Summary

Adds a deterministic pre-drop handoff for the summary-failure fallback path in ContextCompressor.

When the LLM summary cannot be generated but Hermes still compacts a middle window, the fallback summary now includes bounded, redacted source context from the exact removed turns instead of only a generic “summary unavailable” marker.

Problem

Failed summary compression currently preserves the fact that context was lost, but not the actionable state needed to continue safely.

That matters in long-running gateway sessions: durable memory and recent tail context may still exist, but the removed middle window can contain the active workflow trigger, tool state, file/path references, or the last concrete user ask. A generic fallback marker tells the next model that something was dropped, but gives it little evidence for correct continuation.

Skipping a failed compression attempt is safest when the full message list can still be retained. This patch handles the lossy fallback case where Hermes is already compacting a middle window and needs a non-LLM continuity receipt from the exact dropped span.

Changes

• agent/context_compressor.py
  • adds a bounded deterministic fallback-handoff builder;
  • captures recent user requests, tool-call names, tool/result text, file/path mentions, and last dropped turns;
  • runs existing secret redaction before fallback text enters the compacted context;
  • appends the handoff only on the failed-summary fallback path.
• tests/agent/test_context_compressor.py
  • strengthens the existing summary-failure regression to require an actionable fallback handoff;
  • covers tool-call/path preservation;
  • covers secret redaction;
  • covers bounded/truncated fallback output.

Relationship to existing compression fixes

Related to #25585 and the fallback-marker/context-loss reports such as #16670.

This is complementary to PRs like #26051 / #25588 that preserve the original message list when failed compression can be skipped safely. That is the preferred behavior when the full context can still be retained.

This PR covers the remaining hard fallback case: if Hermes is going to replace a middle window anyway, the replacement should contain deterministic source evidence from the dropped span, not just a generic marker.

Invariant this moves toward:

No dropped context window without either a real summary, a native compaction artifact, preserved original messages, or a deterministic pre-drop handoff.

Validation

python -m pytest -o addopts='' tests/agent/test_context_compressor.py::TestSummaryFailureTrackingForGatewayWarning -q
# 4 passed

python -m pytest -o addopts='' tests/agent/test_context_compressor.py -q
# 82 passed

python -m py_compile agent/context_compressor.py tests/agent/test_context_compressor.py
git diff --check

Type of Change

• 🐛 Bug fix (non-breaking change that fixes an issue)
• ✅ Tests (adding or improving test coverage)
• ✨ New feature (non-breaking change that adds functionality)
• 🔒 Security fix
• 📝 Documentation update
• ♻️ Refactor (no behavior change)
• 🎯 New skill (bundled or hub)

Checklist

Code

• My commit messages follow Conventional Commits
• I searched for existing PRs to make sure this is not a duplicate; this is scoped as the hard-fallback handoff layer rather than the fail-closed preserve-original layer
• My PR contains only changes related to this fix
• I added tests for the change
• I tested on Linux/WSL

Documentation & Housekeeping

• Documentation update N/A — behavior is covered by tests and inline comments
• cli-config.yaml.example N/A — no config keys changed
• Tool descriptions/schemas N/A — no tool schemas changed