During long-running tasks such as project analysis, Droid CLI keeps auto-injecting continuation steps like "Continue." while the task is still in progress.
With Cloudflare-backed models, these continuation steps seem to interrupt the model's current reasoning/execution flow. Instead of continuing from the current state, the model restarts its analysis, repeats planning, and loses continuity.
This issue appears mainly in long tasks. Shorter interactions are much less affected.
The same workflow works correctly with non-Cloudflare models, which suggests this is a Cloudflare-specific compatibility issue in Droid's long-task continuation / session handling rather than a general prompt or workflow problem.
Observed pattern:
- Start a long project analysis task
- Droid progressively inspects files and keeps moving forward
- Droid auto-injects "Continue."
- The model loses continuity and starts analyzing from scratch again
Expected:
The model should continue the current task naturally, or fail with a clear error.
Actual:
Droid auto-continues mid-task, and with Cloudflare models this often resets the model's behavior instead of extending it.
During long-running tasks such as project analysis, Droid CLI keeps auto-injecting continuation steps like "Continue." while the task is still in progress.
With Cloudflare-backed models, these continuation steps seem to interrupt the model's current reasoning/execution flow. Instead of continuing from the current state, the model restarts its analysis, repeats planning, and loses continuity.
This issue appears mainly in long tasks. Shorter interactions are much less affected.
The same workflow works correctly with non-Cloudflare models, which suggests this is a Cloudflare-specific compatibility issue in Droid's long-task continuation / session handling rather than a general prompt or workflow problem.
Observed pattern:
Expected:
The model should continue the current task naturally, or fail with a clear error.
Actual:
Droid auto-continues mid-task, and with Cloudflare models this often resets the model's behavior instead of extending it.