fix(run_agent): call should_compress_preflight() for sub-threshold engines (#20316) by Beandon13 · Pull Request #20424 · NousResearch/hermes-agent

Beandon13 · 2026-05-05T21:09:07Z

Summary

run_conversation now consults ContextEngine.should_compress_preflight() when the request is below threshold_tokens, so engines like hermes-lcm can run incremental leaf-chunk compaction (or other deferred maintenance) without waiting for the 75% context fill cutoff.
Default ContextEngine.should_compress_preflight() still returns False — the built-in ContextCompressor is unaffected.
Exceptions raised by the engine hook are caught at debug level and treated as "skip preflight", so a buggy plugin can't break an otherwise-healthy turn.

Closes #20316

Testing

scripts/run_tests.sh tests/run_agent/test_run_agent.py::TestRunConversation::test_engine_preflight_fires_below_threshold tests/run_agent/test_run_agent.py::TestRunConversation::test_engine_preflight_skipped_when_returns_false tests/run_agent/test_run_agent.py::TestRunConversation::test_engine_preflight_exception_does_not_break_turn -q

▶ running pytest with 4 workers, hermetic env, in /tmp/hermes-r2-1-fix
  (TZ=UTC LANG=C.UTF-8 PYTHONHASHSEED=0; all credential env vars unset)
bringing up nodes...
bringing up nodes...

...                                                                      [100%]
3 passed in 4.03s

scripts/run_tests.sh tests/agent/test_context_engine.py -q

...................                                                      [100%]
19 passed in 1.73s

scripts/run_tests.sh tests/run_agent/test_run_agent.py::TestRunConversation::test_context_compression_triggered tests/run_agent/test_run_agent.py::TestRunConversation::test_glm_prompt_exceeds_max_length_triggers_compression -q

..                                                                       [100%]
2 passed in 6.34s

…gines (NousResearch#20316) Context engines that override ``should_compress_preflight()`` (e.g. the hermes-lcm plugin's incremental leaf-chunk compaction) never had their hook fired by ``run_conversation`` because the preflight block exited early once the hardcoded ``>= threshold_tokens`` check failed. As a result, ``LCM_DEFERRED_MAINTENANCE_ENABLED=1`` and friends were inert and accumulated raw_backlog debt indefinitely. Add an ``elif`` branch that delegates to the engine's preflight hook when the legacy threshold check does not fire. The default ``ContextEngine.should_compress_preflight()`` returns ``False`` so the built-in ``ContextCompressor`` is unaffected; engines opting in get a chance to ingest messages and request a single ``compress()`` pass for deferred maintenance. Exceptions are swallowed at debug level so a buggy engine cannot break an otherwise-healthy turn. Closes NousResearch#20316

alt-glitch · 2026-05-05T21:17:31Z

Related: #15806 (open PR addressing same issue — should_compress_preflight() dead code). This PR fixes #20316 with a slightly different approach (preflight call below threshold vs per-turn ingest hook).

barronlroth · 2026-05-13T17:02:32Z

Related UX/API concern from an LCM deployment: #25115.

This PR wires the preflight hook so alternative context engines can run below the normal threshold. That is exactly the path where host status wording becomes important: if should_compress_preflight() is being used for plugin-specific maintenance, the generic 📦 Preflight compression... status can be misleading.

It would be useful if this hook, or a companion ContextEngine capability, allowed the active engine to return/suppress a preflight status label. Default built-in compressor behavior could stay unchanged, while LCM/retrieval engines could use neutral or engine-specific wording.

teknium1

Thanks for identifying the dormant context-engine hook. The premise still holds on current main: agent/turn_context.py:420 only calls _compressor.should_compress(_preflight_tokens), while the optional hook remains defined at agent/context_engine.py:110-116.

Problems

The implementation is stale: run_agent.py:5775-5796 now forwards to the extracted conversation loop, and the live preflight block is agent/turn_context.py:350-463. GitHub currently reports this branch as conflicting.
A port must preserve the newer deferral, cooldown, and Codex-native guards at agent/turn_context.py:399-420; the alternate-engine decision should not bypass those protections.

Suggested changes

Rework the patch against agent/turn_context.py and add current-path coverage for a concrete engine whose preflight hook requests below-threshold maintenance, returns false, and raises.

Automated hermes-sweeper review.

teknium1 · 2026-07-12T23:15:31Z

                    )
                    if _preflight_tokens < self.context_compressor.threshold_tokens:
                        break  # Under threshold
+            elif hasattr(self.context_compressor, "should_compress_preflight"):


Current main extracted this preflight block into agent/turn_context.py:350-463; run_agent.py:5775-5796 is now only a forwarder. Please port this branch into the live turn-context path and preserve its current deferral, cooldown, and Codex-native-compaction guards.

