Skip to content

harness: execute AtlasCloud plan-delegate task tool calls #3935

Description

@asim

Root cause:

  • Harness (E2E) run https://github.com/micro/go-micro/actions/runs/28708394971 failed in the Provider harnesses (live LLM conformance) job on master at 95e7c387b2160cc320d512c2a6d5961dc3a6405e.
  • Only AtlasCloud was configured (ATLASCLOUD_MODEL=minimaxai/minimax-m3); other live providers were skipped for missing API keys.
  • In atlascloud / plan-delegate (plan persistence + delegation + tool call), the model response emitted raw MiniMax-tagged tool-call markup for three task_TaskService_Create invocations instead of producing executable tool calls that the agent loop consumed. The notification delegation did run, but no tasks were created, so the harness failed with unexpected side effects: tasks=0 notify=1.
  • This was not a provider outage/rate-limit/network timeout: the live model returned a response, the harness continued, and the failure was a deterministic side-effect assertion.

Where:

  • The plan-delegate harness path drives the AtlasCloud/minimax-m3 agent through task creation and notification delegation.
  • Failure signature: atlascloud / plan-delegate (plan persistence + delegation + tool call): unexpected side effects: tasks=0 notify=1, with the conductor reply containing MiniMax tool markup for task_TaskService_Create calls.

Acceptance criteria:

  • The AtlasCloud/minimax-m3 plan-delegate path converts/executes the model's task creation tool calls so Design, Build, and Ship tasks are actually created before notification.
  • The stored plan still records completed steps for checking/creating the tasks and delegating notification.
  • The plan-delegate harness observes tasks=3 notify=1 (or the harness's expected successful side effects) and exits successfully.
  • The provider-conformance harness reports atlascloud / plan-delegate as passed in the next live Harness (E2E) run.

Metadata

Metadata

Assignees

No one assigned

    Type

    No type

    Fields

    No fields configured for issues without a type.

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions