fix(compression): preserve latest actionable user turn

[John-Lussier]

Summary

Preserve the newest actionable user event across context compaction, including the observed async completion -> blank platform echo -> long tool run shape.

A blank user-role echo could become the last-user tail anchor. The real completion immediately before it then entered the summarized middle, allowing an older task to become active again. A later compaction could also merge a summary into the completion row itself, so the event stopped being actionable.

Fix

• Distinguish actionable user turns from compaction summaries and genuinely empty platform echoes.
• Treat image, audio, input_audio, and unknown structured content as actionable; only empty content shapes are blank.
• Remove a blank echo only when it is contiguous after the latest actionable user event and an assistant turn follows it.
• Preserve the immediately preceding assistant/tool group as a role bridge when necessary, keeping the latest user event as a separate, exact, non-summary row.
• If no older safe window remains, take the existing ineffective-compression path rather than summarizing or rewriting the latest event.

This preserves provider-valid role/tool ordering and existing tool-call/result groups. It also keeps the latest event actionable through a second compaction.

Tests

PYTHONPATH=. HERMES_HOME=/tmp/hermes-p0-compaction-test-home \
  python -m pytest \
  tests/agent/test_compressor_actionable_tail_anchor.py \
  tests/agent/test_context_compressor.py \
  tests/agent/test_compressor_assistant_tail_anchor.py \
  tests/agent/test_compressor_zero_user_guard.py \
  tests/agent/test_context_compressor_summary_continuity.py \
  tests/agent/test_compressed_summary_metadata.py \
  tests/run_agent/test_compression_persistence.py \
  tests/agent/test_resume_stale_active_task.py \
  -q -o 'addopts='

Result: 212 passed.

git diff --check also passes.

Regression coverage includes:

• async completion followed by blank echo and long tool activity;
• exact boundary with no safely compressible older turn;
• second compaction preserving the same event;
• image-only, audio, input_audio, and unknown structured user content;
• tool-call head/tail groups and visible assistant replies;
• truthful no-op compression accounting.

Compatibility

• The summary is never merged into the latest actionable user row.
• Existing user/assistant/tool schemas are unchanged.
• Underscore-prefixed compression metadata behavior is unchanged.
• This is orthogonal to fix(agent): strip fabricated 'User asked' lines from compaction summary (#62365) #62674, which validates fabricated User asked: summary lines; both touch context_compressor.py but address different failure modes.

Scope

Intentionally not included:

• changes to summary prompting or fabricated-request validation;
• generic message-role normalization;
• changes to gateway/delegation delivery;
• unrelated context-budget or tool-result pruning changes.

[teknium1]

Thanks for isolating a real compaction-anchor failure: current main still anchors any non-summary role="user" row in agent/context_compressor.py:2464, so a blank platform echo can displace the actionable completion.

Problems

• agent/context_compressor.py:2285 initializes the blank-echo scan with user_idx + 1. When no actionable user event exists, user_idx is -1, so this scans and removes index 0 if it is blank and followed by an assistant. That is not an echo after a user event and can leave an assistant-leading transcript.
• agent/context_compressor.py:2254 only recognizes structured type="text" blanks. Runtime paths also treat input_text as textual (run_agent.py:4958), so an empty input_text echo still becomes the actionable tail anchor.

Suggested changes

• Return no removable indices when user_idx < 0, with a leading-blank-user regression test.
• Treat empty input_text as blank and cover it in the parameterized echo tests.

Automated hermes-sweeper review.

[teknium1]

When _find_last_user_message_idx() returns -1, this starts at index 0 and can remove a leading blank user row before an assistant response. That row is not contiguous after an actionable event and may be needed to keep a provider-valid user-leading transcript. Please return an empty set when user_idx < 0 and add that regression case.

[teknium1]

Please include input_text in the text-only blank classification. Current runtime handling recognizes it as textual (run_agent.py:4958); otherwise an empty structured input_text echo remains actionable and can still displace the real latest user turn.

[John-Lussier]

Addressed both sweeper findings in aacd1eed2:

• _blank_echo_indices_after() now returns immediately when there is no actionable user (user_idx < 0), with a leading-blank regression so it cannot produce an assistant-leading transcript.
• Empty input_text blocks now use the same blank classification as empty text blocks.

Strict TDD: the two added cases failed before the source change and now pass. Updated focused compression verification: 214 passed; git diff --check is clean.

[John-Lussier]

The review findings were addressed in aacd1eed2 and remain on the current PR head:

• _blank_echo_indices_after() returns immediately when user_idx < 0, with a leading-blank regression so it cannot produce an assistant-leading transcript.
• Empty input_text blocks now use the same blank classification as empty text blocks, covered by parameterized echo tests.

Focused compression verification: 214 passed; git diff --check clean. Ready for re-review.

[teknium1]

Merged via #69291 (commit 2ee50c6). Both your commits were cherry-picked with authorship preserved — cleanest salvage of the batch. Thanks!