Fix auto compression for tool-heavy sessions

[franksong2702]

Summary

Long single-turn, tool-heavy sessions could repeatedly auto-compress without materially reducing context. Completed tool results in the protected tail could remain large, internal context-compaction markers could be treated like real user anchors, and repeated no-op compression could continue until the provider rejected the oversized context.

This PR makes auto compression fail explicitly when it is exhausted and demotes completed tool results to concise evidence summaries, including completed tool results at the current tail before the next LLM call.

Changes

• Demotes completed tool result bodies to concise evidence summaries even when they sit in the current protected tail.
• Keeps provider-legal assistant/tool pairing while replacing only completed raw tool bodies.
• Skips context-compaction reference markers when finding the latest real user tail anchor.
• Tracks repeated ineffective compression attempts and marks them as compression_exhausted.
• Returns an explicit failed/partial/compression_exhausted result during preflight when compression leaves messages/tokens effectively unchanged.
• Adds regression tests for tool-heavy tail summarization, marker-anchor handling, ineffective compression exhaustion, and preflight no-op exhaustion.

Verification

• python -m pytest tests/test_auto_compression_tool_heavy_tail.py tests/run_agent/test_1630_context_overflow_loop.py tests/run_agent/test_compression_trigger_excludes_reasoning.py tests/test_ctx_halving_fix.py tests/run_agent/test_compression_persistence.py tests/run_agent/test_compressor_fallback_update.py tests/run_agent/test_compression_boundary.py -q
• git diff --check

Related

• Related to Auto compression can exhaust context in tool-heavy sessions #36624.
• Companion WebUI issue: Streaming should show compression-exhausted tool-tail runs as errors nesquena/hermes-webui#3315.
• Companion WebUI PR: Fix compression-exhausted stream finalization nesquena/hermes-webui#3316.

Risks / Follow-ups

• The branch is currently behind upstream main by one commit; no rebase was performed before publishing to avoid changing the already-validated branch state.
• Tool result summarization is intentionally limited to completed role=tool messages; running tools have no result row to summarize.

Model Used

AI-assisted implementation with OpenAI GPT-5 Codex in a local coding workflow. The assistant inspected repository code, wrote regression tests, implemented the fix, and ran the verification commands above.

[teknium1]

Merged via #56165 — #56165 (commit ee710db13, your authorship preserved via rebase).

Salvaged the core fix from this PR: _find_last_user_message_idx now skips context-summary handoff banners (via the existing _is_context_summary_content helper), so _ensure_last_user_message_in_tail no longer anchors the tail to a role="user" summary and rolls the real user turn into the next compaction. Nice catch on the active-task-loss path — verified live on current main.

The PR's other two changes were dropped as superseded by work that landed after you opened this:

• Demoting completed tool results inside the protected tail → covered by the min_tail floor _MAX_TAIL_MESSAGE_FLOOR=8 (fix(agent): preserve recent turns during compression #39170) and no-op compression counting at compress_start >= compress_end ([Bug]: Infinite Context Compaction Loop (messages=N->N) on low context_length / limit configurations #40803).
• Preflight compression_exhausted result → main already returns compression_exhausted on the 413 path and an explicit "auto-compaction disabled" terminal error.

Thanks for the fix and the clear repro in #36624!