Files
PaperClipAI/scripts/runner-api-eval-worker.ts
T
DottaandPaperclip 5bddff0920 feat(runner): add guarded API search and call fallback (#13003)
## Thinking Path

> - Paperclip manages AI agents and their work.
> - The new runner gives agents dedicated tools for common tasks.
> - Some API operations and parameters have no dedicated tool.
> - Agents need a controlled way to find and use those operations.
> - This pull request adds API search and calls through the real server
routes.
> - Existing tools remain the preferred path. The new tools are disabled
by default.
> - Paired tests measure correctness, tool choice, cost and time.

## Linked Issues or Issue Description

**Subsystem affected**

Paperclip Runner contracts, production tool authority and the server API
catalog.

**Problem or motivation**

The runner cannot use much of the API described by the old Paperclip
skill. A generic HTTP client would also let agents bypass runner control
rules.

**Proposed solution**

Add `search_api` and `call_api`. Resolve calls from the mounted API
catalog. Use server-held, run-bound credentials. Preserve route checks
and runner lifecycle rules. Keep the tools disabled until an operator
enables selected companies.

**Alternatives considered**

A dedicated tool for every endpoint would add a large initial prompt. An
unrestricted HTTP tool would weaken authorization and replay controls.

**Roadmap alignment**

This extends the native runner tooling. The repository owner requested
this design and implementation. The roadmap and related open PRs were
checked. No duplicate API escape-hatch PR was found.

## What Changed

- Register two compact fallback tools in canonical contracts and
provider projections.
- Build deterministic API discovery from OpenAPI, mounted experimental
routes and the old skill reference.
- Execute bounded JSON, text, file and download requests through
authenticated HTTP routes.
- Recheck active runs, company access and work modes. Block runner
lifecycle, scheduling, credential and approval bypasses. Keep routine
annotation collaboration available.
- Retain mutation receipts. Report uncertain outcomes without blindly
repeating writes.
- Add a company rollout gate and a durable eval worker with complete
cost accounting checks.
- Record child-task creation in the activity log with the agent and run.
- Add contract, authorization, file, replay and real runnerd/PRP/HTTP
tests.
- Document rollout gates and paid coverage limits. The companion eval
repository retains immutable attempts and reports.

## Verification

- Final app commit `da58370524c3626a744eec20164397c5fb6ba9ef`: all 32
checks passed; the unrelated Storybook visual check was skipped.
Greptile 5/5; no unresolved review threads.

- Full Linux build and recursive typecheck passed. Repository tests were
run by project and serialized shard; all 143 serialized server suites
passed.
- Runner TypeScript: 1,599 passed, two skipped. Rust release: 451
passing test reports. Conformance and replay parity passed. The required
API check passed 837 tests, including runnerd → PRP → authority → real
HTTP.
- Bindings cannot enable API tools without the explicit deployment flag.
Unit and real-authority tests prove the default-off boundary.
- The standalone API check builds and stages its own binary. It passed
after existing staged and debug binaries were removed from the test
container.
- UI and CLI tests passed. Initial environment failures (missing jq,
Docker overlay file identity, and parallel linker memory pressure) and
focused passing reruns are retained. The macOS full runner suite has
platform-specific failures; Linux is the qualified full-check platform.
- Eval harness: 27 tests passed; existing CI discovery ran 86 tests with
two unrelated skips. Credential export rejection is tested against the
actual report command.
- Luna and OpenRouter Sonnet each passed 60 common-workflow runs: ten
workflows, three repetitions per arm, zero unnecessary API fallback.
- Sonnet passed 11 selected capability/contract cases after fixes.
Gemini passed three smoke cases. DeepSeek exceeded the 120-second limit
and remains unqualified.
- Luna's two cost flags received focused follow-up. The original flags
and a later n=1 latency flag remain visible. Sonnet had no cost or
latency increase above 20%.
- The catalog contains 785 entries; 58 were exercised across all stages.
Most operation probes remain unrun and some need additional fixtures.
Authored probes do not establish successful coverage.
- Total conservative accounted cost: $9.875960. Active paid-campaign
time: 88.16/90 minutes. No missing accounting. Later security and
harness fixes have provider-free verification; no paid validation is
claimed for those revisions.
- Inspect the [qualification
report](https://github.com/paperclipai/paperclip-evals/blob/codex/seach-call-api-tools/evals/runner-api-tools/reports/2026-09-07-production/READINESS.md)
and [verification
record](https://github.com/paperclipai/paperclip-evals/blob/codex/seach-call-api-tools/evals/runner-api-tools/reports/2026-09-07-production/verification.json).

