Provider-aware context budgeting, compaction, and usage visibility

Summary

Rowboat supports gateway models, ChatGPT/Codex subscriptions, API keys, OpenRouter, local models, and Code Mode agents. These routes can have different context limits, pricing thresholds, caching rules, and usage allowances.

For example, some GPT API requests become significantly more expensive above roughly 272K input tokens, while recent Claude models offer flat pricing across their 1M context window. Subscription routes may instead consume included allowance faster.

Problem

Users currently cannot easily tell:

• How much context a request will send
• Whether older history and tool results are compacted
• Whether a pricing or allowance threshold is approaching
• How Auto accounts for context size, cost, and provider route
• Which usage belongs to Rowboat chat versus Code Mode

This can cause context-limit errors and unexpected credit consumption.

Proposal

• Apply context budgets by provider route and model
• Show a context meter for instructions, tools, memory, history, and attachments
• Automatically compact older context while preserving the local transcript
• Provide a manual /compact-style action
• Warn before known pricing or allowance thresholds
• Let users compact, switch models, start a handoff chat, or continue
• Separate gateway, subscription, API-key, and Code Mode usage

Similar tools already provide parts of this: Cursor shows context usage, Codex supports compaction, and Cline reports token cost and cache savings.

Related

#514, #281, #392, #768, #346, #388, and PR #680