@Beandon13

…ght flow Relocates the #20424 wiring: the preflight region moved out of run_agent.py into agent/turn_context.py (and through the compression.max_attempts unification, #69315), so the contributor's elif branch is reapplied at its current home as the else arm of the threshold dispatch chain. Integration contracts: - Byte-identical default: the built-in ContextCompressor inherits ContextEngine.should_compress_preflight() -> False, so the default path performs no compression and touches no turn bookkeeping (pinned by test_builtin_compressor_default_sub_threshold_path_unchanged). - Attempt-cap: the engine gets exactly ONE compress() pass per turn, mutually exclusive with the cap-bounded threshold multi-pass loop, so turn-start passes stay within the resolved compression.max_attempts budget in every case. - No-op blocking (#64382 / 377244f): an engine pass that no-ops (_compress_context returns the input list object) neither sets nor clears preflight_compression_blocked and does not re-baseline the flush history — a sub-threshold maintenance no-op proves nothing about over-threshold compressibility. - Engine exceptions are swallowed at debug level; cooldown/defer/ codex-native gates run before the hook is ever consulted. Salvaged from #20424 by @Beandon13. Fixes #20316.

teknium1 · 2026-07-23T15:44:43Z

Merged via #69865 (commit 929c952). Your commit + tests were cherry-picked with authorship preserved; the wiring was reapplied at the code's current home (turn_context) integrated with the attempt cap and no-op-blocking machinery, with a byte-identical-default pin. Fixes #20316.

@Beandon13

…ght flow Relocates the NousResearch#20424 wiring: the preflight region moved out of run_agent.py into agent/turn_context.py (and through the compression.max_attempts unification, NousResearch#69315), so the contributor's elif branch is reapplied at its current home as the else arm of the threshold dispatch chain. Integration contracts: - Byte-identical default: the built-in ContextCompressor inherits ContextEngine.should_compress_preflight() -> False, so the default path performs no compression and touches no turn bookkeeping (pinned by test_builtin_compressor_default_sub_threshold_path_unchanged). - Attempt-cap: the engine gets exactly ONE compress() pass per turn, mutually exclusive with the cap-bounded threshold multi-pass loop, so turn-start passes stay within the resolved compression.max_attempts budget in every case. - No-op blocking (NousResearch#64382 / 0e73a78): an engine pass that no-ops (_compress_context returns the input list object) neither sets nor clears preflight_compression_blocked and does not re-baseline the flush history — a sub-threshold maintenance no-op proves nothing about over-threshold compressibility. - Engine exceptions are swallowed at debug level; cooldown/defer/ codex-native gates run before the hook is ever consulted. Salvaged from NousResearch#20424 by @Beandon13. Fixes NousResearch#20316.

alt-glitch added type/bug Something isn't working P3 Low — cosmetic, nice to have comp/agent Core agent runtime: loop, agent_init, prompt builder, context-compression, responses endpoint comp/plugins Plugin system and bundled plugins labels May 5, 2026

barronlroth mentioned this pull request May 13, 2026

Allow alternative context engines to suppress or customize preflight compression status #25115

Closed

johnnykor82 mentioned this pull request May 20, 2026

User message arriving during preflight compaction aborts the agent loop when an alt context engine is active #29370

Closed

teknium1 reviewed Jul 12, 2026

View reviewed changes

100yenadmin mentioned this pull request Jul 14, 2026

[Bug]: Pre-API pressure loop bypasses context-engine eligibility and retries LCM no-ops #64382

Open

teknium1 mentioned this pull request Jul 23, 2026

fix(agent): wire should_compress_preflight so context engines can run deferred maintenance #69865

Merged

teknium1 closed this Jul 23, 2026

Navigation Menu

Uh oh!

fix(run_agent): call should_compress_preflight() for sub-threshold engines (#20316) - #20424

fix(run_agent): call should_compress_preflight() for sub-threshold engines (#20316)#20424
Beandon13 wants to merge 1 commit into
NousResearch:mainfrom
Beandon13:fix/hermes-20316-lcm-preflight

Beandon13 commented May 5, 2026

Uh oh!

alt-glitch commented May 5, 2026

Uh oh!

barronlroth commented May 13, 2026

Uh oh!

teknium1 left a comment

Uh oh!

teknium1 Jul 12, 2026

Uh oh!

teknium1 commented Jul 23, 2026

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

4 participants

Uh oh!

Conversation

Beandon13 commented May 5, 2026

Summary

Testing

Uh oh!

alt-glitch commented May 5, 2026

Uh oh!

barronlroth commented May 13, 2026

Uh oh!

teknium1 left a comment

Choose a reason for hiding this comment

Problems

Suggested changes

Uh oh!

teknium1 Jul 12, 2026

Choose a reason for hiding this comment

Uh oh!

teknium1 commented Jul 23, 2026

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

4 participants