Repository navigation
Description
CLI version
2.7.2
Skill version (optional)
No response
Node version
v26.13.0
OS
win32 x64, Windows 11 build 28120, invoked from PowerShell 7
Region
cn
Command to reproduce
bl usage token-plan --output jsonExpected behavior
the process exits as soon as the command has answered; the JSON document is on stdout at that point.
Actual behavior
the document appears in about 0.8–1.2 s, then the process stays alive for about 50–56 s before exiting with code 0. Anything that spawns bl and waits for the exit pays that wait: a plugin or an agent tool that reads the quota ends up with a ~50 s child instead of a ~1 s one, and the only workaround outside the CLI is to kill it once the document is complete.
Full output
No error: exit code 0, the document is correct. Timings, with the process allowed to finish:
| form | output complete | process exit |
|---|---|---|
release binary (bl-2.7.2-windows-x64.exe) |
784 ms | 54.7 s |
release binary, NO_UPDATE_NOTIFIER=1 |
786 ms | 56.1 s |
release binary, ~/.bailian/update-state.json set to "checked just now" |
2009 ms | 54.4 s |
npm entry (node node_modules/bailian-cli/dist/bailian.mjs) |
3961 ms | 4.3 s |
npm entry, NO_UPDATE_NOTIFIER=1 |
886 ms | 886 ms |
bl --version (no network call) |
— | 691 ms |
JSON error (optional)
Already tried
- bl update and skill version aligned with CLI
- bl auth status OK for this command
- Different network / region — still reproduces
How often does this happen?
Always
How was bl invoked?
Terminal (manual)
Notes (optional)
Where the exit waits. packages/core/src/client/transport.ts sends the release binary's requests through globalThis.fetch, and the npm package's through undici's global dispatcher. Both keep the connection pooled. undici's keep-alive times out in about 4 s, which is what the npm form's 4.3 s shows; the release binary's pool holds the process instead.
It is not the version check: the binary takes ~54 s with a fresh and an aged ~/.bailian/update-state.json, and with NO_UPDATE_NOTIFIER=1 as well. That flag is what the npm form responds to (4.3 s → 0.9 s, i.e. it only skips the check's own FETCH_TIMEOUT_MS = 3000).
packages/cli/src/main.ts ends with void createCli(...).run();, and process.exit is reached only on errors (runtime/src/error-handler.ts), on SIGINT after flushTelemetry(500) (runtime/src/create-cli.ts:113) and on EPIPE (:118). A successful run therefore waits for every pooled socket before the event loop drains.
Possible fix. Exit after the run resolves, the way the SIGINT path already does:
const code = await createCli({ ...commands, ...commandAliases }, { /* … */ }).run();
process.stdout.write("", () => process.exit(code));For the node path, getGlobalDispatcher().close() would have the same effect; the release binary's fetch has no close, so the explicit exit is the option there. Either way the document is already on the wire, so nothing is lost.
Separate ask: the version check should follow npm's registry configuration. packages/runtime/src/utils/update-checker.ts hardcodes NPM_REGISTRY = "https://registry.npmjs.org". On a machine whose npm is pointed at a mirror or a private proxy (registry=… in .npmrc, or the npm_config_registry environment variable npm exports), the check still calls npmjs directly: here it sits behind a proxy for the whole FETCH_TIMEOUT_MS on every run that is due, and it can never see the mirror's dist-tags, so a mirror-only environment gets a slow check and no update notice.
Suggestion: resolve the registry the way npm does — an explicit npm_config_registry (or NPM_CONFIG_REGISTRY) first, then the nearest .npmrc registry=, then https://registry.npmjs.org — and let the same environment variable win for the standalone binary, which has no npm around it. That also makes the check testable against a mirror instead of the public registry.
Activity