[M3 Bug] M3 stops doing its job suddenly

[arthurrogado]

Hi,

I'm a paying subscriber and I'm running into a consistent behavior issue with the MiniMax M3 model that effectively blocks me from using it for real coding work.

The pattern: When I give M3 a multi-step task (e.g. "add X to state, refactor Y, update UI for Z, auto-apply discount from W"), the model:

1. Outlines a full plan ("Now I will: 1… 2… 3… 4…")
2. Executes only the first 1–2 steps
3. Stops with a "concluding with 2 steps ..." checkpoint
4. Says something like "Now I will rewrite X to load Y:" and then… stops
5. Requires a new prompt to continue

Environment: VS Code Copilot chat, MiniMax M3 selected as the model. Tasks are already broken into small, well-scoped units (I'm using a spec-driven workflow, not asking for 10 things at once).

What I expect: Given an in-scope plan, M3 should execute it end-to-end and only return when done — same behavior I get from other models, like V4 Flash, that iterates until the task is done.

What's attached: Screenshot showing the plan, the partial completion, and the cutoff mid-sentence at the next step's announcement.

Questions

• Is this a known regression in M3 vs M2.x for direct coding tasks?
• Is there a recommended prompt pattern (or system instruction) that prevents the model from checkpoint-narrating between steps?
• When is this expected to be addressed?

Happy to share more examples, my prompt history, or run a controlled test if it helps your team reproduce.
Thanks.