Skip to content

Orbit: agent answers "why did my job fail?" by resuming an inaccessible history (403) and hallucinating a failure explanation #367

Description

@dannon

App: 0.5.1 (linux) · Model: kimi-k2.7-code (openai-compatible). Reported via beta feedback ("Total hallucination — the history does not exist").

What happened

The user asked "why my last galaxy job failed?" The agent worked off the notebook-pages list instead of the user's own jobs:

  1. notebook_list_galaxy_pages → returned a set of Galaxy pages
  2. notebook_resume_from_galaxy(page_id=…, history_id=…) → linked to a history taken from one of those pages
  3. galaxy_history_showtool not found (the agent guessed a tool name that doesn't exist on the server)
  4. galaxy_get_history_contents(history_id=…)403 History is not accessible by user (err_code 403002) — that history isn't owned by the current user/API key

After the 403 it didn't surface the access error; it fabricated an answer about the inaccessible history — hence the reporter's "Total hallucination."

The bugs

  1. Wrong data source for "my last job." "Why did my last job fail?" should resolve against the user's own histories/jobs (get_histories / list_history_ids → recent jobs), not a history pulled from a notebook page that may belong to someone else.
  2. 403 → confident hallucination. A 403 / not-accessible (or a guessed tool that 404s) should be surfaced as "I couldn't access that history," not narrated around.

Suggested direction

  • Ground "my last job/history" queries in the user's accessible histories before answering.
  • Treat a Galaxy 403 / not-accessible as a hard stop that's reported to the user, not something to invent a narrative around.

xref #261 (agent fabricates a tool-failure root cause), #210 (job errors not surfaced proactively).

Metadata

Metadata

Assignees

No one assigned

    Labels

    bugSomething isn't working

    Type

    No type

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions