agentscope-ai / agentscope-ai/ReMe

fix: make metadata-backed step answers complete for LLM consumers

Open
#379 0 comments 0 reactions 1 assignee Claimed by @ployts View on GitHub
bug
Dominant language
Python
Stars
3.4k
Forks
298
Avg merge
19h 52m
Merged PRs (30d)
55

Description

## Context

ReMe MCP tools are designed for direct LLM consumption and intentionally expose Response.answer. Response.metadata remains auxiliary request data for programmatic clients and diagnostics.

Some Steps store information required for the next LLM action only in metadata, leaving answer incomplete. For example, list returns only a count in answer while file paths are available only in metadata, so an agent cannot select a file for a subsequent read call.

This issue follows up on #376 at the Step output layer rather than changing the MCP transport contract.

## Scope

Audit and, where needed, improve the LLM-facing answer produced by these Steps:

- list: include the returned workspace-relative file paths, not only the count.
- search: keep result paths, source text, and relevant link expansion self-contained; return an explicit message when there are no results.
- daily_list: ensure the date, count, note paths, and relevant frontmatter are clear in the answer.
- traverse: ensure the traversal summary and actionable edge details are present in the answer.
- node_search: ensure hit paths and the fields needed for follow-up reads are present in the answer.
- proactive: include the topics and requested content, rather than only a read summary; keep skipped and error outcomes explicit.

Do not copy metadata wholesale into answer. Format only information needed by an LLM to understand the result and perform the next tool call.

If other Steps have the same problem, please add concrete examples to this issue so they can be included in the audit.

## Acceptance criteria

- Each affected answer is self-contained and actionable for direct LLM consumers.
- Existing structured metadata remains available to HTTP and other programmatic consumers.
- MCPService continues to return Response.answer directly.
- Empty, skipped, and error outcomes remain concise and unambiguous.
- Focused unit tests assert the relevant LLM-visible answer content for each changed Step.

Contributor guide

No contributing guide indexed for this repository

Assessment

This issue has not been assessed yet.

Get new issues in your inbox

A short digest of beginner-friendly GitHub issues.