Gateway mode does not have this bug — it evicts the cached agent entirely on /new, so a fresh ContextCompressor is created on the next message with the correct value from the provider.

What happened

The context window size displayed in the TUI status bar comes from agent.context_compressor.context_length, which is populated by calling get_model_context_length() at ContextCompressor construction time (probe + fallback chain). This probes the model metadata for the configured model/provider to determine the actual context length.

In gateway mode, /new calls _evict_cached_agent() which removes the old agent from the cache. The next incoming message creates a brand-new AIAgent with a fresh ContextCompressor, so the new context length is picked up immediately.

In CLI mode, /new calls self.agent.reset_session_state() (cli.py:6288) which resets token counters, session ID, and conversation history — but it does not recreate or reinitialize the context_compressor. The existing agent object is reused, so the stale context_length from the previous session persists.

- Where context length is read: agent/context_compressor.py:544 — ContextCompressor.init calls get_model_context_length()
- Where CLI /new resets state: cli.py:6288 — self.agent.reset_session_state() (resets counters only, does not touch compressor)
- Where gateway correctly handles this: gateway/run.py:9094 — _evict_cached_agent(session_key) evicts the agent entirely

Environment

- Hermes Agent version: 0.14.0
- Platform: CLI (not gateway)
- Provider: LM Studio (local)
- OS: WSL2

python
In cli.py, inside new_session(), after self.agent.reset_session_state():

if hasattr(self.agent, "context_compressor") and self.agent.context_compressor is not None:
    # Re-probe the context length from the provider in case it changed.
    # Create a fresh ContextCompressor with the same parameters.
    from agent.context_compressor import ContextCompressor
    comp = self.agent.context_compressor
    self.agent.context_compressor = ContextCompressor(
        model=self.model,
        threshold_percent=comp.threshold_percent,
        protect_first_n=comp.protect_first_n,
        protect_last_n=comp.protect_last_n,
        summary_target_ratio=comp.summary_target_ratio,
        quiet_mode=True,
        base_url=getattr(self.agent, "base_url", "") or "",
        api_key=getattr(self.agent, "api_key", ""),
        config_context_length=None,  # Let it probe from provider
        provider=getattr(self.agent, "provider", ""),
        api_mode=getattr(self.agent, "api_mode", ""),
    )


Alternative (simpler) approach: add a method like reset_context_length() to the agent class that recreates just the compressor without losing other state.

Reproduction steps

1. Start Hermes CLI with a local provider (e.g., LM Studio at http://127.0.0.1:1234/v1)
2. Note the context window size in the status bar (e.g., 8192 — the default for an unknown model)
3. In LM Studio, change the context length setting for your model to 134400
4. Type /new in Hermes CLI
5. Observe that the status bar still shows the old value (8192) instead of 134400
6. Restart Hermes — the correct value now appears