Files
windmill/ai_evals/core/runSuite.ts
centdix b0ddcf31e4 ci: add path-gated AI agent + ai_evals smoke workflows (#9640)
* ci: add path-gated AI agent integration tests workflow

Runs integration_tests/ai_agent_tests against real LLM providers
(Anthropic/OpenAI/Google) only when AI-agent backend code or the tests
change, since runs make paid LLM calls. Adds a conftest fixture that
skips provider-parametrized cases whose API keys are absent, so CI
exercises only the providers it has secrets for.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

* ci: add path-gated ai_evals global-mode smoke workflow

Runs the global AI chat eval (global-test1) across one cheap model per
provider (Anthropic/OpenAI/Google/DeepSeek) only when the eval harness or
copilot chat code change, since runs make paid LLM calls. Builds Windmill
CE from source as the AI proxy; global tools/drafts run in the Vitest
bridge. Gates on the deterministic draft pipeline (run succeeded +
produced a draft + used write_script), not the variable LLM judge score.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

* ci: run AI smokes on PR ready-for-review instead of every push

Switch the pull_request trigger from `synchronize` (every commit) to
`ready_for_review`, with a job guard skipping draft PRs, so the paid LLM
runs only fire when a PR is marked ready to merge (plus push-to-main and
manual dispatch).

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

* fix(ai_evals): lazily load cli mode so non-cli evals skip the cli toolchain

The entrypoint eagerly imported modes/cli, which pulls the wmill CLI
guidance modules and their JSR deps (@cliffy/*). Global/flow/script/app
runs then crashed with "Cannot find module '@cliffy/ansi/colors'" when
the cli workspace deps were not installed. Import createCliModeRunner
dynamically inside runCliBenchmark instead.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

* test(ai_agent): raise low max_completion_tokens to OpenAI's 16 minimum

OpenAI's /v1/responses rejects max_output_tokens < 16 with a 400, failing
test_low_max_tokens for openai. 16 still exercises a truncated response.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

* ci: run ai_evals workflow on Node 22 for the frontend undici 8.x dep

The Vitest bridge loads frontend/node_modules/undici@8.x, which requires
Node >=22.19; Node 20 failed with "webidl.util.markAsUncloneable is not a
function" when loading vitest.config.ts.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

* fix(ai_evals): run frontend evals autonomously + give global-test1 more turns

Frontend evals (flow/script/app/global) ran the production chat prompt, which
assumes an interactive human — so cheaper models burned their turn budget
asking for confirmation, waiting for approval, or presenting a plan, sometimes
hitting maxTurns without producing a draft. Append a shared autonomy note in
baseEvalRunner (the path all frontend modes share, mirroring cli mode): act
directly on clear requests; only ask on genuinely ambiguous ones (preserving
the askUserQuestion cases). Also raise global-test1's maxTurns 8 -> 10 so a
model that over-explores still converges.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

* ci(ai_evals): watch draft/prompt deps outside copilot/

The global eval runs production frontend code in-process, so the smoke's
behavior depends on files outside frontend/src/lib/components/copilot/**:
the draft model (userDraft.svelte.ts, userDraftDbSyncer.svelte.ts), script
inference (infer.ts), and the chat system prompts ($system_prompts ->
system_prompts/auto-generated). Add them to both push and PR path filters so
a change there actually triggers the smoke that gates on draft production.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

* fix: skip direct provider tests without credentials

* feat: add ai evals skip judge flag

* fix: simplify ai evals ci gate

* fix: simplify ai evals smoke gate

* fix: handle ai eval workflow triggers

---------

Co-authored-by: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
2026-06-22 12:41:31 +02:00

