What changed

This PR makes context compaction (conversation compression) more robust when summary generation fails due to misconfigured auxiliary model routing.

1. Retry summary generation using the main provider/model

   • If the configured summary model/provider fails (e.g. google/* model routed to a Codex endpoint), the compressor retries with:
     • provider="main"
     • model=self.model (the main conversation model)

2. Never fail silently when compaction cannot produce a summary

   • If summary generation still returns None, we insert an explicit compaction warning summary marker (with the standard compaction prefix), explaining that earlier turns may have been dropped without a summary and how to fix configuration.

Why

We observed failures where context compression attempted to use an unsupported summarization model for the active endpoint (e.g. google/gemini-3-flash-preview on a Codex ChatGPT account). The summary call failed, a session split occurred due to “compression”, and continuity degraded (inconsistent answers / lost context).

This PR both:

• recovers automatically in common misconfiguration cases, and
• warns explicitly when compaction cannot summarize.

User impact / behavior

• In the common failure mode (auxiliary summary model incompatible with the active endpoint), Hermes will now still generate a compaction summary using the main model instead of failing compaction silently.
• If summarization still fails, the transcript includes a clear warning marker that explains the continuity risks and points to config knobs (compression.summary_* and auxiliary.compression.provider=main).

How to reproduce (example)

1. Configure context compression to use a model that the active endpoint cannot serve (e.g. compression.summary_model: google/gemini-3-flash-preview while using a Codex endpoint).
2. Generate enough turns/tool output to trigger context compaction.
3. Observe summary generation failure and degraded continuity.

How this fixes it

• On summary failure, compaction performs a second attempt using provider=main + the main model, avoiding the incompatible auxiliary route.
• If the fallback also fails, compaction still injects a warning summary marker to make the failure explicit to both users and future assistants.

Tests

• Added unit tests covering:
  • retry behavior when the first summary call fails and the second succeeds (provider=main)
  • warning marker insertion when summary generation returns None

Command run locally:

• PYTHONPATH="$PWD" pytest -q tests/agent/test_context_compressor.py

Files changed

• agent/context_compressor.py
• tests/agent/test_context_compressor.py