Skip to content

Add general guidance to the orchestrator's own instructions on how to discover and call nested instance tools, and clarify when direct action (vs. handing off to the coder) is appropriate #774

Description

@proagentstore

PROBLEM — This is not a PAGS platform bug; it's a gap in the operating instructions given to an orchestrator (e.g. a Claude project) that drives a PAGS instance through this MCP connector. Two distinct gaps surfaced in the same session:

  1. NO GENERAL GUIDANCE ON NESTED TOOL DISCOVERY — The instructions never state the discover-then-call sequence for an instance's own nested tools (list_instance_tools to get exact names/schemas, then call_instance_tool with those exact names/field names). Without that stated explicitly, the orchestrator guessed tool names and field names repeatedly (e.g. trying issue_number before discovering the real field is number) before landing on a working call. This is a generic sequencing rule that applies to every nested tool, not a fix for one specific case.

  2. UNCLEAR WHEN DIRECT ACTION IS APPROPRIATE VS. HANDING OFF TO THE CODER — The orchestrator's instructions describe an issue-driven workflow ("when we agree on work, it becomes an issue, and the issue goes to the coder"), which was read as implying that even simple, safe, read-only-adjacent actions like filing a GitHub issue must be routed through a full coding run. In fact call_instance_tool can file issues, read issues, comment, etc. directly — no run needed. The instructions describe the philosophy of issue-driven work, but don't clarify that simple direct actions (via call_instance_tool) are appropriate and preferred over spending a coding run when the action doesn't require code changes.

ASK — Update the orchestrator-facing instructions/guidance (not PAGS platform code) to state: (a) always discover a nested tool's exact schema via list_instance_tools before calling it via call_instance_tool, never guess field names from memory or a similar tool's shape; (b) simple direct GitHub/instance actions (reading, filing, commenting on issues) should be done directly via call_instance_tool rather than dispatched as a coding run, reserving coding_loop_start/start_work for work that actually requires code changes or repository access.

NOTE — This may be purely a documentation/instructions change outside this repo (e.g. a project's own system prompt), rather than a platform code change. If so, the actionable output of this issue may just be recommended instruction text rather than a code diff — please confirm scope before starting implementation work.

Activity

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Metadata

Metadata

Assignees

No one assigned

    Labels

    documentationImprovements or additions to documentationenhancementNew feature or requestmcpMCP server surface — the operator's instrument panel, not the product

    Type

    No type

    Projects

    No projects

      Milestone

      No milestone

      Relationships

      None yet

      Development

      No branches or pull requests

      Issue actions