Files
PaperClipAI/tests/runner-e2e/chat-catalog-domains/external-gateway-client.ts
T
f4337bb5fb test(evals): land the chat catalog eval harness on master (#19614)
## Thinking Path

> - Paperclip is the open source app people use to manage AI agents for
work
> - The chat-feature evals measure whether agents complete real chat
requests through the product
> - The chat catalog eval harness and its repairs live on the
`chat-eval/master-candidate` branch, not on master
> - Today's work made 139 of 242 chat evals pass in verified single-case
live runs. Each product fix landed on master as its own pull request
> - The eval-side changes (setups, graders, fault seams, evidence
capture) still exist only on the candidate
> - This pull request lands the harness on master. Master stays the
source of truth for product behavior
> - The benefit is one tree: evals run against master commits, and
product fixes reach the evals as soon as they merge

## Linked Issues or Issue Description

Refs #19441, Refs #19599, Refs #20270, Refs #20334, Refs #20342, Refs
#20344, Refs #20354, Refs #20360, Refs #20373, Refs #20382, Refs #20409

**What existing behavior does this improve?**
The chat catalog eval harness and the product it evaluates live on a
long-running branch outside master. Eval fixes do not reach master, so
master-based eval runs use old setups and graders.

**Subsystem affected**
`tests/runner-e2e` chat catalog harness, plus a small set of product
changes that exist only on the candidate.

**Proposed change**
Land the harness on master. Every product fix from this eval work
already merged to master as its own pull request. Conflicts keep
master's product behavior.

## What Changed

- Most of the diff (808 of 851 files) is under `tests/runner-e2e/`:
chat-catalog domains, scenarios, descriptors, the TestOps manifest,
private evidence tooling, and their tests.
- Graders now check saved state, setup literals and rendered or clicked
links. Many keyword and regex checks on free-form replies are removed.
Each changed grader has a bumped `oracleVersion` and calibration tests
(a valid reply, a reworded valid reply, and broken variants that must
fail).
- The shared link helper `hasLink` no longer treats a list link (for
example `/COM/issues?view=mine`) as a link to every task without a URL
key.
- Setup fixes build real state through supported product routes: routine
history, agent grants through the board grant route, budget holds with a
real cap, and a storage-layer archive fault for
`inbox-search.archive-partial-failure`.
- Browser boundary journeys (sign-in, chat-disabled, onboarding) now
write the standard pass evidence files.
- Retargeted cases, approved by the maintainer:
`agent-chat.chat-disabled` (Conference Room gate),
`agents.restore-instructions` (no saved instruction history),
`access.manage-membership` (member permission edits stay board-only).
- The private evidence trusted files (`manual-private-handoff.ts`,
`manual-private-retention.ts`, `manual-private-report-assets.ts`) are
byte-identical to master.
- Product files outside `tests/` (43 files, +1,328 −49) carry
candidate-only changes that master does not yet have. Most are tests.
The main code changes are in `request-human-input.ts`,
`heartbeat-run-summary.ts`, `native-completion-feedback.ts`,
`native-run-finalizer.ts`, `openapi.ts`, `secrets.ts` and
`IssueChatLongThreadPerf.tsx`.

## Verification

- `pnpm vitest run chat-catalog-domains chat-catalog-contract.test.ts
chat-catalog-consistency.test.ts unit-launcher.test.mjs
--testTimeout=180000` in `tests/runner-e2e`: 269 files pass, 3,550 tests
pass.
- `node --test chat-full-catalog-testops-manifest.test.mjs`: 10 pass.
- `node tests/runner-e2e/source-sdk-public-inputs.mjs <HEAD> --check`:
no change.
- `tsc` for `server` and `tests/runner-e2e`: clean.
- Live single-case verification: 139 of 242 cases passed on paid runs,
and a person read the saved evidence for each. These are single-case
runs, not one full-cohort run.

## Risks

- Large diff. Most of it is test-only, but the 43 non-test files change
product code paths. Review those files first.
- Some graders are now less strict on reply wording. A person must read
replies for meaning on those cases.
- 42 cases need disposable outside accounts (Notion, Slack, GitHub,
Daytona and others) and stay blocked. `access.invite-teammate` stays
failing until local agent isolation lands, because the runner's durable
state keeps a live invite token.
- The paid workflow compares the trusted private-evidence files with
master. They are unchanged, so this risk is low.

## Model Used

Anthropic Claude Opus 5.5 (`claude-opus-5-5`), with tool use and code
execution, in Claude Code. Subagents used the same model.

## Checklist

- [x] I have included a thinking path that traces from project context
to this change
- [x] I have specified the model used (with version and capability
details)
- [x] I have checked ROADMAP.md and confirmed this PR does not duplicate
planned core work
- [x] I have searched GitHub for duplicate or related PRs and linked
them above
- [x] I have either (a) linked existing issues with `Fixes: #` / `Closes
#` / `Refs #` OR (b) described the issue in-PR following the relevant
issue template
- [x] I have not referenced internal/instance-local Paperclip issues or
links (only public GitHub `#NNN` / `github.com/paperclipai/paperclip`
URLs)
- [x] My branch name describes the change (e.g. `docs/...`, `fix/...`)
and contains no internal Paperclip ticket id or instance-derived details
- [x] I have run tests locally and they pass
- [x] I have added or updated tests where applicable
- [x] I have updated relevant documentation to reflect my changes
- [x] I have considered and documented any risks above
- [ ] All Paperclip CI gates are green
- [ ] Greptile is 5/5 with no open P2s, recommendations, or follow-ups
- [x] I will address all Greptile and reviewer comments before
requesting merge

🤖 Generated with [Claude Code](https://claude.com/claude-code)

---------

Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
Co-authored-by: Paperclip <noreply@paperclip.ing>
2026-11-10 08:02:11 -06:00

