`, `span_id`,
28+ `parent_span_id`, `duration`, `_time`.
29+
30+**Span names worth querying** (and their custom attrs):
31+
32+- `executor.tool.execute` — `mcp.tool.name` (full address), and since
33+ PR #992: `executor.tool.outcome` (`ok`/`fail`),
34+ `executor.tool.error_code`, `executor.tool.error_status`,
35+ `executor.tenant`, `executor.subject`.
36+- `mcp.tool.dispatch` — `mcp.tool.name` (sandbox path),
37+ `mcp.tool.integration`, same outcome attrs.
38+- `plugin.openapi.invoke` — `plugin.openapi.method` / `path_template` /
39+ `base_url`, and since PR #992 `http.status_code`.
40+- `mcp.request` (outer) — `mcp.auth.organization_id`,
41+ `mcp.auth.account_id`, `mcp.tool.name`, CF edge fields (`cf.country`…),
42+ MCP client fingerprint (`mcp.client.name`…).
43+
44+**Recipe — error signatures by class (the daily-digest query):**
45+
46+```apl
47+['executor-cloud']
48+| where _time > ago(1d)
49+| where ['status.code'] == "ERROR" and name == "executor.tool.execute"
50+| extend msg = substring(tostring(['status.message']), 0, 120)
51+| extend tool = tostring(['attributes.custom']['mcp.tool.name'])
52+| summarize n = count() by msg, tool
53+| sort by n desc
54+```
55+
56+**Recipe — attribute errors to orgs.** Tool spans now carry
57+`executor.tenant` directly (post-#992). For spans from BEFORE that deploy,
58+join through the outer request span:
59+
60+```apl
61+['executor-cloud']
62+| where name == "mcp.request" and isnotnull(['attributes.custom']['mcp.auth.organization_id'])
63+| project trace_id, org = tostring(['attributes.custom']['mcp.auth.organization_id'])
64+| join kind=inner (
65+ ['executor-cloud']
66+ | where ['status.code'] == "ERROR" and name == "executor.tool.execute"
67+ | project trace_id, msg = substring(tostring(['status.message']), 0, 60)
68+ ) on trace_id
69+| summarize n = count() by org, msg | sort by n desc
70+```
71+
72+**Recipe — upstream failure rate per integration (post-#992 attrs):**
73+
74+```apl
75+['executor-cloud']
76+| where _time > ago(1d) and name == "mcp.tool.dispatch"
77+| extend outcome = tostring(['attributes.custom']['executor.tool.outcome'])
78+| extend integration = tostring(['attributes.custom']['mcp.tool.integration'])
79+| where isnotnull(outcome)
80+| summarize calls = count(), fails = countif(outcome == "fail") by integration
81+| extend failRate = todouble(fails) / todouble(calls)
82+| sort by fails desc
83+```
84+
85+**Known signal caveats** (audited 2026-06-12):
86+
87+- Pre-#992 spans: `ToolResult.fail` outcomes (upstream 4xx/5xx, auth
88+ rejections) are INVISIBLE — they rode the Effect success channel with no
89+ span marker. Don't conclude "no errors" from old data.
90+- Many pre-#992 ERROR spans have an EMPTY `status.message` (tagged errors
91+ without a message field) — group those by `events` exception.type instead.
92+- `[object Object]` status messages are the pre-#992 formatting bug.
93+
94+## Prod database (`planetscale_mcp`)
95+
96+Read tool needs `{organization: "answer-overflow", database: "executor",
97+branch: "main"}`. It returns `ok: true` even when the SQL failed — check the
98+result text for `Error:`. Use for tenant/integration/connection facts that
99+spans don't carry (row sizes, config shapes, counts).
100+
101+## Product analytics (`posthog_api` / `mcp_posthog_com`)
102+
103+Browser-side events only (the ~60-event typed catalog, PR #987; server-side
104+events not built). The org-key `posthog_api` connection covers the REST API;
105+the OAuth MCP connection covers the higher-level tools.
106+
107+## Verifying a deploy's telemetry (Layer-0 canary)
108+
109+After deploying telemetry changes: run a known-failing tool call against
110+prod, then assert the expected attributes arrive in Axiom within ~1 min.
111+Absence of data looks identical to health — query for the NEW attribute
112+explicitly rather than eyeballing dashboards. The e2e equivalent runs on
113+every suite: `e2e/cloud/telemetry-contract.test.ts` via the `Telemetry`
114+service (motel `/api/spans/search?attr.<key>=<value>`).
e2e/AGENTS.md+25−0
@@ -77,6 +77,31 @@ const r = yield * session.call("execute", { code: "return 1 + 1;" }); · 76 unmodified lines
7777 // human-in-the-loop: session.approvePaused(r.text) resumes a paused execution
7878 ```
7979
80+## Telemetry scenarios (cloud)
81+
82+The suite boots a motel OTLP store and points the target's real exporter at
83+it, so a scenario can assert on the spans the server ACTUALLY exported —
84+the layer where "observability silently went dark" bugs live (an attribute
85+stamped on a span the exporter never carries looks identical to health).
86+
87+```ts
88+const telemetry = yield * Telemetry; // skips when motel didn't boot
89+const span =
90+ yield *
91+ telemetry.expectSpan({
92+ operation: "executor.tool.execute",
93+ attributes: { "mcp.tool.name": failAddress }, // exact match, values stringified
94+ });
95+expect(span.span.tags["executor.tool.outcome"]).toBe("fail");
96+```
97+
98+- `expectSpan` polls (~20s): exporters batch, so arrival is
99+ eventually-consistent — "the span reaches the store, soon" IS the contract.
100+- Spec gotcha for fixtures: give operations explicit `tags` — tool addresses
101+ are `group.leaf`, and an untagged op derives its group from the URL path,
102+ so `/fail` does NOT produce a `.fail`-suffixed address.
103+- Prior art: `cloud/telemetry-contract.test.ts`.
104+
80105 ## Running
81106
82107 ```sh
e2e/cloud/telemetry-contract.test.ts+181−0
@@ -0,0 +1,181 @@
1+// Cloud: the telemetry contract, end to end. A tool call that hits an
2+// upstream error wall must be visible in the EXPORTED spans — not just
3+// handled gracefully for the caller. This is the regression class where the
4+// product silently goes dark to operators: `ToolResult.fail` rides the
5+// Effect success channel (a healthy-looking span), and an attribute stamped
6+// on the wrong span simply never arrives in the trace store, which looks
7+// identical to health. So the assertion runs against the OTLP store the dev
8+// stack actually exported to (the suite's motel — the same exporter layer
9+// that ships prod spans to Axiom), driving the whole production topology:
10+// HTTP API → execution engine → sandbox → OpenAPI invoke → a real upstream
11+// returning 502 → span batch → OTLP export.
12+//
13+// Pins two regressions found live in prod (2026-06-12): http.status_code was
14+// annotated inside the inner `OpenApi.invoke` span so the `plugin.openapi.
15+// invoke` span queries target carried it on 0 of ~19.5k spans; and failed
16+// tool calls were indistinguishable from successes on `executor.tool.execute`.
17+import { randomBytes } from "node:crypto";
18+import { createServer } from "node:http";
19+
20+import { expect } from "@effect/vitest";
21+import { Effect } from "effect";
22+import { composePluginApi } from "@executor-js/api/server";
23+import { openApiHttpPlugin } from "@executor-js/plugin-openapi/api";
24+import { AuthTemplateSlug, ConnectionName, IntegrationSlug } from "@executor-js/sdk/shared";
25+
26+import { scenario } from "../src/scenario";
27+import { Api, Target, Telemetry } from "../src/services";
28+
29+const api = composePluginApi([openApiHttpPlugin()] as const);
30+
31+/** Two operations: /ok answers 200, /fail answers 502 — the success and
32+ * expected-upstream-failure outcome classes the telemetry must separate. */
33+const upstreamSpec = (baseUrl: string): string =>
34+ JSON.stringify({
35+ openapi: "3.0.3",
36+ info: { title: "Telemetry Upstream", version: "1.0.0" },
37+ servers: [{ url: baseUrl }],
38+ paths: {
39+ "/ok": {
40+ get: {
41+ operationId: "ok",
42+ summary: "Succeeds",
43+ tags: ["probe"],
44+ responses: { "200": { description: "" } },
45+ },
46+ },
47+ "/fail": {
48+ get: {
49+ operationId: "fail",
50+ summary: "Always 502",
51+ tags: ["probe"],
52+ responses: { "200": { description: "" } },
53+ },
54+ },
55+ },
56+ });
57+
58+/** A real upstream on 127.0.0.1: /ok → 200 JSON, anything else → 502 JSON. */
59+const serveUpstream = Effect.acquireRelease(
60+ Effect.callback<{ readonly baseUrl: string; readonly close: () => void }>((resume) => {
61+ const server = createServer((request, response) => {
62+ const ok = request.url?.startsWith("/ok") ?? false;
63+ response.writeHead(ok ? 200 : 502, { "content-type": "application/json" });
64+ response.end(ok ? '{"fine":true}' : '{"error":{"message":"bad gateway"}}');
65+ });
66+ server.listen(0, "127.0.0.1", () => {
67+ const address = server.address();
68+ const port = typeof address === "object" && address ? address.port : 0;
69+ resume(
70+ Effect.succeed({
71+ baseUrl: `http://127.0.0.1:${port}`,
72+ close: () => {
73+ server.close();
74+ server.closeAllConnections();
75+ },
76+ }),
77+ );
78+ });
79+ }),
80+ (server) => Effect.sync(server.close),
81+);
82+
83+scenario(
84+ "Telemetry · a failing tool call is visible in the exported spans",
85+ { timeout: 180_000 },
86+ Effect.scoped(
87+ Effect.gen(function* () {
88+ const target = yield* Target;
89+ const { client: apiClient } = yield* Api;
90+ const telemetry = yield* Telemetry;
91+ const identity = yield* target.newIdentity();
92+ const client = yield* apiClient(api, identity);
93+
94+ const upstream = yield* serveUpstream;
95+
96+ // Identifier-safe slug: it becomes a property path in the sandbox code.
97+ const slug = IntegrationSlug.make(`telscn${randomBytes(4).toString("hex")}`);
98+ yield* client.openapi.addSpec({
99+ payload: {
100+ spec: { kind: "blob", value: upstreamSpec(upstream.baseUrl) },
101+ slug,
102+ baseUrl: upstream.baseUrl,
103+ authenticationTemplate: [
104+ {
105+ slug: "apiKey",
106+ type: "apiKey",
107+ headers: { Authorization: ["Bearer ", { type: "variable", name: "token" }] },
108+ },
109+ ],
110+ },
111+ });
112+ yield* client.connections.create({
113+ payload: {
114+ owner: "org",
115+ name: ConnectionName.make("main"),
116+ integration: slug,
117+ template: AuthTemplateSlug.make("apiKey"),
118+ value: "telemetry-scenario-token",
119+ },
120+ });
121+
122+ const tools = yield* client.tools.list({ query: {} });
123+ const addressOf = (op: string) => {
124+ const tool = tools.find(
125+ (entry) =>
126+ String(entry.integration) === String(slug) && String(entry.address).endsWith(`.${op}`),
127+ );
128+ expect(tool, `the ${op} tool is in the catalog`).toBeDefined();
129+ return String(tool!.address);
130+ };
131+ const failAddress = addressOf("fail");
132+ const okAddress = addressOf("ok");
133+
134+ // Drive both outcome classes through the full production path. The
135+ // failing call still completes for the caller — that is exactly why
136+ // the exported span is the only place an operator can see it.
137+ for (const address of [okAddress, failAddress]) {
138+ const execution = yield* client.executions.execute({
139+ payload: { code: `return await ${address}({});` },
140+ });
141+ expect(execution.status, `the ${address} execution completes`).toBe("completed");
142+ }
143+
144+ // The failure: outcome attributes on the tool span...
145+ const failSpan = yield* telemetry.expectSpan({
146+ operation: "executor.tool.execute",
147+ attributes: { "mcp.tool.name": failAddress },
148+ });
149+ expect(failSpan.span.tags, "a failed tool call is marked on the exported span").toMatchObject(
150+ {
151+ "executor.tool.outcome": "fail",
152+ "executor.tool.error_code": "upstream_http_error",
153+ "executor.tool.error_status": "502",
154+ },
155+ );
156+ expect(
157+ failSpan.span.tags["executor.tenant"],
158+ "the span carries tenant attribution (no trace-join needed to ask 'whose error?')",
159+ ).toBeTruthy();
160+
161+ // ...and the upstream status on the HTTP span queries actually target.
162+ const invokeSpan = yield* telemetry.expectSpan({
163+ operation: "plugin.openapi.invoke",
164+ attributes: { "plugin.openapi.base_url": upstream.baseUrl, "http.status_code": "502" },
165+ });
166+ expect(
167+ invokeSpan.span.tags["plugin.openapi.method"],
168+ "the invoke span names the method",
169+ ).toBe("GET");
170+
171+ // The success is distinguishable from the failure.
172+ const okSpan = yield* telemetry.expectSpan({
173+ operation: "executor.tool.execute",
174+ attributes: { "mcp.tool.name": okAddress },
175+ });
176+ expect(okSpan.span.tags["executor.tool.outcome"], "a successful call is marked ok").toBe(
177+ "ok",
178+ );
179+ }),
180+ ),
181+);
e2e/setup/cloud.globalsetup.ts+4−0
@@ -28,6 +28,10 @@ export default async function setup(): Promise<(() => Promise<void>) | void> { · 27 unmodified lines
2828
2929 // Suite-owned trace store — every run captures distributed traces.
3030 const motel = await bootMotel();
31+ // Publish to the test workers (they inherit this process's env): scenarios
32+ // that assert on exported spans yield the Telemetry service, which exists
33+ // only when this is set. No motel → those scenarios skip, never fail.
34+ if (motel) process.env.E2E_MOTEL_URL = motel.url;
3135
3236 const publicUrl = `http://127.0.0.1:${ports.E2E_CLOUD_PORT!}`;
3337 let booted;
e2e/src/scenario.ts+7−1
@@ -24,6 +24,7 @@ import { makeApiSurface } from "./surfaces/api"; · 23 unmodified lines
2424 import { makeBrowserSurface } from "./surfaces/browser";
2525 import { makeCliSurface } from "./surfaces/cli";
2626 import { makeMcpSurface } from "./surfaces/mcp";
27+import { makeTelemetrySurface } from "./surfaces/telemetry";
2728 import { completeOAuthConsent, hasOpenCode, makeOpenCodeHome, warmUp } from "./clients/opencode";
2829 import {
2930 Api,
@@ -35,6 +36,7 @@ import { · 5 unmodified lines
3536 Restart,
3637 RunDir,
3738 Target,
39+ Telemetry,
3840 TtlControl,
3941 } from "./services";
4042 import { buildManifest } from "./viewer/manifest";
@@ -62,7 +64,8 @@ type AllServices = · 21 unmodified lines
6264 | Billing
6365 | OpenCode
6466 | TtlControl
65− | Restart;
67+ | Restart
68+ | Telemetry;
6669
6770 /**
6871 * What this target on this host can provide. Services beyond the base are
@@ -94,6 +97,9 @@ const contextFor = (target: TargetShape, dir: string): Context.Context<AllServic · 25 unmodified lines
9497 if (target.restart) {
9598 context = Context.add(context, Restart, target.restart);
9699 }
100+ if (process.env.E2E_MOTEL_URL) {
101+ context = Context.add(context, Telemetry, makeTelemetrySurface(process.env.E2E_MOTEL_URL));
102+ }
97103 return context;
98104 };
99105
e2e/src/services.ts+5−0
@@ -11,6 +11,7 @@ import type { ApiSurface } from "./surfaces/api"; · 10 unmodified lines
1111 import type { BrowserSurface } from "./surfaces/browser";
1212 import type { CliSurface } from "./surfaces/cli";
1313 import type { McpSurface } from "./surfaces/mcp";
14+import type { TelemetrySurface } from "./surfaces/telemetry";
1415 import type { completeOAuthConsent, makeOpenCodeHome, warmUp } from "./clients/opencode";
1516
1617 /** The target under test (always provided). */
@@ -34,6 +35,10 @@ export class Mcp extends Context.Service<Mcp, McpSurface>()("e2e/mcp-oauth") {} · 17 unmodified lines
3435 /** Marker: billing limits are enforced on this target. */
3536 export class Billing extends Context.Service<Billing, true>()("e2e/billing") {}
3637
38+/** Query the suite's OTLP trace store for spans the target actually exported
39+ * (present when the suite booted motel — E2E_MOTEL_URL). */
40+export class Telemetry extends Context.Service<Telemetry, TelemetrySurface>()("e2e/telemetry") {}
41+
3742 /** The real OpenCode binary, hermetically driveable (present when installed on this host). */
3843 export interface OpenCodeClient {
3944 readonly makeHome: typeof makeOpenCodeHome;
e2e/src/surfaces/telemetry.ts+75−0
@@ -0,0 +1,75 @@
1+// Telemetry surface: query the suite's motel OTLP store for the spans the
2+// target ACTUALLY exported. This is how a scenario asserts the observability
3+// contract end-to-end — not "did the code create a span object" but "did the
4+// span leave the server, with the attributes production queries depend on".
5+// The dangerous regression mode here is silence (an attribute stamped on the
6+// wrong span, an error riding the success channel): absent data looks exactly
7+// like health, so the contract has to be pinned where the data is read.
8+import { Effect, Schedule } from "effect";
9+
10+/** One exported span, as motel's /api/spans/search returns it. `tags` is the
11+ * span's attributes with every value stringified. */
12+export interface ExportedSpan {
13+ readonly traceId: string;
14+ readonly rootOperationName: string;
15+ readonly span: {
16+ readonly spanId: string;
17+ readonly operationName: string;
18+ readonly status: "ok" | "error";
19+ readonly durationMs: number;
20+ readonly tags: Readonly<Record<string, string>>;
21+ };
22+}
23+
24+export interface SpanQuery {
25+ /** Exact-match attribute filters (motel `attr.<key>=<value>`). */
26+ readonly attributes?: Readonly<Record<string, string>>;
27+ readonly operation?: string;
28+ readonly traceId?: string;
29+}
30+
31+export interface TelemetrySurface {
32+ /** One-shot search against the trace store. */
33+ readonly searchSpans: (query: SpanQuery) => Effect.Effect<readonly ExportedSpan[], unknown>;
34+ /** Search until at least one span matches. Exporters batch (the app
35+ * flushes ~1s after the request), so arrival is eventually-consistent —
36+ * polling IS the contract: "the span reaches the store, soon". */
37+ readonly expectSpan: (query: SpanQuery) => Effect.Effect<ExportedSpan, unknown>;
38+}
39+
40+export const makeTelemetrySurface = (motelUrl: string): TelemetrySurface => {
41+ const searchSpans = (query: SpanQuery) =>
42+ Effect.gen(function* () {
43+ const params = new URLSearchParams({ lookback: "15m", limit: "100" });
44+ if (query.operation) params.set("operation", query.operation);
45+ if (query.traceId) params.set("traceId", query.traceId);
46+ for (const [key, value] of Object.entries(query.attributes ?? {})) {
47+ params.set(`attr.${key}`, value);
48+ }
49+ const response = yield* Effect.promise(() => fetch(`${motelUrl}/api/spans/search?${params}`));
50+ if (!response.ok) {
51+ return yield* Effect.fail(
52+ `motel span search responded ${response.status}: ${yield* Effect.promise(() => response.text())}`,
53+ );
54+ }
55+ const body = (yield* Effect.promise(() => response.json())) as {
56+ readonly data?: readonly ExportedSpan[];
57+ };
58+ return body.data ?? [];
59+ });
60+
61+ return {
62+ searchSpans,
63+ expectSpan: (query) =>
64+ searchSpans(query).pipe(
65+ Effect.filterOrFail(
66+ (spans) => spans.length > 0,
67+ () => `no exported span matched ${JSON.stringify(query)}`,
68+ ),
69+ Effect.map((spans) => spans[0]!),
70+ // ~20s ceiling (40 × 500ms): BatchSpanProcessor flushes ~1s after the
71+ // request and the dev stack drains on waitUntil; slower is a real bug.
72+ Effect.retry(Schedule.both(Schedule.spaced("500 millis"), Schedule.recurs(40))),
73+ ),
74+ };
75+};
packages/core/execution/src/tool-invoker.ts+5−0
@@ -9,6 +9,7 @@ import type { · 8 unmodified lines
99 ToolSchemaView,
1010 } from "@executor-js/sdk/core";
1111 import {
12+ annotateToolResultOutcome,
1213 authToolFailure,
1314 isToolResult,
1415 ToolResult,
@@ -282,6 +283,10 @@ export const makeExecutorToolInvoker = ( · 267 unmodified lines
282283 // raw success value and wrapped — keeps the sandbox-facing contract
283284 // uniform without forcing every tiny test plugin to import
284285 // `ToolResult.ok`.
286+ // Expected failures resolve through the success channel, so without the
287+ // outcome annotation the dispatch span reads as healthy even when the
288+ // caller hit an upstream error or auth wall.
289+ yield* annotateToolResultOutcome(result);
285290 if (isToolResult(result)) {
286291 return result;
287292 }
packages/core/sdk/src/elicitation.ts+7−1
@@ -62,4 +62,10 @@ export class ElicitationDeclinedError extends Schema.TaggedErrorClass<Elicitatio · 61 unmodified lines
6262 address: ToolAddress,
6363 action: Schema.Literals(["decline", "cancel"]),
6464 },
65−) {}
65+) {
66+ // Derived message so telemetry (span status, logs) labels the failure
67+ // instead of rendering an Error with an empty message.
68+ override get message(): string {
69+ return `Tool approval ${this.action === "cancel" ? "cancelled" : "declined"}: ${this.address}`;
70+ }
71+}
packages/core/sdk/src/errors.ts+52−9
@@ -13,13 +13,24 @@ import { ConnectionName, IntegrationSlug, Owner, ProviderKey, ToolAddress } from · 12 unmodified lines
1313 // Tool lifecycle
1414 // ---------------------------------------------------------------------------
1515
16+/* Tagged errors without an explicit `message` field define a `message` getter:
17+ * `Schema.TaggedErrorClass` instances are real Errors with `message: ""`, and
18+ * an empty message propagates everywhere errors are rendered — span
19+ * status.message in the tracer, Cause.pretty output, log lines — leaving the
20+ * failure unlabeled in telemetry. The getter is derived from the schema fields
21+ * (not an own property), so encoding/serialization is unaffected. */
22+
1623 export class ToolNotFoundError extends Schema.TaggedErrorClass<ToolNotFoundError>()(
1724 "ToolNotFoundError",
1825 {
1926 address: ToolAddress,
2027 suggestions: Schema.optional(Schema.Array(ToolAddress)),
2128 },
22−) {}
29+) {
30+ override get message(): string {
31+ return `Tool not found: ${this.address}`;
32+ }
33+}
2334
2435 export class ToolInvocationError extends Schema.TaggedErrorClass<ToolInvocationError>()(
2536 "ToolInvocationError",
@@ -38,7 +49,11 @@ export class ToolBlockedError extends Schema.TaggedErrorClass<ToolBlockedError>( · 12 unmodified lines
3849 address: ToolAddress,
3950 pattern: Schema.String,
4051 },
41−) {}
52+) {
53+ override get message(): string {
54+ return `Tool blocked by policy "${this.pattern}": ${this.address}`;
55+ }
56+}
4257
4358 /** Tool row exists but its owning plugin isn't loaded in this executor config. */
4459 export class PluginNotLoadedError extends Schema.TaggedErrorClass<PluginNotLoadedError>()(
@@ -47,13 +62,21 @@ export class PluginNotLoadedError extends Schema.TaggedErrorClass<PluginNotLoade · 2 unmodified lines
4762 address: ToolAddress,
4863 pluginId: Schema.String,
4964 },
50−) {}
65+) {
66+ override get message(): string {
67+ return `Plugin "${this.pluginId}" is not loaded for tool: ${this.address}`;
68+ }
69+}
5170
5271 /** Tool was found but its owning plugin has no `invokeTool` handler. */
5372 export class NoHandlerError extends Schema.TaggedErrorClass<NoHandlerError>()("NoHandlerError", {
5473 address: ToolAddress,
5574 pluginId: Schema.String,
56−}) {}
75+}) {
76+ override get message(): string {
77+ return `Plugin "${this.pluginId}" has no invokeTool handler for tool: ${this.address}`;
78+ }
79+}
5780
5881 // ---------------------------------------------------------------------------
5982 // Integration / connection lifecycle
@@ -62,7 +85,11 @@ export class NoHandlerError extends Schema.TaggedErrorClass<NoHandlerError>()("N · 2 unmodified lines
6285 export class IntegrationNotFoundError extends Schema.TaggedErrorClass<IntegrationNotFoundError>()(
6386 "IntegrationNotFoundError",
6487 { slug: IntegrationSlug },
65−) {}
88+) {
89+ override get message(): string {
90+ return `Integration not found: ${this.slug}`;
91+ }
92+}
6693
6794 /** An "add integration" operation targeted a slug (namespace) that is already
6895 * registered. The core `integrations.register` primitive upserts by design
@@ -73,14 +100,22 @@ export class IntegrationAlreadyExistsError extends Schema.TaggedErrorClass<Integ · 4 unmodified lines
73100 "IntegrationAlreadyExistsError",
74101 { slug: IntegrationSlug },
75102 { httpApiStatus: 409 },
76−) {}
103+) {
104+ override get message(): string {
105+ return `Integration already exists: ${this.slug}`;
106+ }
107+}
77108
78109 /** `integrations.remove` was called on an integration declared statically by a
79110 * plugin at startup (`canRemove: false`). */
80111 export class IntegrationRemovalNotAllowedError extends Schema.TaggedErrorClass<IntegrationRemovalNotAllowedError>()(
81112 "IntegrationRemovalNotAllowedError",
82113 { slug: IntegrationSlug },
83−) {}
114+) {
115+ override get message(): string {
116+ return `Integration cannot be removed (declared statically by a plugin): ${this.slug}`;
117+ }
118+}
84119
85120 export class ConnectionNotFoundError extends Schema.TaggedErrorClass<ConnectionNotFoundError>()(
86121 "ConnectionNotFoundError",
@@ -89,7 +124,11 @@ export class ConnectionNotFoundError extends Schema.TaggedErrorClass<ConnectionN · 2 unmodified lines
89124 integration: IntegrationSlug,
90125 name: ConnectionName,
91126 },
92−) {}
127+) {
128+ override get message(): string {
129+ return `Connection not found: ${this.integration}.${this.owner}.${this.name}`;
130+ }
131+}
93132
94133 /** A connection create request was rejected before anything was written: the
95134 * input is structurally invalid (no credential inputs for a credentialed
@@ -106,7 +145,11 @@ export class InvalidConnectionInputError extends Schema.TaggedErrorClass<Invalid · 10 unmodified lines
106145 export class CredentialProviderNotRegisteredError extends Schema.TaggedErrorClass<CredentialProviderNotRegisteredError>()(
107146 "CredentialProviderNotRegisteredError",
108147 { provider: ProviderKey },
109−) {}
148+) {
149+ override get message(): string {
150+ return `Credential provider not registered: ${this.provider}`;
151+ }
152+}
110153
111154 /** A connection's value could not be resolved — the provider returned nothing,
112155 * or an OAuth token refresh failed and the user must re-auth. */
packages/core/sdk/src/executor.ts+25−4
@@ -1,4 +1,4 @@
1−import { Effect, Layer, Option, Predicate, Schema } from "effect";
1+import { Effect, Inspectable, Layer, Option, Predicate, Schema } from "effect";
22 import { FetchHttpClient, type HttpClient } from "effect/unstable/http";
33 import { fumadb } from "@executor-js/fumadb";
44 import { memoryAdapter } from "@executor-js/fumadb/adapters/memory";
@@ -139,6 +139,7 @@ import { · 134 unmodified lines
139139 type OAuthEndpointUrlPolicy,
140140 } from "./oauth-helpers";
141141 import { connectionIdentifier } from "./connection-name-identifier";
142+import { annotateToolResultOutcome } from "./tool-result";
142143
143144 const PLUGIN_STORAGE_DELETE_KEY_BATCH_SIZE = 90;
144145 const MAX_APPROVAL_ARGUMENT_PREVIEW_CHARS = 4_000;
@@ -2719,10 +2720,20 @@ export const createExecutor = <const TPlugins extends readonly AnyPlugin[] = rea · 2574 unmodified lines
27192720 ): Effect.Effect<unknown, ExecuteError> => {
27202721 const handler = pickHandler(options);
27212722 return Effect.gen(function* () {
2723+ // oxlint-disable executor/no-instanceof-error, executor/no-unknown-error-message, executor/no-manual-tag-check -- boundary: normalize arbitrary unknown plugin failures into a human-readable message for ToolInvocationError/telemetry
27222724 const formatInvocationCauseMessage = (cause: unknown): string => {
2723− // oxlint-disable-next-line executor/no-instanceof-error, executor/no-unknown-error-message -- boundary: preserve public execute error message wrapping for unknown plugin failures
2724− return cause instanceof Error ? cause.message : String(cause);
2725+ if (cause instanceof Error && cause.message.length > 0) return cause.message;
2726+ // Non-Error / empty-message causes: `String(plainObject)` renders
2727+ // "[object Object]", which is what telemetry then shows as the only
2728+ // label for the failure. Prefer the tag, else stringify structurally.
2729+ if (typeof cause === "object" && cause !== null) {
2730+ const tag = (cause as { readonly _tag?: unknown })._tag;
2731+ if (typeof tag === "string") return tag;
2732+ return Inspectable.toStringUnknown(cause, 0);
2733+ }
2734+ return String(cause);
27252735 };
2736+ // oxlint-enable executor/no-instanceof-error, executor/no-unknown-error-message, executor/no-manual-tag-check
27262737 const wrapInvocationError = <A, E>(
27272738 effect: Effect.Effect<A, E>,
27282739 ): Effect.Effect<A, ToolInvocationError> =>
@@ -2877,8 +2888,18 @@ export const createExecutor = <const TPlugins extends readonly AnyPlugin[] = rea · 148 unmodified lines
28772888 }),
28782889 );
28792890 }).pipe(
2891+ // Expected tool failures (`ToolResult.fail`) resolve through the
2892+ // success channel, so the tracer alone would record them as healthy
2893+ // spans. Stamp the outcome + error code so telemetry can distinguish
2894+ // "tool ran fine" from "user hit an upstream error / auth wall"
2895+ // without parsing response bodies.
2896+ Effect.tap(annotateToolResultOutcome),
28802897 Effect.withSpan("executor.tool.execute", {
2881− attributes: { "mcp.tool.name": String(address) },
2898+ attributes: {
2899+ "mcp.tool.name": String(address),
2900+ "executor.tenant": tenant,
2901+ ...(subject != null ? { "executor.subject": subject } : {}),
2902+ },
28822903 }),
28832904 );
28842905 };
packages/core/sdk/src/index.ts+7−1
@@ -324,7 +324,13 @@ export { buildToolTypeScriptPreview } from "./schema-types"; · 323 unmodified lines
324324 export { InternalError } from "./api-errors";
325325
326326 // ToolResult — typed value-based discriminated union for tool outcomes.
327−export { ToolResult, isToolResult, type ToolError, type ToolHttpMeta } from "./tool-result";
327+export {
328+ ToolResult,
329+ annotateToolResultOutcome,
330+ isToolResult,
331+ type ToolError,
332+ type ToolHttpMeta,
333+} from "./tool-result";
328334
329335 // Stamped boot-time data-migration ledger for the libSQL-backed apps.
330336 export {
packages/core/sdk/src/tool-result.ts+28−1
@@ -5,7 +5,7 @@ · 4 unmodified lines
55 // the Effect failure channel.
66 // ---------------------------------------------------------------------------
77
8−import { Schema } from "effect";
8+import { Effect, Schema } from "effect";
99
1010 export const ToolErrorSchema = Schema.Struct({
1111 code: Schema.String,
@@ -56,3 +56,30 @@ const isUnknownToolResult = Schema.is(ToolResultSchema); · 44 unmodified lines
5656
5757 export const isToolResult = (value: unknown): value is ToolResult<unknown> =>
5858 isUnknownToolResult(value);
59+
60+/**
61+ * Annotate the current span with the outcome of a tool invocation.
62+ *
63+ * `ToolResult.fail` rides the Effect *success* channel by design (expected
64+ * failures are values, not defects), which means the tracer records those
65+ * spans as healthy. Without this, "user keeps hitting 4xx walls" is invisible
66+ * to telemetry — the exact class of signal that lets us catch product issues
67+ * before they're reported. Stamped attributes:
68+ *
69+ * - `executor.tool.outcome` — "ok" | "fail" (always, on ToolResults)
70+ * - `executor.tool.error_code` — ToolError.code (fail only)
71+ * - `executor.tool.error_status` — upstream HTTP status (fail, when present)
72+ *
73+ * Codes/statuses are enumerable identifiers, never user content — safe span
74+ * attributes. Non-ToolResult values (raw success payloads) annotate "ok".
75+ */
76+export const annotateToolResultOutcome = (value: unknown): Effect.Effect<void> => {
77+ if (isToolResult(value) && !value.ok) {
78+ return Effect.annotateCurrentSpan({
79+ "executor.tool.outcome": "fail",
80+ "executor.tool.error_code": value.error.code,
81+ ...(value.error.status != null ? { "executor.tool.error_status": value.error.status } : {}),
82+ });
83+ }
84+ return Effect.annotateCurrentSpan({ "executor.tool.outcome": "ok" });
85+};
packages/plugins/openapi/src/sdk/invoke.ts+5−0
@@ -625,6 +625,11 @@ export const invokeWithLayer = ( · 624 unmodified lines
625625
626626 return invoke(operation, args, resolvedHeaders, sourceQueryParams).pipe(
627627 Effect.provide(clientWithBaseUrl),
628+ // `invoke` annotates http.status_code on ITS span (`OpenApi.invoke`,
629+ // via Effect.fn) — annotateCurrentSpan inside it never reaches this
630+ // wrapper span. Stamp the status here too so queries against
631+ // `plugin.openapi.invoke` see the upstream outcome directly.
632+ Effect.tap((result) => Effect.annotateCurrentSpan({ "http.status_code": result.status })),
628633 Effect.withSpan("plugin.openapi.invoke", {
629634 attributes: {
630635 "plugin.openapi.method": operation.method.toUpperCase(),

Keyboard shortcuts

General

Show keyboard shortcuts
?
Focus the current search
/

Issues and pull requests

Open filters
f

Issues, pull requests and Inbox

Select next item
j↓
Select previous item
k↑
Open selected item
Enter
Dismiss filters or clear selection
Esc

Repository code

Go to file
t