openpencil/scripts/ab-corpus/clients/codex-cli.ts
2026-05-03 21:00:00 +08:00

133 lines
5 KiB
TypeScript

/**
* Codex CLI client — drives `codex exec` as a subprocess to reach
* GPT-5.4 (the CLI's default) via the user's Codex Pro subscription.
*
* Trade-offs:
* + No per-token API billing (counts against Codex subscription)
* + Works offline of API key management
* - Each call incurs Codex's own system-prompt overhead (~20k tokens
* of codex agent framing) which the model still sees alongside
* our design-generation system prompt. Interpret A/B deltas
* against GPT-5.4/codex with that caveat; they're not a clean
* "raw GPT-5.4 completion" measurement.
* - `codex exec` cold-start adds a few seconds per call → 48 runs
* at 10-20s each = 10-15 min wall time for full GPT-5.4 pass.
*
* Output shape: we pass `--output-last-message` to write just the
* assistant's final message to disk, then read + delete. Skips the
* CLI's pretty-printed session log (which is noisy and harder to
* parse reliably).
*/
import { spawn } from 'node:child_process';
import { mkdtempSync, readFileSync, rmSync } from 'node:fs';
import { tmpdir } from 'node:os';
import { join } from 'node:path';
import type { ChatCallResult } from './openai-compat';
export interface CallCodexArgs {
model: string;
system: string;
user: string;
/**
* Timeout in ms. Default 10 minutes — composite ab-v3 prompts plus
* Codex's ~20k-token agent framing plus thinking can blow the older
* 5-minute cap (saw this on ab-v4 dashboard-search-filters-composite,
* where gpt-5.4 was mid-stream when codex SIGKILL'd it). Override via
* `AB_CORPUS_CODEX_TIMEOUT_MS` env var if you want to surface
* single-call slowness as a hard fail instead of waiting it out.
*/
timeoutMs?: number;
}
const DEFAULT_CODEX_TIMEOUT_MS = (() => {
const env = process.env.AB_CORPUS_CODEX_TIMEOUT_MS;
if (env && /^\d+$/.test(env)) {
const n = Number(env);
if (n > 0) return n;
}
return 10 * 60 * 1000;
})();
export async function callCodex(args: CallCodexArgs): Promise<ChatCallResult> {
const timeoutMs = args.timeoutMs ?? DEFAULT_CODEX_TIMEOUT_MS;
const dir = mkdtempSync(join(tmpdir(), 'ab-codex-'));
const outPath = join(dir, 'last-message.txt');
// Codex CLI has its own system prompt (coding agent framing) that
// we can't override. Fold OUR system prompt into the user-visible
// prompt so it at least reaches the model. Tag the boundary so the
// model treats the system block as authoritative context, not as
// part of the task to answer.
const prompt = `SYSTEM CONTEXT (treat as authoritative system prompt):\n<<<\n${args.system}\n>>>\n\nUSER REQUEST:\n${args.user}`;
try {
const content = await new Promise<string>((resolve, reject) => {
// Pin reasoning effort to `medium` so ab-corpus stays comparable
// across model bumps. Codex CLI's per-model default isn't fixed:
// gpt-5.4 historically ran ~medium; gpt-5.5 defaults to `xhigh`,
// which sent ab-v5 wall time from ~15 min (ab-v4) to a projected
// ~13 h. Override via `AB_CORPUS_CODEX_REASONING` if you ever
// need to measure full-effort ceiling.
const reasoningEffort = process.env.AB_CORPUS_CODEX_REASONING || 'medium';
const proc = spawn(
'codex',
[
'exec',
'--skip-git-repo-check',
'--ephemeral',
'--sandbox',
'read-only',
'-c',
`model_reasoning_effort=${reasoningEffort}`,
'-m',
args.model,
'--output-last-message',
outPath,
prompt,
],
{ stdio: ['ignore', 'pipe', 'pipe'] },
);
let stderr = '';
proc.stderr.on('data', (chunk) => {
stderr += chunk.toString();
});
const killTimer = setTimeout(() => {
proc.kill('SIGKILL');
reject(new Error(`codex timed out after ${timeoutMs}ms`));
}, timeoutMs);
proc.on('error', (err) => {
clearTimeout(killTimer);
reject(err);
});
proc.on('close', (code) => {
clearTimeout(killTimer);
if (code !== 0) {
reject(new Error(`codex exit ${code}: ${stderr.slice(0, 300)}`));
return;
}
try {
const out = readFileSync(outPath, 'utf-8');
if (out.trim().length === 0) {
reject(new Error('codex produced empty output-last-message'));
return;
}
resolve(out);
} catch (err) {
reject(err instanceof Error ? err : new Error(String(err)));
}
});
});
// Codex CLI doesn't surface token usage in --output-last-message
// mode (the streaming JSONL session log has it, but parsing that
// adds fragility for marginal value). Report 0/0 — the harness
// aggregator skips zero rows, so codex columns show '—' instead
// of looking falsely cheap.
return { content, usage: { promptTokens: 0, completionTokens: 0 } };
} finally {
try {
rmSync(dir, { recursive: true, force: true });
} catch {
// Best-effort cleanup
}
}
}