435 lines
15 KiB
TypeScript

import { randomBytes } from "node:crypto";
import type { ChatCatalogContext } from "../chat-catalog-contract.js";
import type {
ToolProfileWithDetails,
ToolProfileEffectiveSummary,
ToolRunDecisionLookup,
} from "../../../packages/shared/src/types/tool-access.js";
import { pollUntil } from "../api.js";
import {
definition,
externalObservation,
observeExternal,
root,
checks,
verifyRenderedReply,
} from "./external-support.js";
import {
connectNotes,
startNotesProvider,
type NotesProvider,
} from "./external-provider.js";
interface PrivateFixture {
provider: NotesProvider;
key?: { id: string; token: string };
session?: { sessionId: string; token: string };
}
const fixtures = new WeakMap<ChatCatalogContext, PrivateFixture>();
async function privateRequest(
ctx: ChatCatalogContext,
pathname: string,
token: string,
body?: unknown,
session = false,
) {
const url = new URL(pathname, ctx.input.api.baseURL);
if (
url.origin !== ctx.input.api.baseURL ||
!url.pathname.startsWith("/api/tool-gateway/")
)
throw new Error("Scoped session probe escaped its fixture API");
return fetch(url, {
method: body === undefined ? "GET" : "POST",
redirect: "error",
signal: AbortSignal.timeout(15_000),
headers: {
...(session
? { "x-paperclip-tool-gateway-token": token }
: { Authorization: `Bearer ${token}` }),
"content-type": "application/json",
},
...(body === undefined ? {} : { body: JSON.stringify(body) }),
});
}
/** Failure diagnostics are closed control-plane facts, never arbitrary upstream
* error text. Unknown/secret-bearing fields are deliberately omitted. */
export function scopedSessionFailureDiagnostic(
status: number,
context: {
companyId: string;
agentId: string;
runId: string;
issueId: string;
observedRunStatusAtAdmission: string;
},
response: unknown,
) {
const allowed = [
"run_company_mismatch",
"run_agent_mismatch",
"run_inactive",
"run_context_mismatch",
];
const reason =
response && typeof response === "object"
? (response as { reasonCode?: unknown }).reasonCode
: undefined;
return {
httpStatus: status,
companyId: context.companyId,
agentId: context.agentId,
runId: context.runId,
issueId: context.issueId,
observedRunStatusAtAdmission: context.observedRunStatusAtAdmission,
reasonCode:
typeof reason === "string" && allowed.includes(reason)
? reason
: "unclassified",
responseBodyRetained: false,
};
}
export function gradeGatewayClient(value: unknown) {
const c = checks(),
o = externalObservation(value);
c.add(
"complete-scoped-client-evidence",
!!o,
"Actual assigned profile, run/task admission, independent discovery/denial and provider-read receipts are required.",
);
if (!o) return c.result();
const effective = o.state.effective as
ToolProfileEffectiveSummary | undefined;
c.add(
"intended-effective-profile",
effective?.agentId === o.leadId &&
effective.profiles.some((row) => row.id === o.state.profileId) &&
effective.allowedTools.some((row) => row.id === o.state.readEntryId) &&
!effective.allowedTools.some((row) => row.id === o.state.writeEntryId),
"The saved launch profile grants its read action to this agent while excluding the write action.",
);
c.add(
"actual-active-run-scoped-session",
o.state.sessionCreated === true &&
o.runs.some(
(run) =>
run.id === o.state.probeRunId &&
run.agentId === o.leadId &&
run.companyId === o.companyId,
) &&
o.state.discoveredRead === true &&
o.state.discoveredWrite === false,
"A real agent bearer creates a session for the admitted source run/task and discovers only authorized actions.",
);
const decisions = o.state.decisions as ToolRunDecisionLookup[] | undefined;
c.add(
"model-performed-bounded-read",
!!decisions &&
decisions.some((row) =>
row.decisions.some(
(decision) =>
decision.invocation.runId === row.runId &&
decision.invocation.agentId === o.leadId &&
decision.invocation.issueId === o.state.sourceIssueId &&
decision.invocation.connectionId === o.state.connectionId &&
decision.invocation.catalogEntryId === o.state.readEntryId &&
decision.invocation.status === "succeeded",
),
) &&
Number(o.state.providerReadCalls) >= 1 &&
Number(o.state.providerReadCalls) <= 3 &&
typeof o.state.marker === "string" &&
o.reply.includes(o.state.marker),
"The actual model run uses the real provider read and reports its unpredictable result; the oracle never performs that read.",
);
c.add(
"disallowed-write-denied",
o.state.writeProbeStatus === 403 &&
["deny_default", "deny_policy_block"].includes(
String(o.state.writeProbeReason),
) &&
o.state.providerWriteCalls === 0 &&
o.state.written === null,
"A real scoped-session write attempt is denied before provider execution.",
);
c.add(
"no-private-token-disclosure",
o.state.tokenScanComplete === true && o.state.privateTokenLeaked === false,
"Scoped session and probe credentials never enter the persisted reply/run result.",
);
return c.result();
}
export const gatewayClientExternalCases = [
definition(
"gateways-profiles.connect-client",
async (ctx) => {
const marker = `Launch fixture result ${randomBytes(16).toString("hex")}`;
const provider = await startNotesProvider(marker);
const fixture: PrivateFixture = { provider };
fixtures.set(ctx, fixture);
try {
const connection = await connectNotes(ctx, provider, "read-only");
const profile = await ctx.input.api.post<ToolProfileWithDetails>(
`${root(ctx)}/tools/profiles`,
{
profileKey: "launch",
name: "Launch profile",
defaultAction: "deny",
entries: [
{
selectorType: "catalog_entry",
effect: "include",
connectionId: connection.connectionId,
catalogEntryId: connection.read[0],
},
{
selectorType: "catalog_entry",
effect: "exclude",
connectionId: connection.connectionId,
catalogEntryId: connection.write[0],
},
],
},
);
await ctx.input.api.post(
`${root(ctx)}/tools/profiles/${profile.id}/bind`,
{ targetType: "agent", targetId: ctx.input.fixtures.agent.id },
);
const access = await ctx.input.api.get<{
access: {
tools: Array<{ toolName: string; gatewayToolName: string }>;
};
}>(
`/api/tool-connections/${connection.connectionId}/test-agents/${ctx.input.fixtures.agent.id}/access`,
);
const write = access.access.tools.find(
(row) => row.toolName === "launch_update",
);
if (!write)
throw new Error(
"The excluded write action has no independently discoverable catalog identity",
);
return {
connectionId: connection.connectionId,
profileId: profile.id,
readEntryId: connection.read[0],
writeEntryId: connection.write[0],
writeGatewayName: write.gatewayToolName,
marker,
};
} catch (error) {
fixtures.delete(ctx);
await provider.close();
throw error;
}
},
gradeGatewayClient,
{
sourceFiles: [
"chat-catalog-domains/external-gateway-client.ts",
"chat-catalog-domains/external-provider.ts",
"chat-catalog-domains/external-support.ts",
"chat-catalog-domains/internal-support.ts",
"chat-scenarios/gateways-profiles.yaml",
],
async act(ctx, state) {
const fixture = fixtures.get(ctx)!;
const active = await pollUntil({
label: "admitted source run for independent scoped denial probe",
deadlineAt: Date.now() + 60_000,
intervalMs: 100,
load: async () => {
await ctx.refreshIssue();
return ctx.allRuns();
},
accept: (rows) =>
rows.some(
(row) =>
row.agentId === ctx.input.fixtures.agent.id &&
row.status === "running",
),
reject: (rows) =>
rows.length &&
rows.every((row) =>
["succeeded", "failed", "cancelled", "timed_out"].includes(
row.status,
),
)
? "Source run settled before scoped-session probe admission"
: undefined,
});
const run = active.find(
(row) =>
row.agentId === ctx.input.fixtures.agent.id &&
row.status === "running",
)!;
const source = ctx.issue();
if (!source || !ctx.input.privateSecrets)
throw new Error(
"Scoped client requires source identity and private redaction",
);
state.probeRunId = run.id;
state.sourceIssueId = source.id;
const key = await ctx.input.api.postSensitive<{
id: string;
token: string;
}>(`/api/agents/${ctx.input.fixtures.agent.id}/keys`, {
name: `scoped-client-probe-${ctx.input.nonce}`,
});
fixture.key = key;
ctx.input.privateSecrets.push(key.token);
const admitted = await privateRequest(
ctx,
"/api/tool-gateway/sessions",
key.token,
{
companyId: ctx.input.fixtures.company.id,
agentId: ctx.input.fixtures.agent.id,
runId: run.id,
issueId: source.id,
},
);
if (admitted.status !== 201) {
// Parse only in private memory and retain only the exact finite classifier.
// In particular, never serialize arbitrary body/error/token fields.
const response = (await admitted
.json()
.catch(() => undefined)) as unknown;
const diagnostic = scopedSessionFailureDiagnostic(
admitted.status,
{
companyId: ctx.input.fixtures.company.id,
agentId: ctx.input.fixtures.agent.id,
runId: run.id,
issueId: source.id,
observedRunStatusAtAdmission: run.status,
},
response,
);
await ctx.input.evidence(
"chat-catalog-scoped-session-admission-failure.json",
diagnostic,
);
throw new Error(
`Active-run scoped session admission failed (HTTP ${diagnostic.httpStatus}; ${diagnostic.reasonCode}); response withheld`,
);
}
fixture.session = (await admitted.json()) as {
sessionId: string;
token: string;
};
ctx.input.privateSecrets.push(fixture.session.token);
state.sessionCreated = true;
const listed = await privateRequest(
ctx,
"/api/tool-gateway/tools",
fixture.session.token,
undefined,
true,
);
if (!listed.ok) {
await listed.body?.cancel();
throw new Error("Scoped tool discovery failed; response withheld");
}
const tools = (await listed.json()) as Array<{
catalogEntryId?: string;
}>;
state.discoveredRead = tools.some(
(row) => row.catalogEntryId === state.readEntryId,
);
state.discoveredWrite = tools.some(
(row) => row.catalogEntryId === state.writeEntryId,
);
const denied = await privateRequest(
ctx,
"/api/tool-gateway/tools/call",
fixture.session.token,
{
tool: state.writeGatewayName,
parameters: { text: "This oracle write must be denied" },
},
true,
);
state.writeProbeStatus = denied.status;
// Retain only the allowlisted policy code, never an arbitrary error body.
const denial = (await denied.json()) as { reasonCode?: unknown };
state.writeProbeReason = ["deny_default", "deny_policy_block"].includes(
String(denial.reasonCode),
)
? denial.reasonCode
: "other";
},
async observe(ctx, state) {
const fixture = fixtures.get(ctx);
if (!fixture) throw new Error("Scoped synthetic fixture missing");
state.effective = await ctx.input.api.get<ToolProfileEffectiveSummary>(
`${root(ctx)}/tools/profiles/effective/agents/${ctx.input.fixtures.agent.id}`,
);
state.decisions = await Promise.all(
(await ctx.allRuns()).map((run) =>
ctx.input.api.get<ToolRunDecisionLookup>(
`${root(ctx)}/tools/runs/${run.id}/decisions`,
),
),
);
state.providerReadCalls = fixture.provider.calls.filter(
(row) =>
row.method === "tools/call" &&
row.name === "launch_read" &&
row.authorized,
).length;
state.providerWriteCalls = fixture.provider.calls.filter(
(row) => row.method === "tools/call" && row.name === "launch_update",
).length;
state.written = fixture.provider.written() ?? null;
const o = await observeExternal(ctx, state);
const tokens = [fixture.key?.token, fixture.session?.token].filter(
(token): token is string => typeof token === "string",
);
state.tokenScanComplete = tokens.length === 2;
state.privateTokenLeaked = tokens.some((token) =>
JSON.stringify([o.reply, o.runs]).includes(token),
);
return { ...o, state };
},
async verifyResult(ctx, _state, value) {
await verifyRenderedReply(ctx, value);
},
async cleanup(ctx, state) {
const fixture = fixtures.get(ctx);
fixtures.delete(ctx);
try {
if (fixture?.session && fixture.key) {
const response = await privateRequest(
ctx,
`/api/tool-gateway/sessions/${fixture.session.sessionId}/revoke`,
fixture.key.token,
{ companyId: ctx.input.fixtures.company.id },
);
if (!response.ok && response.status !== 404)
throw new Error("Scoped fixture session revocation failed");
await response.body?.cancel();
}
if (fixture?.key)
await ctx.input.api.delete(
`/api/agents/${ctx.input.fixtures.agent.id}/keys/${fixture.key.id}`,
{ allowNotFound: true },
);
if (typeof state.profileId === "string")
await ctx.input.api.post(
`${root(ctx)}/tools/profiles/${state.profileId}/unbind`,
{ targetType: "agent", targetId: ctx.input.fixtures.agent.id },
);
if (typeof state.connectionId === "string")
await ctx.input.api.delete(
`/api/tool-connections/${state.connectionId}`,
{ allowNotFound: true },
);
} finally {
await fixture?.provider.close();
}
},
},
),
];