## Risks

- This is a broad authenticated API surface. Keep the default-off gate
until an operator selects initial rollout companies.
- Paid coverage is incomplete. Small regression samples do not prove all
workflows are unchanged.
- A timeout or server failure can follow a committed mutation. The
result reports an unknown outcome and requires state inspection.
- The new definitions add prompt tokens. The report retains cost flags
and cache variation.
- No database migration is required.
- Repository rules require code-owner approval before merge. Technical
CI and automated review are complete.

## Model Used

OpenAI Codex based on GPT-6 assisted with code, tests and review. The
exact serving model ID and context window are not exposed in this
session. It used reasoning, tool calls and code execution.

Eval models: `gpt-5.6-luna` with low reasoning,
`openrouter/anthropic/claude-sonnet-5`,
`openrouter/google/gemini-3.8-flash`, and
`openrouter/deepseek/deepseek-v4-flash-0731`. Attempts retain runtime
versions, model identity, usage and source provenance.

## 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
- [x] All Paperclip CI gates are green
- [x] Greptile is 5/5 with no open P2s, recommendations, or follow-ups
- [x] I will address all Greptile and reviewer comments before
requesting merge

---------

Co-authored-by: Paperclip <noreply@paperclip.ing>
2026-09-07 14:14:43 -05:00

218 lines
17 KiB
TypeScript
Raw Blame History

This file contains ambiguous Unicode characters
This file contains Unicode characters that might be confused with other characters. If you think that this is intentional, you can safely ignore this warning. Use the Escape button to reveal them.
/** JSONL worker for the companion paperclip-evals API suite. Never selects cases or retries. */
import { createHash, randomUUID } from "node:crypto";
import { createReadStream, realpathSync } from "node:fs";
import { execFileSync } from "node:child_process";
import { mkdir, writeFile, readFile } from "node:fs/promises";
import { join, resolve } from "node:path";
import { createInterface } from "node:readline";
import { homedir } from "node:os";
import { runnerApiCatalog } from "../server/src/services/native-runtime/runner-api-catalog.js";
import { startRunnerApiTestServer } from "../server/src/__tests__/helpers/runner-api-server.js";
import { registerRunnerPrpAuthority } from "../server/src/realtime/runner-prp-ws.js";
import { createRunnerdCodexTransport, defaultCapabilityRunnerdBinary } from "../packages/paperclip-runner/src/live/runnerd-codex-transport.js";
import { createSkilllessCodexThreadConfig } from "../packages/paperclip-runner/src/drivers/codex/codex-app-server-driver.js";
import { AttemptJournal } from "../packages/paperclip-runner/src/evals/attempt-journal.js";
import { estimateModelCostNanodollars } from "../packages/paperclip-runner/src/evals/model-pricing.js";
if (process.argv.includes("--catalog")) {
process.stdout.write(JSON.stringify(runnerApiCatalog()) + "\n");
process.exit(0);
}
if (!process.argv.includes("--jsonl")) throw new Error("Use --catalog or --jsonl; there is no default campaign");
process.env.PAPERCLIP_AGENT_JWT_SECRET = randomUUID() + randomUUID();
const output = (value: unknown) => process.stdout.write("RUNNER_API_EVAL " + JSON.stringify(value) + "\n");
const OPENROUTER_MODELS = new Set(["openrouter/anthropic/claude-sonnet-5", "openrouter/deepseek/deepseek-v4-flash-0731", "openrouter/google/gemini-3.8-flash"]);
// The controller selects and injects one provider credential. The worker never
// reads ambient home credential files or desktop keychains.
function openRouterEnvironment() {
const key = process.env.OPENROUTER_API_KEY;
if (!key || /\s/.test(key)) throw new Error("Controller must inject OPENROUTER_API_KEY");
return { PATH: process.env.PATH, OPENROUTER_API_KEY: key };
}
const server = await startRunnerApiTestServer();
const fileDigests = new Map<string, Promise<string | null>>();
function fileDigest(path: string): Promise<string | null> {
let digest = fileDigests.get(path);
if (!digest) {
digest = (async () => {
const hash = createHash("sha256");
for await (const chunk of createReadStream(path)) hash.update(chunk);
return hash.digest("hex");
})().catch(() => null);
fileDigests.set(path, digest);
}
return digest;
}
const runtimeBuild = {
runnerBinarySha256: await fileDigest(defaultCapabilityRunnerdBinary()),
