Stop wasting API calls: keep the whole project in context

[jwadow]

The problem

Right now, Roo Code constantly re-reads files to "understand" the project or file.
That means:

• Every file access = at least 1 API request, often 2 (reason -> read -> process).
• For providers that charge per request instead of per token, this kills quotas extremely fast. This especially hurts on free plans (Gemini 2.5 Pro, Copilot, etc.).
• In the end, requests are wasted -- and the model still makes dumb mistakes, because it only sees isolated files with zero awareness of the rest of the project.

In practice, users are literally paying (or burning quotas) for unnecessary round-trips.

---

The proposal

Introduce an optional "project-in-context" mode:

• Load the entire project into the model’s context (we already have 1M-token windows available).
• Dynamically update changed files in the context, instead of re-reading them every time.
• For small and medium-sized projects, this is completely realistic.

---

Benefits

• Massive reduction in wasted API calls.
• Saves money and quotas for users on request-limited providers.
• The model gains a holistic view of the codebase -> fewer dumb mistakes.
• Competitive edge: many modern tools are already moving toward full-context reasoning.

---

Why it matters

This is not just a personal annoyance - it’s a structural inefficiency.
As long as Roo Code keeps "forgetting" the project and re-reading files, everyone pays extra for worse results.
We have models with massive context windows today. Let’s use them properly.

---

Suggestion

Keep the current approach as the default, but allow an opt-in persistent context mode for users who want it.
That way, no one loses, and those who want efficiency + accuracy can get it.