Context length issue / better context inspector

[geril07]

There is an example of pretty simple session.json
There is a 1 call saying hello to the agent inside the opencode sources. And even with that, the context length is already 10k.
If I dump the entire session.json file into the OpenRouter chat, context length will be 5k, half of it.

Are there any hidden data sent to the provider/model which is not tracked inside session file?

[github-actions]

This issue might be a duplicate of existing issues. Please check:

• [Question] detailed per-interaction telemetry (token counts and input composition) for opencode-model communications #2666: [Question] detailed per-interaction telemetry (token counts and input composition) for opencode-model communications - discusses detailed token tracking and visibility into what's sent to the model
• [Feature Request] Configuration to Limit Input Tokens per Request #2637: [Feature Request] Configuration to Limit Input Tokens per Request - specifically addresses context length management and control over token consumption

Feel free to ignore if none of these address your specific case.

[rekram1-node]

@geril07 what is likely causing extra context that isn't necessarily in the session.json is the tool definitions (along with mcp server tools if you have any).

opencode ships with a variety of tools by default, each one has a JSON input schema, a name, a description, and then often a description for the different fields that can be passed to it.

Due to that these do require additional tokens / cost and wouldn't be present when interacting just directly with llm without any tools. So that probably explains the context difference.

[geril07]

Yeah, that's logical. It would be nice to have an option to inspect the entire context, which was requested in #2666.
Closing as a duplicate then, thanks.