TUI context usage can exceed 100% (percent falls back to a hardcoded 64k window when the model window can't be resolved)

[arelchan]

Summary

The TUI context-usage indicator can read over 100% (e.g. ~150%), and messages still send fine after it does. Root cause: the percentage is computed against a denominator that silently falls back to a hardcoded 64k default whenever the model's real context window can't be resolved — so once usage passes 64k but stays under the model's true (larger) window, the bar exceeds 100% while requests keep succeeding.

Where

raven/agent/loop/main.py:1447-1456:

context_max = resolve_context_window(call_model) or self.context_window_tokens
context_used = prompt_tokens + completion_tokens
usage_sink["context_percent"] = round(100 * context_used / context_max) if context_max else 0

Same fallback pattern in raven/tui_rpc/methods/session.py:111-116 (context_max = config...context_window_tokens, replaced by the live window only if resolvable).

self.context_window_tokens defaults to 65_536 (raven/config/schema.py:252). So when resolve_context_window(model) returns None, context_max becomes 64k. If the model's real window is larger (128k / 200k / 1M), a conversation between 64k and the real window renders as >100% yet is still well within the model's limit — hence "150% but still sends".

Why it's intermittent (and how to trigger)

resolve_context_window -> _try_litellm_context_window reads litellm's model map (raven/token_wise/pricing.py:166-201). It returns None when:

• the model isn't in litellm's cost/window map (version-dependent — e.g. deepseek/deepseek-v4-flash was "isn't mapped yet" in the fix: Model request failed on Windows #119 environment, though it resolves to 1M here), or
• litellm's remote cost-map fetch fails and the local backup doesn't cover the model (the fix: Model request failed on Windows #119 log shows Failed to fetch remote model cost map ... SSL handshake timed out. Falling back to local backup.).

On this machine resolution currently succeeds (claude-sonnet-4-5 -> 200000), so it does not repro here — but the fallback path makes the percentage meaningless whenever resolution misses.

Note: the per-turn emit uses call_model (the model of that specific call), so a sub-call on a cheaper/unmapped model can momentarily set a wrong context_max.

Expected

A usage percentage should never exceed 100%, and should not be computed against a guessed denominator presented as authoritative.

Suggested fix

• When resolve_context_window returns None, don't render a percentage against the 64k guess. Instead show used-tokens with the window marked unknown/estimated (or hide the bar), so the number never implies "over budget" when it isn't.
• Optionally clamp/annotate when context_used > context_max, and allow a per-model context_window_tokens override for models litellm can't map.

Environment

macOS TUI. Reported: bar once showed ~150% while sending continued to work.