workerSourceSha256: await fileDigest(resolve("scripts/runner-api-eval-worker.ts")),
lockfileSha256: await fileDigest(resolve("pnpm-lock.yaml")),
};
output({ ready: true });
const record = (value: unknown): Record<string, any> => value && typeof value === "object" ? value as Record<string, any> : {};
try {
for await (const line of createInterface({ input: process.stdin })) {
const request = JSON.parse(line);
if (request.shutdown) break;
const directory = resolve(request.outputDirectory);
await mkdir(directory, { recursive: true, mode: 0o700 });
const journal = new AttemptJournal(join(directory, "journal.jsonl"));
const startedAt = new Date().toISOString();
journal.append({ kind: "attempt_started", attemptId: request.attemptId, model: request.model, startedAt });
const started = performance.now();
const isOpenRouter = OPENROUTER_MODELS.has(request.model);
const provider = isOpenRouter ? "opencode" : request.model === "claude-sonnet-5" ? "acpx" : "codex";
let providerVersion: string | null = null;
const fixture = await server.fixture({ mode: request.mode, apiToolsEnabled: request.arm !== "baseline", reset: true });
const substitutions = Object.fromEntries(Object.entries(fixture).filter(([, value]) => typeof value === "string"));
const expand = (value: any): any => typeof value === "string" ? value.replace(/\{\{(\w+)\}\}/g, (_, key) => String(substitutions[key] ?? (() => { throw new Error(`Unknown fixture variable ${key}`); })())) : Array.isArray(value) ? value.map(expand) : value && typeof value === "object" ? Object.fromEntries(Object.entries(value).map(([key, entry]) => [key, expand(entry)])) : value;
const calls: Record<string, any>[] = [], notifications: unknown[] = [];
const observedToolCallIds = new Set<string>();
const definitions = await fixture.authority.definitions();
const prompt = expand(request.prompt ?? "");
let error: string | null = null, evidence: unknown = null, thread: Record<string, any> = {}, usage: Record<string, any> | null = null;
let usageUpdates = 0, lastUsageFingerprint = "", terminalSeen = false, providerTurnStarted = false;
let diagnosticTail: string | null = null;
let bundle: ReturnType<typeof createRunnerdCodexTransport> | undefined;
let timer: ReturnType<typeof setTimeout> | undefined;
const invoke = async (tool: string, args: unknown, callId: string) => {
observedToolCallIds.add(callId);
if (observedToolCallIds.size > 12) throw new Error("Attempt tool-call ceiling reached");
const call: Record<string, any> = { tool, arguments: args, callId, startedMs: performance.now() - started };
calls.push(call);
journal.append({ kind: "tool_dispatch", ...call });
try { call.result = await fixture.authority.execute({ tool, arguments: args, callId }); }
catch (caught) { call.error = caught instanceof Error ? caught.message : String(caught); call.status = record(caught).status; }
call.durationMs = performance.now() - started - call.startedMs;
journal.append({ kind: "tool_result", ...call });
return call;
};
try {
if (request.calls) {
for (const [i, call] of request.calls.entries()) await invoke(call.tool, expand(call.arguments), call.callId ?? `direct-${i}`);
} else {
if (request.model === "claude-sonnet-5" && (process.platform !== "linux" || process.arch !== "x64")) throw new Error("Qualified ACPX Claude requires Linux x64; no provider turn was dispatched");
if (!request.reservationId || request.maxCostUsd !== 0.5 || !["gpt-5.6-luna", "claude-sonnet-5", ...OPENROUTER_MODELS].includes(request.model)) throw new Error("Paid attempt requires ledger reservation and qualified model");
if (isOpenRouter) {
providerVersion = execFileSync(resolve("packages/paperclip-runner/node_modules/opencode-ai/bin/opencode.exe"), ["--version"], { encoding: "utf8" }).trim();
if (providerVersion !== "1.18.29") throw new Error("OpenCode profile requires version 1.18.29");
}
const providerEnvironment = isOpenRouter ? openRouterEnvironment() : request.model === "claude-sonnet-5" ? (() => {
const token = process.env.CLAUDE_CODE_OAUTH_TOKEN;
if (!token) throw new Error("Controller must inject CLAUDE_CODE_OAUTH_TOKEN");
return { PATH: process.env.PATH, CLAUDE_CODE_OAUTH_TOKEN: token };
})() : undefined;
bundle = createRunnerdCodexTransport({
provider, acpxAgent: "claude", acpxPermissionMode: "approve-reads",
environment: { ...providerEnvironment, PAPERCLIP_PROVIDER_TRACE_PATH: join(directory, "provider-trace.jsonl"), PAPERCLIP_PROVIDER_TRACE_MAX_BYTES: String(32 * 1024 * 1024) },
codexCommand: request.model === "gpt-5.6-luna" ? realpathSync(execFileSync("which", ["codex"], { encoding: "utf8" }).trim()) : undefined,
sourceCodexHome: process.env.CODEX_HOME ?? join(homedir(), ".codex"),
runnerBinary: defaultCapabilityRunnerdBinary(), stateDirectory: join(server.root, `runner-${request.attemptId}`),
lifecyclePolicy: { mode: "per_turn", idleTimeoutMs: null },
prpIdentity: { runnerInstanceId: `runner-${fixture.runId}`, environmentLeaseId: `lease-${fixture.runId}`, runId: fixture.runId, normalizedSessionId: `session-${fixture.runId}`, turnId: `turn-${fixture.runId}`, itemId: `item-${fixture.runId}` },
controlPlaneRegistration: (prp) => registerRunnerPrpAuthority({ companyId: fixture.companyId, runId: fixture.runId, authority: prp }),
});
const active = bundle;
timer = setTimeout(() => { error = "Attempt exceeded 120 seconds"; void active.transport.close(error); }, 120_000);
bundle.transport.setServerRequestHandler(async (request) => {
const params = record(request.params);
const result = await invoke(String(params.tool), params.arguments, String(params.callId));
return { success: !result.error, contentItems: [{ type: "inputText", text: JSON.stringify(result.error ? { ok: false, error: result.error, status: result.status } : { ok: true, result: result.result }) }] };
});
await bundle.transport.request("initialize", {});
thread = await bundle.transport.request("thread/start", {
cwd: fixture.workspace, model: request.model,
completionContract: { revision: "runner-api-eval-v1", criterionIds: ["objective"] },
config: { ...createSkilllessCodexThreadConfig(fixture.workspace), model_reasoning_effort: "low" },
permissions: "paperclip-runner-workspace-only", runtimeWorkspaceRoots: [fixture.workspace], approvalPolicy: "never",
baseInstructions: "You are operating a disposable real Paperclip company. Use the provided tools to do the user's task. Do not use shell, network, skills, or credentials. Stop when the requested work is verified. " + (request.arm === "baseline" ? "" : "Prefer available dedicated tools. Only use search_api and call_api when no dedicated tool supports the required operation or parameters. Do not search before ordinary dedicated tool use."),
dynamicTools: definitions, experimentalRawEvents: true, persistExtendedHistory: true,
});
if (request.preflight) {
terminalSeen = true;
usage = { inputTokens: 0, outputTokens: 0, cachedInputTokens: 0, estimatedCostNanodollars: 0, providerRequests: 0, preflight: true };
} else {
journal.append({ kind: "provider_turn_dispatch", thread, at: new Date().toISOString() });
providerTurnStarted = true;
await bundle.transport.request("turn/start", { input: [{ type: "text", text: prompt }] });
for await (const notification of bundle.transport.notifications()) {
const retained = { ...notification, observedAt: new Date().toISOString() };
journal.append({ kind: "notification", ...retained });
notifications.push(retained);
const item = record(record(notification.params).item);
if (["dynamicToolCall", "tool_call", "tool_result"].includes(item.type) && typeof item.id === "string") {
observedToolCallIds.add(item.id);
if (observedToolCallIds.size > 12) { error = "Attempt tool-call ceiling reached, including provider-rejected calls"; break; }
}
if (notification.method === "thread/tokenUsage/updated") {
const params = record(notification.params);
const tokenUsage = record(params.tokenUsage);
const cumulative = record(tokenUsage.total ?? tokenUsage.totalTokenUsage ?? tokenUsage.total_token_usage ?? tokenUsage);
if (request.model === "claude-sonnet-5" && params.runDeltaAvailable !== true) continue;
const total = request.model === "claude-sonnet-5" ? record(tokenUsage.runDelta) : cumulative;
const cacheWriteTokens = total.cacheWriteTokens ?? 0;
const inputTokens = (total.inputTokens ?? total.input_tokens) + (request.model === "claude-sonnet-5" ? (total.cacheReadTokens ?? 0) : isOpenRouter ? (total.cacheReadTokens ?? total.cachedInputTokens ?? 0) : 0);
const outputTokens = total.outputTokens ?? total.output_tokens;
const cachedInputTokens = total.cacheReadTokens ?? total.cachedInputTokens ?? total.cache_read_input_tokens ?? 0;
if ([inputTokens, outputTokens, cachedInputTokens, cacheWriteTokens].every(value => Number.isSafeInteger(value) && value >= 0)) {
