CopilotClient.start() is not single-flight — concurrent auto-start spawns duplicate CLI processes

[DonJayamanne]

TL;DR

Several concurrent first calls on one CopilotClient can each start a separate
Copilot CLI process. The client only tracks the last process handle
(this.cliProcess), so stop() cannot terminate the earlier ones — a duplicate
cold-start cost plus a real process leak. The fix makes start() single-flight:
all concurrent callers await one in-progress startup. A three-call reproduction
then spawns one CLI (instead of three) and leaves zero orphans after stop().

Impact

• One logical startup can pay for multiple CLI boots, native-addon loads, and
  settings reads (CPU/memory/disk contention).
• Orphaned CLI child processes survive stop(), and temp/session resources they
  own can remain in use.
• Triggered without any explicit start() call — e.g. firing several
  createSession() calls with Promise.all() on a fresh client.

Root cause

• start() returns early only when the state is already "connected". The
  "connecting" state has no stored in-flight promise, so a second concurrent
  caller re-enters the whole start body.
• createSession() and resumeSession() auto-start when !this.connection, so
  the race happens even when the app never calls start() directly.
• startCLIServer() assigns this.cliProcess = spawn(...); a second spawn
  overwrites it and orphans the first, which stop() never terminates.

Reproduction

const client = new CopilotClient(); // autoStart defaults to true
await Promise.all([
  client.createSession({ onPermissionRequest: approveAll }),
  client.createSession({ onPermissionRequest: approveAll }),
  client.createSession({ onPermissionRequest: approveAll }),
]);
// Expected: 1 CLI process; after stop(), 0 remain.
// Actual:   up to 3 CLI processes; after stop(), the extras are orphaned.

Proposed fix

• Add a private startPromise. In start(): if connected, return; if a start is
  in flight, return that same promise; otherwise run the start body once and
  clear startPromise in a finally so a failed start can be retried.
• Move the existing start body into a private doStart(); startup order is
  unchanged.

Acceptance criteria

• Concurrent createSession/resumeSession/start() on a fresh client spawn
  exactly one CLI process.
• After stop(), zero CLI processes remain.
• A failed start can be retried by a later start().
• Regression tests covering the concurrent-start and retry-after-failure cases.

[github-actions]

Investigation

I reviewed nodejs/src/client.ts, specifically start() (lines ~933-992), startCLIServer() (~2593), and stop()/forceStop() (~1159, ~1290).

Findings confirm the bug as described:

• start() only short-circuits when this.state === "connected" (line 934). When a second caller invokes start() (directly, or transitively via createSession()/resumeSession()) while the first call is still in the "connecting" state, there is no stored in-flight promise to await — the second caller falls straight through and re-executes the entire start body, including startCLIServer().
• startCLIServer() assigns the spawned child process to this.cliProcess = spawn(...) (lines ~2666/2673). Since this is a single field (not a set/array), each concurrent invocation overwrites the previous handle.
• stop()/forceStop() only reference this.cliProcess (lines ~1159-1160, ~1290-1296), so any overwritten (earlier) process handles become unreachable and are never terminated — a real orphaned child process leak, plus wasted CPU/memory/disk from redundant CLI boots.
• createSession()/resumeSession() auto-start when !this.connection, so this race is reachable without the caller ever calling start() explicitly (as shown in the issue's Promise.all() repro).

Conclusion: This is a genuine concurrency bug, not working-as-designed. The root cause and proposed fix in the issue are accurate — start() needs a single-flight guard (store the in-flight promise, return it for concurrent callers, clear it in a finally to allow retry after failure).

I confirmed the linked PR #2561 implements exactly this fix. Labeling as bug.

Generated by Bug Handler for #2560 · copilot · auto · 20.2 AIC · ⌖ 5.73 AIC · ⊞ 7.6K · ◷