Files
PaperClipAI/tests/runner-e2e/chat-catalog-domains/runtime-failed-runs.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

116 lines
10 KiB
TypeScript

import { expect } from "@playwright/test";
import { pollUntil } from "../api.js";
import { collectRunEvents } from "../run-observations.js";
import { readTaskRunHistory } from "./internal-run-history.js";
import { readRegisteredArtifacts } from "../registered-artifact.js";
import { checks, type Entity } from "./internal-support.js";
import { runtimeDefinition, runtimeWorker, runtimeObservation, observeRuntime } from "./runtime-support.js";
import { unavailableNativeModelFailure } from "./runtime-native-failure.js";
import type { HeartbeatRun } from "../../../packages/shared/src/types/heartbeat.js";
export function exactFailedRunLink(href: string | null, origin: string, companyPrefix: string, worker: { id: string; urlKey?: string }, runId: string) {
if (!href) return false;
try {
const url = new URL(href, origin), parts = url.pathname.split("/").map(decodeURIComponent);
return url.origin === new URL(origin).origin && !url.search && !url.hash && parts.length === 6
&& parts[1] === companyPrefix && parts[2] === "agents" && [worker.id, worker.urlKey].includes(parts[3])
&& parts[4] === "runs" && parts[5] === runId;
} catch { return false; }
}
/** The reply's own link to the failed attempt's run page: the company's prefix, the worker, then the run. */
function failedRunHref(o: { origin: string; links: string[]; state: Record<string, unknown> }) {
const failed = o.state.failedRun as { id?: string } | undefined, worker = o.state.worker as Entity | undefined;
const prefix = String((o.state.task as Entity | undefined)?.identifier ?? "").split("-")[0] ?? "";
if (!failed?.id || !worker?.id || !prefix) return undefined;
return o.links.find(href => exactFailedRunLink(href, o.origin, prefix, worker as { id: string; urlKey?: string }, failed.id!));
}
export function gradeFailedAttempt(value: unknown) {
const c = checks(); const o = runtimeObservation(value);
c.add("actual-failed-and-successful-native-history", !!o, "An actually admitted native failure and a later real successful execution must remain separately readable."); if (!o) return c.result();
const failed = o.state.failedRun as HeartbeatRun | undefined, retained = o.state.retainedFailure as HeartbeatRun | undefined;
const success = o.state.successfulRun as HeartbeatRun | undefined;
c.add("original-failure-not-erased-by-success", !!failed && !!retained && failed.id === retained.id && failed.status === "failed"
&& retained.status === "failed" && retained.errorCode === failed.errorCode && retained.error === failed.error && !!failed.errorCode
&& !!success && success.status === "succeeded" && success.id !== failed.id && o.state.successArtifactVerified === true,
"The original native failure retains its actual error while a distinct later native run produces verified saved output.");
// The reply's wording is not matched. The answer is graded by the run's own
// product read of the failed attempt (whose response carries its error, cost
// and log fields) and by an exact link to that attempt, not the later success.
c.add("read-failed-run-evidence", o.state.failedRunRead === true,
"The chat run read the failed attempt's recorded details (or the task's run history) through a product tool before answering.");
c.add("real-run-source-link", !!failed && !!failedRunHref(o),
"The failure explanation links the actual retained failed run, not only the later success.");
// Absence guard only: an unrecorded cost must not be reported as a dollar
// zero. How the reply describes unavailable data is not matched.
c.add("unrecorded-cost-not-reported-as-zero", failed?.costUsd != null || !/\$\s?0(?:\.0+)?(?!\.?\d)/.test(o.reply),
"Unrecorded usage is never turned into a zero-dollar bill.");
c.add("no-new-work-to-explain-history", o.tasks.length === 1 && o.tasks[0]?.id === (o.state.task as Entity).id
&& o.runs.length === 3 && o.runs.filter(run => run.status === "failed").length === 1 && o.runs.filter(run => run.status === "succeeded").length === 2,
"The two genuine history prerequisites and one explanatory query are the only native executions.");
return c.result();
}
export const runtimeFailedRunCases = [runtimeDefinition("runs-adapters.inspect-failed-run", async ctx => {
const worker = await runtimeWorker(ctx); let task: Entity | undefined;
const failure = await unavailableNativeModelFailure(ctx, worker, async () => {
task = await ctx.input.api.post<Entity>(`/api/companies/${ctx.input.fixtures.company.id}/issues`, { title: "Website task", status: "todo", assigneeAgentId: worker.id,
description: `Create a small semantic HTML launch website with headline Acorn Notes and reference ${ctx.input.nonce}. Save it as website.html and register that actual file as a downloadable artifact work product titled website.html. Verify it and complete this task normally. No deployment, email, new task or remote operation is authorized.` });
return { issueId: task.id };
});
if (!task) throw new Error("Actual failed native task was not created");
const source = ctx.input.page.url();
await ctx.input.page.goto(`/${ctx.input.fixtures.company.issuePrefix}/issues/${task.identifier ?? task.id}`, { waitUntil: "domcontentloaded" });