const fingerprint = `${inputTokens}:${outputTokens}:${cachedInputTokens}`;
if (fingerprint !== lastUsageFingerprint) { usageUpdates++; lastUsageFingerprint = fingerprint; }
usage = { inputTokens, outputTokens, cachedInputTokens, cacheWriteTokens, providerRequests: (total.requests ?? total.providerRequests ?? 0) > 0 ? total.requests ?? total.providerRequests : null, providerUsageUpdateCount: usageUpdates, providerRequestCountSource: (total.requests ?? total.providerRequests ?? 0) > 0 ? "provider" : "not reported; providerUsageUpdateCount is a proxy", providerReportedCostNanodollars: cumulative.providerCostUsd > 0 ? Math.round(cumulative.providerCostUsd * 1e9) : null, ...estimateModelCostNanodollars(request.model, { inputTokens, outputTokens, cachedInputTokens }) };
// A conservative one-hour cache-write ceiling (2× standard input)
// covers either cache duration when the runtime omits its TTL.
if (request.model === "claude-sonnet-5" || request.model === "openrouter/anthropic/claude-sonnet-5") {
usage.estimatedCostNanodollars += cacheWriteTokens * 4_000;
usage.cacheWriteUsdPerMillionTokens = 4;
usage.cacheWritePricing = "Conservative one-hour cache-write rate; TTL not exposed";
}
if (Math.max(usage.estimatedCostNanodollars, usage.providerReportedCostNanodollars ?? 0) >= 500_000_000) { error = "Attempt cost ceiling reached"; break; }
}
}
if (notification.method === "turn/completed") {
terminalSeen = true;
const params = record(notification.params);
const terminal = record(params.turn);
if ((params.status ?? terminal.status) === "failed") error = String(record(params.error ?? terminal.error).message ?? "Provider turn failed");
break;
}
}
}
evidence = bundle.evidence();
if (!request.preflight && (!usage || usage.inputTokens + usage.outputTokens === 0)) {
usage = null;
error ??= "Missing provider accounting; block further paid attempts";
}
}
} catch (caught) { error ??= caught instanceof Error ? caught.message : String(caught); }
finally {
if (timer) clearTimeout(timer);
if (bundle) {
evidence = bundle.evidence();
diagnosticTail = await readFile(join(server.root, `runner-${request.attemptId}`, "diagnostics", "runnerd.stderr.log"), "utf8").then(text => text.slice(-16000).replace(/sk-(?:ant-|or-v1-)[A-Za-z0-9_-]+/g, "[redacted]").replace(/Bearer\s+[^\s"']+/gi, "Bearer [redacted]").replace(/eyJ[A-Za-z0-9_-]+\.[A-Za-z0-9_-]+\.[A-Za-z0-9_-]+/g, "[redacted JWT]")).catch(() => null);
}
await bundle?.transport.close();
}
if (!providerTurnStarted && !request.calls) {
usage = { inputTokens: 0, outputTokens: 0, cachedInputTokens: 0, estimatedCostNanodollars: 0, providerRequests: 0, accountingProvenance: "No provider turn/start was dispatched" };
}
const artifact = {
schema: "paperclip-runner/eval-session-artifact/v1", attemptId: request.attemptId,
requestedModel: request.model ?? "provider-free", provider, driver: "real-server-api-tools",
providerSessionId: record(thread.thread).id ?? null, effectiveModel: record(thread.thread).model ?? null,
providerVersion: request.model === "gpt-5.6-luna" ? execFileSync("codex", ["--version"], { encoding: "utf8" }).trim() : providerVersion ?? record(evidence).providerVersion ?? null,
runtimeVersions: { node: process.versions.node, acpx: record(evidence).providerVersion, agentServer: record(evidence).agentServerVersion, agentRuntime: record(evidence).agentRuntimeVersion },
runtimeBuild,
timing: { startedAt, finishedAt: new Date().toISOString(), durationMs: performance.now() - started },
usage, accountingComplete: Boolean(request.calls) || !providerTurnStarted || terminalSeen, providerTurnStarted, observedProviderToolCalls: observedToolCallIds.size, diagnosticTail, error, calls, notifications, evidence, thread, prompt, arm: request.arm ?? "treatment",
fixture: substitutions, state: await fixture.snapshot(),
toolSchemaBytes: Buffer.byteLength(JSON.stringify(definitions)), definitions,
};
// The controller supplies a unique retained attempt directory; never overwrite evidence.
journal.append({ kind: "attempt_finished", error, usage, terminalSeen, at: new Date().toISOString() });
journal.close();
await writeFile(join(directory, "artifact.json"), JSON.stringify(artifact, null, 2), { flag: "wx", mode: 0o600 });
output({ attemptId: request.attemptId, artifactPath: join(directory, "artifact.json"), error, usage });
}
} finally { await server.close(); }