fix(context-compressor): clamp out-of-range tail-cut and summary-scan bounds (#75588)

[Ahmett101]

Summary

A short conversation whose protected head / tool-group alignment reached the
end of the message list previously made _find_tail_cut_by_tokens() return
len(messages) + 1. compress() then passed that value into
_find_latest_context_summary() / _find_context_summaries(), which indexed
messages[idx] without clamping and raised IndexError, escaping the
compression path and failing the active gateway turn.

Three defensive bounds:

1. _find_tail_cut_by_tokens() early-returns len(messages) when
   head_end >= len(messages) so the caller's existing
   compress_start >= compress_end no-op path fires.
2. _find_context_summaries() clamps start/end to [0, len(messages)]
   and tolerates non-dict rows.
3. _find_latest_context_summary() inherits the same bounds via the helper.

Changes

• agent/context_compressor.py: three minimal bounds guards.
• tests/agent/test_context_compressor_short_transcript_tail_index_75588.py:
  11 new regression tests pinning the post-fix invariants.

How to Test

PYTHONHASHSEED=0 ~/.hermes/hermes-agent/venv/bin/python -m pytest   tests/agent/test_context_compressor_short_transcript_tail_index_75588.py -v
# ✅ 11/11 passed

PYTHONHASHSEED=0 ~/.hermes/hermes-agent/venv/bin/python -m pytest   tests/agent/test_context_compressor.py   tests/agent/test_compressor_assistant_tail_anchor.py
# ✅ 114/114 passed (no regressions in adjacent compressor suites)

Checklist

• Tests pass — 11/11 new + 114/114 existing
• Follows Conventional Commits
• Changes scoped to this fix only
• Cross-platform impact assessed (Linux / macOS / WSL2 / Windows / Termux)
  — compressor is platform-neutral Python; no path/signal/process
  primitives touched.
• profile-safe paths used — no paths touched.
• .env not used for non-credential settings — no config touched.

Risk & Impact

Low. Three bounds guards inside an already-defensive compression path; no
behaviour change for any transcript that previously succeeded. The
head_end >= n early-return is strictly more conservative than the prior
max(cut_idx, head_end + 1) floor — it returns n instead of n + 1,
which is exactly the value the existing compress_start >= compress_end
guard expects.

Closes #75588

[alt-glitch]

This was generated by AI during triage.

Duplicate of #75604: both current patches repair the same short tool-tail out-of-range boundary that causes the #75588 IndexError.

[teknium1]

Thanks for the focused defensive fix. The helper-level premise remains valid on current main: agent/context_compressor.py:4940 can return len(messages) + 1, while agent/context_compressor.py:4172-4174 scans and indexes without bounding the range. The proposed guards address both.

Suggested changes

• Add a regression for the remaining current-main public caller: ContextCompressor.has_content_to_compress() at agent/context_compressor.py:4953-4955, which the gateway checks at gateway/slash_commands.py:3975. For the reported fully protected shape it should return False. The new compress() assertions at tests/agent/test_context_compressor_short_transcript_tail_index_75588.py:115 already pass through the existing short-transcript return at agent/context_compressor.py:5028-5049, before the changed tail-cut path runs.

This is an automated hermes-sweeper review.

[teknium1]

On current main this call returns through compress()'s existing short-transcript guard at agent/context_compressor.py:5028-5049, before _find_tail_cut_by_tokens() executes. Please add a has_content_to_compress(messages) is False assertion for this shape, since the gateway uses that preflight at gateway/slash_commands.py:3975 and it exercises the changed helper path.