Skip to content

Gemini 4 Neon and Hydrogen: Usage records report medium for xhigh task requests #5518

Backlog

Description

@MattPD

Describe the bug

Copilot CLI 1.0.114 lists gemini-4-neon and gemini-4-hydrogen under xhigh in its task tool schema (reasoning_effort extras). However, two observations show xhigh unavailable or reported as medium:

  • In the /model picker's interactive reasoning-effort switcher for gemini-4-neon, high is the highest available option (xhigh is absent).
  • In four synchronous explore sub-agent runs (one xhigh request and one high control per model), three relevant records from each xhigh run's event chain reported different reasoningEffort values: "reasoningEffort": "medium" in subagent.configured, "reasoningEffort": "xhigh" in session.model_change, and "reasoningEffort": "medium" in session.usage_record. Neither xhigh run logged a warning event. In both high control runs, all three records reported "reasoningEffort": "high".

Google documents an xhigh thinking level for these models (labeled Extra in the Antigravity model selector): the Gemini 4 Neon announcement, the Gemini 4 Hydrogen announcement, and the Antigravity model documentation.

Related issue: #3823 reports a silent downgrade from xhigh to medium for models not advertising xhigh in CAPI capabilities. Here, gemini-4-neon and gemini-4-hydrogen appear under xhigh in the CLI's own task tool schema. These records do not establish whether CAPI advertises xhigh for either model or which effort Copilot sent to the provider.

Affected version

GitHub Copilot CLI 1.0.114

Steps to reproduce the behavior

  1. Interactive picker check: In Copilot CLI 1.0.114, open /model and highlight gemini-4-neon, then step through its reasoning-effort levels with the left and right arrow keys. A separate interactive check for gemini-4-neon showed high as the highest available option in the picker, and xhigh was absent.
  2. Sub-agent check: In a Copilot CLI 1.0.114 session already running gemini-4-neon at long_context (the recorded parent session state; fresh-launch setup for long_context was not separately verified), prompt Copilot to dispatch four synchronous explore sub-agents via the task tool (asking each to reply with OK without calling tools), giving each a distinct name and omitting context_tier on both gemini-4-hydrogen calls:
    • name: "probe-neon-xhigh", model: "gemini-4-neon", reasoning_effort: "xhigh", context_tier: "long_context"
    • name: "probe-neon-high", model: "gemini-4-neon", reasoning_effort: "high", context_tier: "long_context"
    • name: "probe-hydrogen-xhigh", model: "gemini-4-hydrogen", reasoning_effort: "xhigh" (omit context_tier; inherited long_context from the parent session)
    • name: "probe-hydrogen-high", model: "gemini-4-hydrogen", reasoning_effort: "high" (omit context_tier; inherited long_context from the parent session)
  3. Obtain the session ID with /session id and open ~/.copilot/session-state/<session-id>/events.jsonl (under $COPILOT_HOME when set). Use agentId to link each run's subagent.started event to three records: subagent.configured, session.model_change, and session.usage_record. Confirm "contextTier": "long_context" in each subagent.configured record before comparing results.

Observed field excerpts from the four runs:

probe-neon-xhigh (gemini-4-neon, requested xhigh):
  subagent.configured:  {"model": "gemini-4-neon", "reasoningEffort": "medium", "contextTier": "long_context"}
  session.model_change: {"source": "agent", "contextTier": "long_context", "newModel": "gemini-4-neon", "reasoningEffort": "xhigh"}
  session.usage_record: {"model": "gemini-4-neon", "responseModel": "gemini-4-neon", "initiator": "sub-agent", "reasoningEffort": "medium", "maxPromptTokens": 983040}

probe-neon-high (gemini-4-neon, requested high):
  subagent.configured:  {"model": "gemini-4-neon", "reasoningEffort": "high", "contextTier": "long_context"}
  session.model_change: {"source": "agent", "contextTier": "long_context", "newModel": "gemini-4-neon", "reasoningEffort": "high"}
  session.usage_record: {"model": "gemini-4-neon", "responseModel": "gemini-4-neon", "initiator": "sub-agent", "reasoningEffort": "high", "maxPromptTokens": 983040}

probe-hydrogen-xhigh (gemini-4-hydrogen, requested xhigh):
  subagent.configured:  {"model": "gemini-4-hydrogen", "reasoningEffort": "medium", "contextTier": "long_context"}
  session.model_change: {"source": "agent", "contextTier": "long_context", "newModel": "gemini-4-hydrogen", "reasoningEffort": "xhigh"}
  session.usage_record: {"model": "gemini-4-hydrogen", "responseModel": "gemini-4-hydrogen", "initiator": "sub-agent", "reasoningEffort": "medium", "maxPromptTokens": 983040}

probe-hydrogen-high (gemini-4-hydrogen, requested high):
  subagent.configured:  {"model": "gemini-4-hydrogen", "reasoningEffort": "high", "contextTier": "long_context"}
  session.model_change: {"source": "agent", "contextTier": "long_context", "newModel": "gemini-4-hydrogen", "reasoningEffort": "high"}
  session.usage_record: {"model": "gemini-4-hydrogen", "responseModel": "gemini-4-hydrogen", "initiator": "sub-agent", "reasoningEffort": "high", "maxPromptTokens": 983040}

Expected behavior

For gemini-4-neon, if Copilot CLI supports interactive xhigh, the /model picker's interactive reasoning-effort switcher should offer xhigh.

For each model (gemini-4-neon and gemini-4-hydrogen), if task supports xhigh, Copilot CLI should honor and record it across three points: request xhigh from the provider, log "reasoningEffort": "xhigh" in subagent.configured, and report "reasoningEffort": "xhigh" in session.usage_record when the provider applies that level.

If task does not support xhigh for a model, the task tool schema should omit that model from xhigh. In that case, Copilot CLI should reject the request or warn the caller.

Additional context

  • Operating system: Red Hat Enterprise Linux 9.6 (Linux 5.14.0-570.19.1.el9_6.x86_64)
  • CPU architecture: x86_64
  • Verbatim task tool reasoning_effort extras on 1.0.114: xhigh='claude-fable-5','claude-fable-5.1','claude-fable-5.2','claude-haiku-5.5','claude-opus-4.8-fast','claude-opus-5.5','claude-sonnet-5','claude-sonnet-5.5','gemini-4-hydrogen','gemini-4-neon','gpt-5.3-codex','gpt-5.6-luna','gpt-5.6-sol','gpt-6-astra','gpt-6-luna','gpt-6-sol','gpt-6.1-sol','gpt-6.1-terra','gpt-6.2-astra','gpt-6.2-sol','grok-4.6','grok-4.7','mai-code-2','mai-code-2-flash'
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

    Type

    No type

    Projects

    No projects

      Milestone

      No milestone

      Relationships

      None yet

      Development

      No branches or pull requests

      Issue actions