Back to Oh My Openagent

Senpi task category model status QA

.omo/evidence/20260722-senpi-task-model-status/README.md

4.19.12.1 KB
Original Source

Senpi task category model status QA

What was tested

  • Failing-first renderer proof for a resolved ultrabrain category.
  • Planner propagation of the registry model's friendly name.
  • Registry-boundary rejection of inherited/accessor/control-bearing/oversized model names.
  • Backward-compatible provider deduplication for persisted human-readable displays.
  • Plain task result grammar and the real width-aware task result component.
  • The full senpi-task package suite, Senpi adapter gate, package/root type checks, and the pinned repository build.
  • A real xterm.js browser-rendered 72-column terminal surface.
  • The isolated live Senpi task lifecycle probe for spawn, batch, category error, isolation, and process cleanup.

What was observed

Before: task category:ultrabrain (GPT-5.6 Sol reasoning:xhigh) ...

After, plain result: task category:ultrabrain (openai GPT-5.6 Sol reasoning:xhigh) ...

After, 72-column component: task category:ultrabrain (openai GPT-5.6 Sol xhigh) background pending

The compact row keeps all requested identity at 72 columns by rendering the reasoning level as xhigh; the expanded/plain result retains the explicit reasoning:xhigh label.

Why this is enough

The failing tests pin both data plumbing and output grammar: the planner must preserve the friendly name, and both plain and width-aware renderers must combine it with provider and reasoning. Package-wide and adapter-wide gates cover neighboring state, persistence, sanitization, error, and task-tool behavior. The xterm artifact proves the exact user-visible terminal row at the width used by the existing task renderer tests.

The broad live Senpi probe's three follow-up/output-sequencing failures were reproduced on unmodified origin/dev@499b52358; see live-baseline.txt.

What was omitted

  • Raw Senpi logs and environment dumps were not copied because they may contain private session data.
  • The broad live lifecycle probe's three unrelated follow-up/output-sequencing failures are summarized in verification.txt; they are not presented as a pass for those behaviors.
  • Unrelated generated Codex/codegraph diffs from the repository build are not part of this change.