347 lines
11 KiB
TypeScript

import { judgeOutput, DEFAULT_JUDGE_MODEL } from "./judge";
import type {
BenchmarkAttemptResult,
BenchmarkCaseResult,
BenchmarkCheck,
EvalCase,
FrontendBenchmarkProgressEvent,
ModeRunner,
} from "./types";
import { validateToolExpectations } from "./validators";
export async function runSuite<TInitial, TExpected, TActual>(input: {
modeRunner: ModeRunner<TInitial, TExpected, TActual>;
cases: EvalCase[];
runs: number;
runModel: string | null;
judgeModel?: string | null;
executionOnly?: boolean;
concurrency?: number;
verbose?: boolean;
onProgress?: (event: FrontendBenchmarkProgressEvent) => void;
}): Promise<BenchmarkCaseResult[]> {
const judgeModel =
input.judgeModel === undefined ? DEFAULT_JUDGE_MODEL : input.judgeModel;
const concurrency = Math.max(1, input.concurrency ?? input.modeRunner.concurrency);
const results = new Array<BenchmarkCaseResult>(input.cases.length);
let cursor = 0;
if (input.modeRunner.mode !== "cli") {
input.onProgress?.({
type: "run-start",
surface: input.modeRunner.mode,
totalCases: input.cases.length,
runs: input.runs,
concurrency,
});
}
async function worker(): Promise<void> {
while (true) {
const caseIndex = cursor++;
if (caseIndex >= input.cases.length) {
return;
}
const evalCase = input.cases[caseIndex];
results[caseIndex] = {
id: evalCase.id,
prompt: evalCase.prompt,
initialPath: evalCase.initialPath,
expectedPath: evalCase.expectedPath,
attempts: await runCaseAttempts({
caseIndex,
evalCase,
runs: input.runs,
judgeModel,
judgeThreshold: input.modeRunner.judgeThreshold ?? 80,
executionOnly: input.executionOnly ?? false,
modeRunner: input.modeRunner,
totalCases: input.cases.length,
verbose: input.verbose ?? false,
onProgress: input.onProgress,
}),
};
}
}
await Promise.all(
Array.from({ length: Math.min(concurrency, input.cases.length) }, () => worker())
);
return results;
}
async function runCaseAttempts<TInitial, TExpected, TActual>(input: {
caseIndex: number;
evalCase: EvalCase;
runs: number;
judgeModel: string | null;
judgeThreshold: number;
executionOnly: boolean;
modeRunner: ModeRunner<TInitial, TExpected, TActual>;
totalCases: number;
verbose: boolean;
onProgress?: (event: FrontendBenchmarkProgressEvent) => void;
}): Promise<BenchmarkAttemptResult[]> {
const attempts: BenchmarkAttemptResult[] = [];
const surface = input.modeRunner.mode === "cli" ? null : input.modeRunner.mode;
for (let attempt = 1; attempt <= input.runs; attempt += 1) {
if (surface) {
input.onProgress?.({
type: "attempt-start",
surface,
caseId: input.evalCase.id,
caseNumber: input.caseIndex + 1,
totalCases: input.totalCases,
attempt,
runs: input.runs,
});
}
const startedAt = Date.now();
try {
const initial = await input.modeRunner.loadInitial(input.evalCase.initialPath);
const expected = input.executionOnly
? undefined
: await input.modeRunner.loadExpected(input.evalCase.expectedPath);
const run = await input.modeRunner.run(input.evalCase.prompt, initial, {
evalCase: input.evalCase,
caseId: input.evalCase.id,
caseNumber: input.caseIndex + 1,
totalCases: input.totalCases,
attempt,
runs: input.runs,
verbose: input.verbose,
onAssistantMessageStart: input.verbose && surface
? () =>
input.onProgress?.({
type: "assistant-message-start",
surface,
caseId: input.evalCase.id,
caseNumber: input.caseIndex + 1,
totalCases: input.totalCases,
attempt,
runs: input.runs,
})
: undefined,
onAssistantChunk: input.verbose && surface
? (chunk: string) =>
input.onProgress?.({
type: "assistant-chunk",
surface,
caseId: input.evalCase.id,
caseNumber: input.caseIndex + 1,
totalCases: input.totalCases,
attempt,
runs: input.runs,
chunk,
})
: undefined,
onAssistantMessageEnd: input.verbose && surface
? () =>
input.onProgress?.({
type: "assistant-message-end",
surface,
caseId: input.evalCase.id,
caseNumber: input.caseIndex + 1,
totalCases: input.totalCases,
attempt,
runs: input.runs,
})
: undefined,
onToolCall: input.verbose && surface
? ({ toolName, argumentsText }) =>
input.onProgress?.({
type: "tool-call",
surface,
caseId: input.evalCase.id,
caseNumber: input.caseIndex + 1,
totalCases: input.totalCases,
attempt,
runs: input.runs,
toolName,
argumentsText,
})
: undefined,
});
const checks: BenchmarkCheck[] = [
buildCheck("run succeeded", run.success, run.error),
];
if (!input.executionOnly) {
checks.push(
...input.modeRunner.validate({
evalCase: input.evalCase,
prompt: input.evalCase.prompt,
initial,
expected,
actual: run.actual,
run,
}),
...validateToolExpectations({
run,
toolExpect: input.evalCase.toolExpect,
})
);
}
const artifactFiles = input.modeRunner.buildArtifacts?.(run.actual) ?? [];
if (
run.success &&
!input.executionOnly &&
input.modeRunner.backendValidate
) {
try {
const backendValidation = await input.modeRunner.backendValidate({
evalCase: input.evalCase,
prompt: input.evalCase.prompt,
initial,
expected,
actual: run.actual,
run,
context: {
evalCase: input.evalCase,
caseId: input.evalCase.id,
caseNumber: input.caseIndex + 1,
totalCases: input.totalCases,
attempt,
runs: input.runs,
verbose: input.verbose,
onAssistantMessageStart: undefined,
onAssistantChunk: undefined,
onAssistantMessageEnd: undefined,
},
});
if (backendValidation) {
checks.push(...backendValidation.checks);
artifactFiles.push(...(backendValidation.artifactFiles ?? []));
}
} catch (error) {
checks.push(
buildCheck(
"backend validation succeeded",
false,
error instanceof Error ? error.message : String(error)
)
);
}
}
let judgeScore: number | null = null;
let judgeSummary: string | null = null;
if (
run.success &&
!input.executionOnly &&
input.judgeModel !== null &&
!input.evalCase.skipJudge
) {
const judge = await judgeOutput({
mode: input.modeRunner.mode,
prompt: input.evalCase.prompt,
checklist: input.evalCase.judgeChecklist,
initial,
expected: input.modeRunner.mode === "cli" ? undefined : expected,
actual: input.modeRunner.prepareJudgeActual
? input.modeRunner.prepareJudgeActual(run.actual)
: run.actual,
model: input.judgeModel,
});
judgeScore = judge.success ? judge.score : null;
judgeSummary = judge.summary;
checks.push(buildCheck("judge succeeded", judge.success, judge.error));
checks.push(
buildCheck(
`judge score >= ${input.judgeThreshold}`,
(judgeScore ?? 0) >= input.judgeThreshold,
judge.success ? `score=${judgeScore}` : judge.error
)
);
}
const attemptResult: BenchmarkAttemptResult = {
attempt,
passed: checks.every((check) => check.passed),
durationMs: Date.now() - startedAt,
assistantMessageCount: run.assistantMessageCount,
toolCallCount: run.toolCallCount,
toolsUsed: uniqueStrings(run.toolsUsed),
toolCallDetails: run.toolCallDetails,
skillsInvoked: uniqueStrings(run.skillsInvoked),
checks,
judgeScore,
judgeSummary,
error: run.error ?? null,
tokenUsage: run.tokenUsage ?? null,
finalContextTokens: run.finalContextTokens ?? null,
artifactsPath: null,
artifactFiles,
};
if (surface) {
input.onProgress?.({
type: "attempt-finish",
surface,
caseId: input.evalCase.id,
caseNumber: input.caseIndex + 1,
totalCases: input.totalCases,
attempt,
runs: input.runs,
passed: attemptResult.passed,
durationMs: attemptResult.durationMs,
judgeScore: attemptResult.judgeScore,
error: attemptResult.error,
});
}
attempts.push(attemptResult);
} catch (error) {
const message = error instanceof Error ? error.message : String(error);
const failedAttempt: BenchmarkAttemptResult = {
attempt,
passed: false,
durationMs: Date.now() - startedAt,
assistantMessageCount: 0,
toolCallCount: 0,
toolsUsed: [],
skillsInvoked: [],
checks: [buildCheck("run crashed", false, message)],
judgeScore: null,
judgeSummary: null,
error: message,
tokenUsage: null,
finalContextTokens: null,
};
if (surface) {
input.onProgress?.({
type: "attempt-finish",
surface,
caseId: input.evalCase.id,
caseNumber: input.caseIndex + 1,
totalCases: input.totalCases,
attempt,
runs: input.runs,
passed: false,
durationMs: failedAttempt.durationMs,
judgeScore: null,
error: message,
});
}
attempts.push(failedAttempt);
}
}
return attempts;
}
function buildCheck(name: string, passed: boolean, details?: string): BenchmarkCheck {
return details ? { name, passed, details } : { name, passed };
}
function uniqueStrings(values: string[]): string[] {
return [...new Set(values)];
}