paperclip/packages/paperclip-runner/scripts/render-runner-workflow-eval...

364 lines
11 KiB
JavaScript

import { createHash } from "node:crypto";
import { spawn } from "node:child_process";
import { access, mkdir, readFile, rm, writeFile } from "node:fs/promises";
import { resolve } from "node:path";
const EVAL_PROGRAM_RELATIVE_PATH =
"evals/paperclip-runner/tools/eval_program.py";
function json(value) {
return `${JSON.stringify(value, null, 2)}\n`;
}
function sha256(value) {
return createHash("sha256").update(value).digest("hex");
}
function safeSegment(value) {
const segment = String(value)
.trim()
.replaceAll(/[^0-9A-Za-z._-]+/g, "-")
.replaceAll(/^-+|-+$/g, "");
if (!segment) throw new Error("Evalbook attempt identity is empty");
return segment;
}
function candidateDescriptor(report, candidateId) {
const descriptor =
report.bundle.providerVersions?.[candidateId] ?? candidateId;
const separator = descriptor.indexOf(":");
return separator < 0
? { driver: "unknown driver", model: descriptor }
: {
driver: descriptor.slice(0, separator),
model: descriptor.slice(separator + 1),
};
}
function scoreChecks(result) {
const dimensionChecks = Object.values(result.scorecard.dimensions).map(
(dimension) => ({
id: dimension.dimension,
kind: "workflow_dimension",
passed: dimension.passed === true,
detail:
dimension.score === null
? dimension.reasons.join("; ") || "not scored"
: `score ${dimension.score}${
dimension.reasons.length === 0
? ""
: `; ${dimension.reasons.join("; ")}`
}`,
evidenceRefs: [],
}),
);
const observationChecks = [
["lifecycle", result.observation.lifecycle?.checks ?? []],
["continuation", result.observation.continuation?.checks ?? []],
["presentation", result.observation.presentation?.checks ?? []],
].flatMap(([group, checks]) =>
checks.map((check) => ({
id: `${group}.${check.id}`,
kind: `${group}_check`,
passed: check.passed === true,
detail: check.reason ?? (check.passed ? "passed" : "failed"),
evidenceRefs: [],
})),
);
return [...dimensionChecks, ...observationChecks];
}
function disposition(result) {
if (result.scorecard.overall.passed === true) return "passed";
if (
result.observation.classification === "infrastructure_failure" ||
result.observation.classification === "skipped"
) {
return "infrastructure_failure";
}
return "behavior_failure";
}
function infrastructureErrors(result) {
if (disposition(result) !== "infrastructure_failure") return [];
return [
result.observation.failure?.message ??
`workflow execution was ${result.observation.classification}`,
];
}
export function runnerWorkflowEvalbookAttempt({
report,
result,
caseDefinition,
}) {
const { driver, model } = candidateDescriptor(report, result.candidateId);
const identity = {
generatedAt: report.generatedAt,
bundleId: report.bundle.id,
caseId: result.scenarioId,
candidateId: result.candidateId,
};
const timestamp = safeSegment(report.generatedAt);
const attemptId = `${timestamp}-${safeSegment(result.scenarioId)}-${safeSegment(
result.candidateId,
)}-${sha256(JSON.stringify(identity)).slice(0, 10)}`;
const checks = scoreChecks(result);
const costUsd = result.observation.metrics.costUsd;
const infrastructure =
disposition(result) === "infrastructure_failure"
? result.observation.failure
: undefined;
const artifact = {
schema: "paperclip-runner/workflow-eval-artifact/v1",
attemptId,
createdAt: report.generatedAt,
requestedModel: model,
provider: result.observation.provider,
driver,
providerVersion:
report.bundle.runnerBuild ?? report.bundle.runnerVersion ?? "unknown",
providerSessionId: result.observation.base.trace.sessionId,
retainedSession: false,
retainedSessionStatus: "not retained by safe workflow eval projection",
usage: {
agentTurns: result.observation.metrics.attempts,
inputTokens: result.observation.metrics.totalTokens,
cachedInputTokens: 0,
outputTokens: 0,
reasoningTokens: 0,
...(costUsd === undefined
? {}
: { estimatedCostNanodollars: Math.round(costUsd * 1_000_000_000) }),
},
turn: {
status:
result.observation.classification === "completed"
? "completed"
: "failed",
},
snapshot: {
createdAt: report.generatedAt,
providerModel: {
id: model,
provider: result.observation.provider,
},
transcript: [],
evidence: [],
},
devtools: { revisions: [] },
...(infrastructure === undefined
? {}
: {
infrastructureFailure: {
class: infrastructure.code,
category: infrastructure.category,
retryable: infrastructure.retryable,
},
}),
workflow: {
bundle: report.bundle,
observation: result.observation,
scorecard: result.scorecard,
},
};
const score = {
schema: "paperclip-runner/eval-score/v1",
attemptId,
caseId: result.scenarioId,
checks,
disposition: disposition(result),
infrastructureErrors: infrastructureErrors(result),
passed: result.scorecard.overall.passed === true,
digest: `sha256:${sha256(JSON.stringify({ identity, checks }))}`,
};
const evalCase = {
schema: "paperclip-runner/workflow-eval-case/v1",
id: result.scenarioId,
title: caseDefinition?.title ?? result.scenarioId,
description: caseDefinition
? `Runner workflow case. Tags: ${caseDefinition.tags.join(", ")}.`
: "Runner workflow case.",
prompt:
"Redacted by the safe Runner workflow eval projection; inspect the authored workflow case for orchestration steps.",
fixture: "runner-workflow-live-harness",
authority: {
controlPlaneOwned: result.observation.base.controlPlaneOwned,
},
checks: Object.entries(caseDefinition?.assertions ?? {}).map(
([id, expected]) => ({
id,
kind: "workflow_assertion",
expected,
}),
),
...(caseDefinition === undefined
? {}
: { workflowDefinition: caseDefinition }),
};
const config = {
schema: "paperclip-runner/workflow-eval-config/v1",
id: result.candidateId,
model,
provider: result.observation.provider,
driver,
runnerVersion: report.bundle.runnerVersion,
runnerBuild: report.bundle.runnerBuild,
promptPolicyId: report.bundle.promptPolicyId,
};
return { attemptId, artifact, score, case: evalCase, config };
}
async function writeImmutable(path, value) {
const content = json(value);
try {
const existing = await readFile(path, "utf8");
if (existing !== content) {
throw new Error(`Immutable Evalbook record changed: ${path}`);
}
} catch (error) {
if (error?.code !== "ENOENT") throw error;
await writeFile(path, content, { flag: "wx", mode: 0o600 });
}
}
export async function writeRunnerWorkflowEvalbookAttempts({
report,
runsRoot,
caseForId,
}) {
await mkdir(runsRoot, { recursive: true });
const attempts = [];
for (const result of report.results) {
const attempt = runnerWorkflowEvalbookAttempt({
report,
result,
caseDefinition: caseForId?.(result.scenarioId),
});
const directory = resolve(runsRoot, attempt.attemptId);
await mkdir(directory, { recursive: true, mode: 0o700 });
await Promise.all([
writeImmutable(resolve(directory, "artifact.json"), attempt.artifact),
writeImmutable(resolve(directory, "score.json"), attempt.score),
writeImmutable(resolve(directory, "case.json"), attempt.case),
writeImmutable(resolve(directory, "config.json"), attempt.config),
]);
attempts.push(attempt.attemptId);
}
return attempts;
}
async function existingPath(paths) {
for (const path of paths.filter(Boolean)) {
try {
await access(path);
return resolve(path);
} catch {
// Continue through the explicit and conventional checkout locations.
}
}
return null;
}
export async function resolveCanonicalEvalProgram(packageRoot, environment) {
const fromRoot = environment.PAPERCLIP_EVALS_ROOT
? resolve(environment.PAPERCLIP_EVALS_ROOT, EVAL_PROGRAM_RELATIVE_PATH)
: null;
const program = await existingPath([
environment.PAPERCLIP_EVALBOOK_PROGRAM,
fromRoot,
resolve(packageRoot, "../../.paperclip-evals", EVAL_PROGRAM_RELATIVE_PATH),
resolve(
packageRoot,
"../../../paperclip-evals",
EVAL_PROGRAM_RELATIVE_PATH,
),
resolve(
packageRoot,
"../../../../paperclip-evals",
EVAL_PROGRAM_RELATIVE_PATH,
),
]);
if (program !== null) return program;
throw new Error(
`Canonical Evalbook generator not found. Set PAPERCLIP_EVALBOOK_PROGRAM to ${EVAL_PROGRAM_RELATIVE_PATH} in a paperclip-evals checkout.`,
);
}
async function run(command, args) {
await new Promise((accept, reject) => {
const child = spawn(command, args, { stdio: "inherit" });
child.once("error", reject);
child.once("exit", (code, signal) => {
if (code === 0) accept();
else
reject(
new Error(
`${command} exited ${code ?? `for signal ${signal ?? "unknown"}`}`,
),
);
});
});
}
export async function renderRunnerWorkflowWithCanonicalEvalbook({
packageRoot,
outputDirectory,
report,
caseForId,
environment = process.env,
}) {
const program = await resolveCanonicalEvalProgram(packageRoot, environment);
const viewerRoot = resolve(
environment.PAPERCLIP_EVAL_VIEWER_ROOT ??
resolve(packageRoot, "dist-issue-thread"),
);
await access(resolve(viewerRoot, "index.html")).catch(() => {
throw new Error(
"Evalbook requires the chat viewer. Run pnpm --filter @paperclipai/paperclip-runner build:issue-thread first.",
);
});
const runsRoot = resolve(outputDirectory, "evalbook-runs");
await rm(runsRoot, { recursive: true, force: true });
const attempts = await writeRunnerWorkflowEvalbookAttempts({
report,
runsRoot,
caseForId,
});
await Promise.all([
rm(resolve(outputDirectory, "attempts"), { recursive: true, force: true }),
rm(resolve(outputDirectory, "tests"), { recursive: true, force: true }),
rm(resolve(outputDirectory, "index.html"), { force: true }),
rm(resolve(outputDirectory, "latest.html"), { force: true }),
rm(resolve(outputDirectory, "live-report.html"), { force: true }),
rm(resolve(outputDirectory, "deterministic-report.html"), { force: true }),
]);
await run(environment.PYTHON ?? "python3", [
program,
"report",
"--runs-root",
runsRoot,
"--output",
outputDirectory,
"--viewer-root",
viewerRoot,
]);
const programBytes = await readFile(program);
const manifest = {
schema: "paperclip.runner.workflow-evalbook-render.v1",
generatedAt: report.generatedAt,
generator: {
path: program,
sha256: `sha256:${sha256(programBytes)}`,
},
sourceReport: "live-report.json",
attempts,
index: resolve(outputDirectory, "index.html"),
};
await writeFile(
resolve(outputDirectory, "evalbook-manifest.json"),
json(manifest),
);
return manifest;
}