fix(compressor): restore summary_model and reset fallen_back flag on session reset

[nftpoetrist]

What does this PR do?

_generate_summary() clears self.summary_model = "" when the configured auxiliary summary model fails (404, 503, timeout, etc.) and falls back to the main model for that compression call. This mutation is permanent for the lifetime of the ContextCompressor instance.

reset_session_state() in run_agent.py calls on_session_reset() on the same compressor instance on every /new and /reset — the compressor is never recreated. on_session_reset() already resets _summary_failure_cooldown_until to 0 so transient errors do not block a fresh session (fix #15548 / salvage #19622). The same logic applies to summary_model: a per-session fallback should not permanently degrade compression quality for every subsequent session.

Without this fix: after one aux model failure in a long session, every /reset silently leaves summary_model = "". The configured model (e.g. aux_models.compressor.model = "gpt-4o-mini") is never used again — even if the aux model recovered — until the gateway is restarted.

Root cause: on_session_reset() did not include summary_model in its reset list, and _summary_model_fallen_back was also left as True, preventing the one-shot fallback from firing again in the new session.

Fix:

1. Store the originally configured value in _configured_summary_model at __init__ time.
2. Restore self.summary_model from it in on_session_reset().
3. Reset _summary_model_fallen_back = False so the fallback logic can fire once more if the model fails again in the new session.

Symmetric with the _summary_failure_cooldown_until = 0.0 reset already present in the same method.

Related Issue

Companion to #15548 (same lifecycle class — on_session_reset() not resetting a compressor field).

Type of Change

• 🐛 Bug fix
• ✨ New feature
• 🔒 Security fix
• 📝 Documentation update
• ✅ Tests
• ♻️ Refactor
• 🎯 New skill

Changes Made

• agent/context_compressor.py: store _configured_summary_model at init (+1 line); restore in on_session_reset() (+2 lines)
• tests/agent/test_context_compressor.py: 3 new regression tests in TestOnSessionResetSummaryModel (+35 lines)

How to Test

python3.11 -m pytest tests/agent/test_context_compressor.py::TestOnSessionResetSummaryModel -v --override-ini="addopts="

Expected: 3 passed.

Manual reproduction:

1. Configure aux_models.compressor.model to a model that will fail (wrong name / offline).
2. Have a long session that triggers compression — the aux model fails, falls back to main.
3. Run /reset.
4. Have another long session that triggers compression.
5. Before fix: main model used (aux model silently absent). After fix: aux model tried again.

Checklist

Code

• Contributing Guide read
• Conventional Commits
• No duplicate PR
• This fix only
• pytest run — 72 passed, 0 failed
• Tests added
• Platform: macOS

Documentation & Housekeeping

• Docs updated — N/A
• cli-config.yaml.example — N/A
• CONTRIBUTING.md/AGENTS.md — N/A
• Cross-platform impact — affects all platforms; no platform-specific code
• Tool descriptions — N/A

[alt-glitch]

Revives the fix from #15549 (closed without merge) and #16152 (closed as duplicate of #15549). The bug is confirmed present in main — on_session_reset() does not restore summary_model or reset _summary_model_fallen_back.

[teknium1]

Thanks for preserving this regression fix. The premise still holds on current main: _fallback_to_main_for_compression() sets _summary_model_fallen_back and clears summary_model at agent/context_compressor.py:1733-1745, while on_session_reset() does not restore either value at agent/context_compressor.py:726-745.

Problems

• The submitted patch no longer applies to current main: git apply --check <(gh pr diff 21783 --repo NousResearch/hermes-agent --patch) fails for both changed files because the compressor and test file have moved substantially.

Suggested changes

• Salvage the same two state resets into the current ContextCompressor layout, preserving the regression tests; no semantic defect was found in the proposed behavior.

Automated hermes-sweeper review.