uncaged/workflow
Archived
10
0
Fork 0
Code Issues 6 Pull Requests Actions Packages Projects Releases Wiki Activity

feat: generate agentic summary in __end__ node via ReAct LLM call #187

New Issue
Closed
opened 2026-05-11 09:59:06 +00:00 by xiaoju · 0 comments
xiaoju commented 2026-05-11 09:59:06 +00:00
Owner
Copy Link
Copy Source

Background

Currently end node summary is hardcoded as 'completed: moderator returned END', which is useless for downstream consumers (especially workflow-as-agent, where the parent thread needs a meaningful result).

Proposal

Keep end as a pseudo-role (no RoleDefinition change). In the engine finalize path, before writing the end node, make a single ReAct LLM call to generate a summary:

  1. Moderator returns END
  2. Engine gathers context: original prompt + all step results (role outputs from the DAG)
  3. One-shot ReAct LLM call -> produces a natural-language summary that answers the original prompt
  4. Summary stored in end node payload.content (ContentMerkleNode)
  5. Thread finalized as usual

What doesn't change

  • Moderator END semantics unchanged -- still means terminate
  • No new RoleDefinition needed for end
  • No schema/extract for the end node
  • Existing workflow definitions unaffected

Benefits

  • workflow-as-agent can read result.summary directly instead of traversing child DAG
  • Extract from workflow output becomes trivial (summary is already a coherent answer)
  • Reuses engine existing LLM provider (same one used for extract)

Key files

  • packages/workflow-engine/src/engine.ts -- finalize path, end node creation
  • packages/workflow-agent-hermes/src/workflow-as-agent.ts -- consumes result.summary
## Background Currently __end__ node summary is hardcoded as 'completed: moderator returned END', which is useless for downstream consumers (especially workflow-as-agent, where the parent thread needs a meaningful result). ## Proposal Keep __end__ as a pseudo-role (no RoleDefinition change). In the engine finalize path, before writing the end node, make a single ReAct LLM call to generate a summary: 1. Moderator returns END 2. Engine gathers context: original prompt + all step results (role outputs from the DAG) 3. One-shot ReAct LLM call -> produces a natural-language summary that answers the original prompt 4. Summary stored in __end__ node payload.content (ContentMerkleNode) 5. Thread finalized as usual ## What doesn't change - Moderator END semantics unchanged -- still means terminate - No new RoleDefinition needed for __end__ - No schema/extract for the end node - Existing workflow definitions unaffected ## Benefits - workflow-as-agent can read result.summary directly instead of traversing child DAG - Extract from workflow output becomes trivial (summary is already a coherent answer) - Reuses engine existing LLM provider (same one used for extract) ## Key files - packages/workflow-engine/src/engine.ts -- finalize path, end node creation - packages/workflow-agent-hermes/src/workflow-as-agent.ts -- consumes result.summary
This repo is archived. You cannot comment on issues.
No Branch/Tag Specified
Branches Tags
Labels
No items
No Label
Milestone
No items
No Milestone
Projects
Clear projects
No project
Assignees
Clear assignees
aobing jiashuang (Jia Shuang) jiayi (Jiayi) luming scottwei tuanzi xiaoju xiaomo xiaonuo xingyue
No Assignees
1 Participants
Due Date
No due date set.
Dependencies

No dependencies set.

Reference: uncaged/workflow#187
Delete Branch "%!s()"

Deleting a branch is permanent. Although the deleted branch may continue to exist for a short time before it actually gets removed, it CANNOT be undone in most cases. Continue?