[Bug] Context compaction missing

[CMZCoder]

Self check

• I'm on the latest version and searched existing issues (incl. closed) — no duplicate.

Environment

WSL2 via Codex model provider switch

What happened?

Agent Cow doesn't deterministically inherit upstream tools and practices, like such an iimportant model compaction mechanism, resulting in model crashing due to context window fillup.

Logs

[zhayujie]

Hi @CMZCoder, thanks for the report.

Context compaction is actually supported in CowAgent. When the conversation history approaches the model's context window, the agent automatically performs smart compression: it trims and summarizes older turns with an LLM, injects the summary back into the retained context so key information is preserved, truncates oversized tool outputs, and falls back to aggressive trimming (and, if necessary, a reset) when the API reports a context-overflow error. You can tune this via agent_max_context_tokens and agent_max_context_turns in the config.

In addition, the latest version now supports manual compaction via the /compact command, so you can summarize older turns and free up context on demand at any point in a conversation. There's also /clear to fully reset the current context.

Please upgrade to the latest version and let us know if you still run into any context issues. We'll keep improving token estimation and context-window detection for custom model providers going forward. Appreciate the feedback!

command docs: https://docs.cowagent.ai/cli/general#compact