Summary
GitHub Copilot CLI currently caps Claude Opus 4.6 at 200K tokens of context, despite the model natively supporting 1M tokens. This is an 80% reduction in available context that causes frequent automatic compaction (summarization) during deep technical sessions.
Problem
When working on complex projects that involve reading large specs, multiple deliverables, and conducting multi-turn analysis, the 200K limit is hit quickly. Automatic compaction then discards context that was expensive to build (both in time and premium requests), forcing re-reads and losing nuance.
Example session: Reading a ~3,200-line protocol spec, ~56KB of research deliverables, a ~50KB wave plan, plus web searches and sub-agent outputs. This is well within the model's 1M capability but exceeds the 200K Copilot budget.
Request
- Preferred: Allow users to configure the context budget (e.g.,
copilot --context-budget 500k or a setting in ~/.copilot/config.json). Users who need deep sessions can opt into higher token usage and accept the cost/latency tradeoff.
- Minimum: Expose the current context budget in
/context output alongside a note explaining the cap vs model capability, so users understand the constraint.
Environment
- Copilot CLI version: 1.0.49-1
- Model: Claude Opus 4.6 (claude-opus-4.6)
- OS: Windows 11
/context reports 200K token budget
References
Summary
GitHub Copilot CLI currently caps Claude Opus 4.6 at 200K tokens of context, despite the model natively supporting 1M tokens. This is an 80% reduction in available context that causes frequent automatic compaction (summarization) during deep technical sessions.
Problem
When working on complex projects that involve reading large specs, multiple deliverables, and conducting multi-turn analysis, the 200K limit is hit quickly. Automatic compaction then discards context that was expensive to build (both in time and premium requests), forcing re-reads and losing nuance.
Example session: Reading a ~3,200-line protocol spec, ~56KB of research deliverables, a ~50KB wave plan, plus web searches and sub-agent outputs. This is well within the model's 1M capability but exceeds the 200K Copilot budget.
Request
copilot --context-budget 500kor a setting in~/.copilot/config.json). Users who need deep sessions can opt into higher token usage and accept the cost/latency tradeoff./contextoutput alongside a note explaining the cap vs model capability, so users understand the constraint.Environment
/contextreports 200K token budgetReferences