@love-moon/ai-manager

Manage local AI CLI tools (codex, claude, kimi, copilot) for conductor. Runs inside the daemon process so it can read local credential files, query the Keychain, and make outbound requests over the host's network/VPN.

Features

• Install status — detect whether codex, claude, and kimi are on PATH, and whether the bundled Copilot SDK is available
• Network reachability — probe chatgpt.com, api.anthropic.com, api.kimi.com, api.githubcopilot.com (VPN sanity check)
• Quota — pull usage for all tools via authenticated probes against each provider's own rate-limit channel; results are cached on disk with a TTL
• Codex account switching — list configured ~/.codex/auth.json profiles and atomically swap between them

Config

Reads from ~/.conductor/config.yaml:

ai_manager:
  codex:
    auth_json:
      - /abs/path/auth_accountA.json
      - /abs/path/auth_accountB.json

The codex.auth_json list controls account switching. Quota and install detection require nothing in the conductor config.

How quota is fetched

Each provider exposes usage on a different surface. There is no stable local command, so we hit the same channel the official CLI uses.

| Tool | Source | Auth | | ------ | -------------------------------------------------------------------- | ----------------------------------------------------------------------------------------------------------- | | codex | POST chatgpt.com/backend-api/codex/responses rate-limit headers | tokens.access_token from ~/.codex/auth.json; model resolved from ~/.codex/config.toml | | claude | POST api.anthropic.com/v1/messages rate-limit headers | ANTHROPIC_API_KEY env, then macOS Keychain entry Claude Code-credentials, then ~/.claude/.credentials.json | | kimi | GET api.kimi.com/coding/v1/usages | ~/.kimi/credentials/kimi-code.json; expired access tokens are refreshed automatically and persisted back | | copilot | @github/copilot-sdk RPC account.getQuota | SDK logged-in auth discovery: stored Copilot login or gh auth; explicit tokens are only used when passed through API options |

For codex the body stream is aborted as soon as headers arrive, so the probe costs at most one token of quota. Quota responses are cached at ~/.conductor/cache/ai-manager/quota-<tool>-<fp>.json with a 60s TTL by default; pass forceRefresh: true to bypass.

API

import { AiManager } from "@love-moon/ai-manager";

const m = new AiManager();

// 1) install
await m.checkInstall("codex");
await m.checkInstallAll(); // { codex, claude, kimi, copilot }

// 2) network
await m.checkNetwork("kimi");
await m.checkNetworkAll();

// 3) quota
await m.getCodexQuota();
await m.getClaudeQuota();
await m.getKimiQuota();
await m.getCopilotQuota();
//   { fiveHour, weekly, source: 'fresh'|'cached'|'stale'|'unknown', ... }

// 4) codex accounts
await m.listCodexAccounts();          // [{ name, email, planType, isCurrent }, ...]
await m.getCurrentCodexAccount();
await m.switchCodexAccount("accountB"); // atomic rename + 0o600, backs up to auth.json.bak

Lower-level functions (checkInstall, getKimiQuota, loadAiManagerConfig, parseAuthFile, …) are exported individually if you want to bypass the aggregator.

Identity / safety

• Codex account identity is email|account_id from the JWT, not a token prefix — OpenAI access tokens share a long header prefix and would otherwise collide.
• switchCodexAccount writes to auth.json.tmp and renames over ~/.codex/auth.json so partial writes can never produce a corrupt file. The previous content is copied to auth.json.bak first.
• A switch does not affect codex processes that are already running; they hold the old token in memory until they restart.
• Quota probes never log raw bearer tokens.

Testing

pnpm install
pnpm test       # node --test, ~14 unit tests
pnpm typecheck
pnpm build