Adds one-time warnings when context usage crosses critical thresholds:
- 80%: suggests /compress or /new if responses degrade
- 95%: warns of imminent errors/truncation, suggests /new
Each threshold fires at most once per session to avoid spam.
Warnings show actual token counts and percentage. Suppressed for
subagents (delegate_depth > 0) where the user can't act on them.
Always shown in CLI mode regardless of quiet_mode setting.
Inspired by OpenCode PR #152 (context window warning).
Bug fix found during live testing:
- Anthropic prompt caching reports input tokens across three fields
(input_tokens, cache_read_input_tokens, cache_creation_input_tokens).
The existing code only counted input_tokens, causing the context
compressor to see ~0 tokens when caching was active. Fixed by summing
all three fields. This also fixes context % display in the status bar
for Anthropic users.
Changes:
- agent/context_compressor.py: add check_context_warning() with
_warned_80/_warned_95 state tracking
- run_agent.py: call check_context_warning() after each API response,
fix Anthropic cached token counting
- tests/test_context_warning.py: 8 tests covering thresholds,
one-shot behavior, escalation, edge cases
Live tested with:
- Nous Portal (chat_completions mode) ✔
- Anthropic direct (anthropic_messages mode) ✔
- Interactive CLI session ✔
885c5dc5e6