Skills dont work, agents derail a lot

[ZoRDoK]

just dont work:

The user is right - the work was done by the current model (me), not by specialized subagents configured in .pi/skills. The ce_subagent calls I made were to 01-brainstorm and 02-plan skills, but those are actually the same agent's skills running in the same process. Let me check if there's a proper subagent configuration.

And deepseek-v4-pro (f8) dont create full plan after 02 skill.

[leing2021]

Hi @ZoRDoK, thanks a lot for reporting this — and sorry for the late reply.

You were right that the current subagent behavior was confusing and too easy to misuse.

After looking into it, the root problem is that ce_subagent / ce_parallel_subagent were being used to invoke full pipeline-stage skills such as 01-brainstorm and 02-plan. Those stages are interactive workflow stages: they rely on context, artifacts, handoffs, user confirmation, and sometimes multiple turns. Running them inside a one-shot subagent process is not reliable and can lead to incomplete plans or agent derailment.

I’ve made a minimal fix following the “do not add what is unnecessary” principle:

• ce_subagent now rejects pipeline-stage skills:
  • 01-brainstorm
  • 02-plan
  • 03-work
  • 04-review
  • 05-learn
• ce_parallel_subagent has the same guard.
• 03-work now documents that inline execution is the default path.
• CE subagents are now explicitly scoped to bounded, non-interactive, easily verifiable leaf tasks.
• README / README_CN now clarify that Super Pi is not a general-purpose multi-agent executor. Its core purpose is workflow discipline: staged execution, durable artifacts, TDD gates, checkpoints, review, and learning loops.

Tests are passing:

212 pass
0 fail

This should prevent the specific failure mode where an agent tries to run 02-plan inside a subagent and produces an incomplete plan.

I’ll publish this as 0.23.8.

Thanks again for the feedback — it helped clarify an important boundary in the design.