Skip to content

contextTier=long_context is not applied on startup / no CLI flag for long context tier #3481

@lossyrob

Description

@lossyrob

Describe the bug

The CLI accepts contextTier: "long_context" in ~/.copilot/settings.json, but new/default sessions launched non-interactively do not reliably start with the long context tier. The session starts with the smaller/default context even when settings specify model, effortLevel, and contextTier.

Selecting the same model interactively through /model switches the session from gpt-5.5 (xhigh) to gpt-5.5 (xhigh) (1.1M context), which suggests the context tier is recognized by the model picker but not applied during startup/model resolution.

This blocks autonomous or automatically launched sessions that start with an initial prompt, because those sessions cannot interactively choose the long-context tier after launch.

Affected version

GitHub Copilot CLI 1.0.52-4

Observed installed package: 1.0.52-4 on Windows.

Steps to reproduce the behavior

  1. Configure ~/.copilot/settings.json with:

    {
      "model": "gpt-5.5",
      "effortLevel": "xhigh",
      "contextTier": "long_context"
    }
  2. Start a new default Copilot CLI session, or launch an automated session with an initial prompt.

  3. Run /context or inspect the displayed model/context information.

  4. Observe that the session uses the smaller/default context rather than the long-context tier.

  5. In an interactive session, run /model and select the same model/tier. The display changes to the long-context variant, for example gpt-5.5 (xhigh) (1.1M context).

Expected behavior

A new session should honor contextTier: "long_context" from settings.json during startup/model resolution. For automation, there should also be a launch flag such as:

copilot --model gpt-5.5 --reasoning-effort xhigh --context-tier long_context

or equivalent.

Additional context

I inspected the installed 1.0.52-4 bundle locally and found:

  • contextTier is schema-recognized with values "default" and "long_context".
  • The interactive model picker has a context_tier phase and writes contextTier when selecting the long-context tier.
  • The documented CLI flags include --model and --reasoning-effort, but not a context-tier flag.
  • The startup/model-resolution path appears to read model and effortLevel, while the UI path applies context-tier capability overrides later.
  • The plain /model <id> command path appears to clear contextTier, while the picker path can write it.

Operating system: Windows 11 / Windows Terminal / PowerShell.

Metadata

Metadata

Assignees

No one assigned

    Labels

    area:configurationConfig files, instruction files, settings, and environment variablesarea:modelsModel selection, availability, switching, rate limits, and model-specific behaviorarea:non-interactiveNon-interactive mode (-p), CI/CD, ACP protocol, and headless automation

    Type

    No fields configured for Bug.

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions