fix(compression): skip empty post-handoff summary windows (salvage #59526)

[teknium1]

Skips the wasted aux-LLM summary call (and its noisy abort) when the newest handoff summary consumes the entire compressible window — compress() now early-returns the transcript unchanged instead of dispatching _generate_summary on empty input.

Salvages PR #59526 by @WOLIKIMCHENG — cherry-picked onto current main with authorship preserved (fix(compression): skip empty post-handoff summary windows, including the regression test). @ms-alan's #59737 independently identified the same bug and the same correct early-return site — credit to both; #59737 was opened 7h later without a test, so #59526 is the salvage vehicle.

Fixes the #59496 shape.

The bug (verified on main @ c2c2449)

After the post-handoff slice recompute (turns_to_summarize = pre_summary_turns + messages[summary_idx+1:compress_end], ~L4490-4503), an empty turns_to_summarize still reached _generate_summary(...) unconditionally (~L4548). Post-#69297 a standalone handoff strips to None in _window_row, so a window containing only the handoff row genuinely produces an empty summarizer input — burning an aux LLM call that aborts noisily for zero benefit, every turn the shape recurs.

Not covered by the existing guards: the compress_start >= compress_end guard (#40803, ~L4396) fires at an earlier, different site; the semantic no-op check in conversation_compression.py runs after the wasted call.

Adaptations from the original diff (main moved under it)

• Ineffective-counter strike routed through _record_ineffective_compression_verdict() instead of the PR's stale bare _ineffective_compression_count += 1 — fix(compression): persist anti-thrash state across process restarts #69872 made the anti-thrash counter durable, and all strike/clear sites must write through the helper so the persisted copy stays in sync.
• The skip DOES count as an ineffective strike (deliberate choice, documented here): this mirrors the sibling compress_start >= compress_end guard's behavior exactly — the empty-window shape cannot shrink, so without a strike should_compress() keeps saying yes and every subsequent turn re-fires the same no-op loop. That is the precise failure [Bug]: Infinite Context Compaction Loop (messages=N->N) on low context_length / limit configurations #40803's guard exists to stop; this guard inherits the same contract.
• Guard placement: after the region recording / chunk_count telemetry, well below the top-of-function per-call state reset — so the fix(compression): reset in-place compaction state on every abort path #69853 _last_compression_attempt_* per-attempt abort-state semantics are untouched (verified by tests/run_agent/test_compression_abort_state_reset.py, all green).
• _previous_summary rehydration is deliberately kept (not rolled back as the summary-abort path does for fix(agent): decay protected summaries after restart #57835): the rehydrated summary came from a handoff genuinely present in the transcript, which is returned unchanged.
• New telemetry failure_class = "empty_post_handoff_window" distinguishes this skip from no_compressible_window / insufficient_messages in the attempt log.
• Regression test adapted to current main: the original fixture no longer produced an empty window on today's boundary logic (a live probe showed _generate_summary being called and a normal compaction succeeding). The fixture was reshaped so the standalone handoff is the only window row, and the test now also pins the failure_class and the strike-through-helper behavior. Kept in the contributor's commit alongside their guard.

Note: the skip returns the input transcript per the standard caller contract for no-op skips (callers gate post-compaction bookkeeping on result is not input; Phase-1 tool-result pruning may still return a pruned copy, exactly as on the sibling guard paths — the contract is unchanged by this PR).

Validation

Check	Result
Counterfactual: regression test vs pre-fix origin/main source	FAILS (_generate_summary called once on empty window) — bug reproduced
Regression test on this branch	PASS
tests/agent/test_context_compressor*.py + tests/run_agent/test_compression_abort_state_reset.py	246 tests, 0 failed
Full sibling sweep: tests/agent/test_compress*.py + test_context_compressor*.py + abort-state suite (rebased tip)	498 tests, 0 failed
ruff check on changed files	clean
ast.parse after conflict resolution	clean

Attribution

• fix(compression): skip empty post-handoff summary windows — author @WOLIKIMCHENG (kinsonnee@gmail.com), cherry-picked from fix(compression): skip empty post-handoff summary windows #59526 with authorship preserved; window-guard hunk surgically reapplied at its current home (the region was rewritten by fix(compression): dedupe persisted compaction handoffs #69297/fix(compression): decay head protection after restart when a persisted summary exists #69342/fix(compression): persist anti-thrash state across process restarts #69872 since the PR branched).
• Contributor email mapping committed via scripts/add_contributor.py (contributors/emails/kinsonnee@gmail.com).
• @ms-alan (fix(agent): skip compression when latest summary leaves no messages to summarize #59737): independent identification of the same fix — please close with credit when this lands.

Infographic

[github-actions]

૮ >ﻌ< ა ci review

_{ran on 25769d2}

ℹ️ Info

Desktop E2E visual evidence · View test artifacts · View job

1 visual diff.

_{inline evidence upload failed.}

Failed to upload diff-665a0833239e-onboarding-overlay-diff.png with gh image (exit code 1): Error uploading /home/runner/work/_temp/e2e-evidence/diff-665a0833239e-onboarding-overlay-diff.png: step 0 (get upload token): uploadToken not found on repo page — do you have write access to NousResearch/hermes-agent? (or, if NousResearch enforces SAML SSO, authorize at https://github.com/orgs/NousResearch/sso)

[isak-ialogics]

CI triage: the blocking slice-4 failure appears unrelated to this compression diff. The same exact test and assertion (tests/gateway/test_config.py::TestApiServerEnvOverride::test_env_key_does_not_reenable_explicitly_disabled_api_server, None == 'secret-key') also fail on current main at c4f5a45d (main run); #70249 and #70247 hit it independently as well. Suggested next action: rebase/rerun after the gateway-config regression on main is fixed rather than changing this patch for that failure.