Skip to content

fleet nodes advertise spawn capabilities for harness executables unavailable to the service #1819

Description

@kjgbot

Summary

A fleet broker advertised spawn:gemini and accepted Gemini placements even though its service environment could not execute gemini. Failures happened only after dispatch during MCP setup.

Evidence

On finn-mini running 12.3.1, AGENT_RELAY_NODE_HARNESSES contained claude,codex,gemini,opencode. The daemon PATH resolved Claude and Codex but not Gemini. The node nevertheless advertised spawn:gemini.

The broker repeatedly logged:

failed to spawn cloud-issue-audit-06x: failed to configure Agent Relay MCP for gemini: could not run gemini mcp add: No such file or directory (os error 2)

Several issue and PR audit workers failed this way, and the coordinator retried under replacement names.

Impact

The scheduler treats an unusable node as eligible, consumes invocation and identity state, and creates retry/name-collision churn.

Acceptance

Before publishing spawn:<harness>, preflight the configured command in the exact daemon environment. Missing or non-executable commands must withhold the capability and emit one actionable startup diagnostic. Add doctor coverage for commands removed after startup.

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

    bugSomething isn't workingworkflow-readyReady for workflow implementation

    Type

    No type

    Projects

    No projects

      Milestone

      No milestone

      Relationships

      None yet

      Development

      No branches or pull requests

      Issue actions