Agent should prevent sending prompts that exceed model context window

[bhunt-krs]

Summary

Zed's agent panel does not enforce the model's context window limit before sending requests. When a conversation accumulates more tokens than the model supports, Zed sends the full payload anyway, the API rejects it, and the user gets an error. This wastes time and creates a frustrating retry loop.

Current behavior

With Amazon Bedrock and allow_extended_context: true, the effective limit is 1M tokens. In long agent sessions, Zed repeatedly sends prompts exceeding this limit (1.5M+ tokens against a 1M maximum), resulting in errors like:

ERROR [agent::thread] Turn execution failed: invalid request format to Amazon Bedrock's API:
The model returned the following errors: {"type":"error","error":{"type":"invalid_request_error",
"message":"prompt is too long: 1581464 tokens > 1000000 maximum"}}

Zed retries the same oversized prompt multiple times in a row, each time getting the same rejection. The user has no signal that the context is too large until the error appears, and no recourse other than manually starting a new thread.

A secondary issue: the "generate summary" feature uses a separate model call with a 64K token limit and fails silently on long threads (model_max_prompt_tokens_exceeded). This is tracked in #37025.

Expected behavior

Zed should:

1. Refuse to send (or warn before sending) when the estimated token count exceeds max_token_count() for the selected model
2. Auto-truncate or summarize old context (tool outputs, file contents, earlier conversation turns) when approaching the limit, similar to how Claude Code auto-compacts conversation history
3. Show a visible token counter in the agent panel so users can see how close they are to the limit before hitting it
4. Not retry the same oversized prompt multiple times — if the API returns a token limit error, do not send the same payload again

Related issues

• allow_extended_context does not update UI context window limit (still shows 200k) #49617 — allow_extended_context does not update UI context window limit (still shows 200k). The gauge is inaccurate, which makes it impossible for users to self-manage context size.

Environment

• Zed: v0.233.9 stable
• Provider: Amazon Bedrock (SSO auth, allow_extended_context: true)
• Models: Claude Opus 4.6, Claude Opus 4.7
• OS: macOS (Apple Silicon)