const marker = ctx.input.page.locator("#main-content").getByTestId("task-chat-collapsible-marker")
.filter({ has: ctx.input.page.getByRole("button", { name: /^Model unavailable(?: ·|$)/ }) });
await expect(marker).toHaveCount(1);
await marker.getByRole("button", { name: /^Model unavailable(?: ·|$)/ }).click();
const runLink = marker.getByRole("link", { name: "View run", exact: true });
const href = await runLink.getAttribute("href");
if (!exactFailedRunLink(href, ctx.input.api.baseURL, ctx.input.fixtures.company.issuePrefix!, worker, failure.failedRun.id))
throw new Error("Visible failed-run detail link does not identify the original company, worker and attempt");
await runLink.click();
await ctx.input.page.waitForURL(url => url.pathname === new URL(href!, ctx.input.api.baseURL).pathname);
const original = await ctx.input.api.get<HeartbeatRun>(`/api/heartbeat-runs/${failure.failedRun.id}`);
if (original.id !== failure.failedRun.id || original.status !== "failed" || original.companyId !== worker.companyId || original.agentId !== worker.id)
throw new Error("The visible retry target does not retain the genuine original failure");
const retry = ctx.input.page.locator("#main-content").getByRole("button", { name: "Retry", exact: true });
await expect(retry).toBeVisible();
await retry.click();
const done = await pollUntil({ label: "Real successful task retry after restoring the admitted model", deadlineAt: Date.now() + 240_000,
load: async () => ({ task: await ctx.input.api.get<Entity>(`/api/issues/${task!.id}`), runs: await ctx.allRuns() }),
accept: value => value.task.status === "done" && value.runs.length === 2 && value.runs.some(run => run.id !== failure.failedRun.id && run.status === "succeeded"),
reject: value => value.runs.length > 2 || value.runs.some(run => run.id !== failure.failedRun.id && ["failed", "cancelled", "timed_out"].includes(run.status)) ? "The one declared successful history retry did not succeed" : undefined });
const successfulRun = done.runs.find(run => run.id !== failure.failedRun.id)!;
const [artifact] = await readRegisteredArtifacts(ctx.input.api, task.id, successfulRun.id, ["website.html"]);
if (!artifact?.contentVerified || !artifact.content.includes(ctx.input.nonce) || !/<html|<!doctype html/i.test(artifact.content)) throw new Error("The later run did not produce the genuine expected HTML delivery");
const response = await ctx.input.api.request.get(`/api/heartbeat-runs/${failure.failedRun.id}/log?offset=0&limitBytes=4096`);
const failureLogAvailable = response.ok(); await response.dispose();
await ctx.input.page.goto(source, { waitUntil: "domcontentloaded" });
return { worker, task: done.task, ...failure, successfulRun, successArtifactVerified: true, failureLogAvailable };
}, gradeFailedAttempt, {
sourceFiles: ["chat-catalog-domains/runtime-failed-runs.ts", "chat-catalog-domains/runtime-native-failure.ts", "chat-catalog-domains/runtime-support.ts", "chat-catalog-domains/internal-run-history.ts", "chat-catalog-domains/api-read-receipts.ts", "chat-scenarios/runs-adapters.yaml", "../../ui/src/components/TaskChatThread.tsx", "../../ui/src/components/task-chat/TaskChatMarker.tsx", "../../ui/src/components/task-chat/TaskChatThreadView.tsx", "../../ui/src/pages/AgentDetail.tsx"],
oracleVersion: "paperclip.chat-catalog.runtime.inspect-failed-run.v4", expectedRunCount: 3, attemptTimeoutMs: 720_000,
runPolicy: { kind: "terminal-counts", terminalStatusCounts: { failed: 1, succeeded: 2 } },
async observe(ctx, state) {
state.retainedFailure = await ctx.input.api.get(`/api/heartbeat-runs/${(state.failedRun as { id: string }).id}`);
const o = await observeRuntime(ctx, state); const taskId = (state.task as Entity).id, failedId = (state.failedRun as { id: string }).id;
let failedRunRead = false;
for (const run of o.runs.filter(row => row.agentId === o.leadId && row.contextSnapshot?.issueId === ctx.issue()?.id)) {
const events = await collectRunEvents<Record<string, unknown>>(afterSeq => ctx.input.api.get(`/api/heartbeat-runs/${run.id}/events?afterSeq=${afterSeq}&limit=1000`));
failedRunRead = failedRunRead || readTaskRunHistory(events, taskId, [failedId], (state.task as Entity).identifier);
}
o.state.failedRunRead = failedRunRead;
return o;
},
async verifyResult(ctx, state, value) {
const source = ctx.input.page.url(); const failed = state.failedRun as { id: string };
// Follow the reply's own link, so a route the person cannot open fails here.
const o = runtimeObservation(value), href = o && failedRunHref(o);
if (!href) throw new Error("Reply did not link the failed attempt's run page");
await ctx.input.page.goto(new URL(href, ctx.input.api.baseURL).href, { waitUntil: "domcontentloaded" });
await expect(ctx.input.page.getByText(/failed/i).first()).toBeVisible();
const actual = await ctx.input.api.get<HeartbeatRun>(`/api/heartbeat-runs/${failed.id}`);
if (actual.id !== failed.id || actual.status !== "failed") throw new Error("Failure result UI did not retain the intended failed attempt");
await ctx.input.evidence("runtime-retained-failed-attempt-ui.json", { runId: failed.id, errorCode: actual.errorCode, taskId: (state.task as Entity).id, url: ctx.input.page.url() });
await ctx.input.page.goto(source, { waitUntil: "domcontentloaded" });
},
})];