Skip to content

Route task decomposition through the plan tracker - #30

Merged
Luhaozhu merged 1 commit into
mainfrom
feat/plan-aware-task-decomposition
Jul 26, 2026
Merged

Route task decomposition through the plan tracker#30
Luhaozhu merged 1 commit into
mainfrom
feat/plan-aware-task-decomposition

Conversation

@Luhaozhu

Copy link
Copy Markdown
Contributor

Summary

Follow-up to #29. The default system prompt's task-decomposition rule predated the update_plan plan tracker and still forced the model to write a text checklist into the reply body — so complex tasks could end up with both a body checklist and the plan bar showing the same list.

  • System prompt workflow §3 now routes decomposition through update_plan when the tool is available: the checklist lives in the plan bar above the chat input and the reply body no longer repeats it. Contexts without the tool (automations, channels, sub-agents) keep the original in-reply checklist behavior.
  • Pure-text compound questions (e.g. "summarize A and B, then compare") no longer force a checklist; instead the answer must cover every requested item in its own section.
  • The update_plan prompt section gains the matching "don't repeat the checklist in the body" instruction, closing the loop from both sides.
  • The deliverable-completeness rule — every requested item must have its own section in the produced document/slides/spreadsheet — is unchanged.

Testing

  • Multi-deliverable task (Excel + chart): plan bar checklist, no duplicated text checklist in the reply.
  • Pure-text compare/summarize question: no plan bar, answer covers all requested items in sections.
  • Manual plan mode still generates a trailing verify-and-fix step.

- System prompt workflow §3: multi-step execution tasks (multiple tool
  calls / multiple deliverables) now surface their checklist via the
  update_plan tool — rendered in the plan bar above the chat input —
  instead of duplicating a text checklist in the reply body; contexts
  without the tool keep the original in-reply checklist behavior
- Pure-text compound questions no longer force a checklist, but the
  answer must cover every requested item in its own section
- update_plan prompt section now tells the model not to repeat the
  checklist in the reply body, closing the loop with the system prompt
- Deliverable-completeness rule (every item gets its own section) is
  unchanged
@Luhaozhu
Luhaozhu merged commit 5d66c29 into main Jul 26, 2026
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant