Composer 2 either hallucinating or sharing from other clients

Discussions
,
1

Has anyone noticed composer 2 fast model, either severely hallucinating, or possibly bringing in knowledge/data from other clients?

It started trying to fix an issue as follows. The problem is that this is nothing that I have ever used in any project. There is no instance of “governor”, “GLOBAL_MAX_BROWSER_CONTEXTS”, “_scrape_targets”, or any other concepts below in my entire source tree over all projects, and there never has been.

This is not to do with any of my projects. It’s either hallucinated, or it has brought in data from another person’s project.

Has anyone else experienced this?

Adding a brief "Governor vs parallelism" note — GLOBAL_MAX_BROWSER_CONTEXTS=2 can cap concurrent work if we switch to per-task governor acquires (possible cause of "only 2 tabs").

### Governor (`global_max_browser_contexts`) vs parallel HTTP

Today, `_scrape_targets`-style functions **`await governor.acquire()` once per batch** then run several concurrent `scrape_target_via_node_*` calls **without** each acquiring the governor — so **`GLOBAL_MAX_BROWSER_CONTEXTS=2` does not cap in-flight Node POSTs within one job** (only one permit is held for the whole batch).

If implementation is refactored to **acquire the governor inside each concurrent leg** (one permit per in-flight Node scrape), then **`global_max_browser_contexts` must be ≥ the max concurrent telemetry writes you want** (e.g. raise it or set it ≥ `default_job_max_concurrency` for that scope). Otherwise you will see **at most `GLOBAL_MAX` concurrent “tabs”** regardless of `job_max_concurrency`. Call this out when wiring so “only 2” is not mistaken for a DDB key bug.
3

Hey, thanks for the report. This is a known issue with Composer models. Sometimes the model hallucinates content that looks like someone else’s project, but it’s actually generating plausible, fully made-up code. This isn’t a data leak from another user. The names, commands, and configs in the model output don’t match anyone’s real codebase.

The team is aware and tracking it. For now, the workaround is to start a new chat if you see this behavior. Switching to a non-Composer model like Claude or GPT can also help.

If it happens again, please share the Request ID right corner of the chat > Copy Request ID. That’ll help us track how often this happens.