Skip to content

Hermes Agent context engine: the main model writes the compaction handoff on its cached prompt prefix

Prevent stale compaction workers from replacing current status

Elevatormusic080f6de42b45
diff --git a/README.md b/README.md
index 74c2d5a..d15dec2 100644
--- a/README.md
+++ b/README.md
@@ -163,7 +163,7 @@ The warm request uses the key that Hermes gives to the engine. The fallback requ
## Check the result
-Hermes writes one metadata line for each compaction attempt to `logs/agent.log` in the Hermes home folder:
+Hermes writes compaction metadata to `logs/agent.log` in the Hermes home folder:
```text
warm_compaction: path=warm reason=accepted elapsed_s=10.656 prompt_tokens=108021 cached_tokens=107968
@@ -173,7 +173,7 @@ warm_compaction: path=warm reason=accepted elapsed_s=10.656 prompt_tokens=108021
- `reason` is `accepted` or the refusal code of the warm request: `disabled`, `no_capture`, `api_mode_unsupported`, `route_changed`, `settings_unsupported`, `source_transform_unsupported`, `history_changed`, `capacity`, `cancelled`, `middleware_unavailable`, `middleware_refused`, `middleware_rewrite`, `middleware_repeated`, `middleware_after_capture`, `middleware_order_unknown`, `headers_unknown`, `tls_unknown`, `tls_unverified`, `middleware_changed_reply`, `provider_error`, `timeout`, `incomplete_response`, or `gate:<reason>`. The warm request goes through the Hermes `llm_request` and `llm_execution` middleware, as a main request does.
- `cached_tokens` is `None` when the server does not report it. Then the cache reuse is unknown.
-The engine status (`get_status()`) has the same values for the latest attempt in `warm_last`. The log never contains message text, request bodies, or keys.
+The engine status (`get_status()`) has the same values in `warm_last`. After a new attempt starts, an older worker cannot replace this status or write a final metadata line. The log never contains message text, request bodies, or keys.
Manual `/compress` can report `No changes from compression` when all messages fit in the recent-history tail. Then the plugin has no older history to summarize. The request estimate also includes system instructions and tool definitions, which the plugin does not compress. Use `/context` to see the token counts for each part. This unchanged result does not count as a warm-path failure.
diff --git a/evidence/plugin-integration.json b/evidence/plugin-integration.json
index 34b2232..9b53d89 100644
--- a/evidence/plugin-integration.json
+++ b/evidence/plugin-integration.json
@@ -1,6 +1,6 @@
{
"check": "standalone warm_compaction plugin on unpatched Hermes Agent",
- "started_utc": "2026-10-07T03:43:35Z",
+ "started_utc": "2026-10-07T03:59:22Z",
"hermes": {
"commit": "45871e100feceb89769536c88e5e6e265226a409",
"clean": true
@@ -10,7 +10,7 @@
"plugin_files_sha256": {
"warm_compaction/__init__.py": "760484ace80b9da77d7397828f69178905c9ac95df7b8114956bde727bd98aef",
"warm_compaction/capture.py": "b7ba65676334da63dacc5cebdd2399f46f85e35545be22fccae9443cebcf7ce0",
- "warm_compaction/engine.py": "56c44d835f5149a9d39ee8235f5cbc8fc778e0b3e0791198103af5edae5c6646",
+ "warm_compaction/engine.py": "33910bde19aa7b9a63778a82a9a1a4ddf863dc4c46835fd9561fbcc216ea7cbb",
"warm_compaction/fallback.py": "dc2b7fbfdee292601e52549d23d9b7224d40da1d8c06f553f3d0b8e10a23534d",
"warm_compaction/handoff.py": "fcab51def8fbcc2b1928957930c9a5c58a4ad202c949fd9f870a77f3b30276ae",
"warm_compaction/layout.py": "b49314c5236268247c337d9fae939135c97b47384300dcdfd00b0da0c040db93",
@@ -79,11 +79,11 @@
"threshold_tokens": 100000,
"context_length": 200000
},
- "compress_seconds": 0.143,
+ "compress_seconds": 0.142,
"warm_last": {
"path": "warm",
"reason": "accepted",
- "elapsed_s": 0.042,
+ "elapsed_s": 0.041,
"prompt_tokens": 33855,
"cached_tokens": 16927
},
@@ -192,7 +192,7 @@
"install",
"run"
],
- "process_seconds": 73.363
+ "process_seconds": 70.006
},
{
"scenario": "manual_fallback",
@@ -254,11 +254,11 @@
"threshold_tokens": 100000,
"context_length": 200000
},
- "compress_seconds": 0.301,
+ "compress_seconds": 0.298,
"warm_last": {
"path": "fallback",
"reason": "provider_error",
- "elapsed_s": 0.14,
+ "elapsed_s": 0.156,
"prompt_tokens": null,
"cached_tokens": null
},
@@ -266,7 +266,7 @@
"final_session_id": "wc-manual_fallback",
"continuation_keeps_custom_system_text": false,
"fence_denied": {
- "child_process_blocked": 6,
+ "child_process_blocked": 5,
"network_host_blocked": 1
},
"fence_blocked_programs": [
@@ -367,7 +367,7 @@
"install",
"run"
],
- "process_seconds": 40.173
+ "process_seconds": 41.015
},
{
"scenario": "manual_fixed",
@@ -428,7 +428,7 @@
"threshold_tokens": 100000,
"context_length": 200000
},
- "compress_seconds": 0.302,
+ "compress_seconds": 0.299,
"warm_last": {
"path": "fixed",
"reason": "provider_error",
@@ -541,7 +541,7 @@
"install",
"run"
],
- "process_seconds": 41.571
+ "process_seconds": 39.251
},
{
"scenario": "auto_tool_loop",
@@ -629,14 +629,14 @@
"warm_last_1": {
"path": "warm",
"reason": "accepted",
- "elapsed_s": 0.032,
+ "elapsed_s": 0.031,
"prompt_tokens": 34449,
"cached_tokens": 17224
},
"warm_last_2": {
"path": "warm",
"reason": "accepted",
- "elapsed_s": 0.016,
+ "elapsed_s": 0.031,
"prompt_tokens": 18744,
"cached_tokens": 9372
},
@@ -744,7 +744,7 @@
"install",
"run"
],
- "process_seconds": 43.307
+ "process_seconds": 42.949
},
{
"scenario": "auto_warm_failures",
@@ -1036,7 +1036,7 @@
"install",
"run"
],
- "process_seconds": 45.512
+ "process_seconds": 44.747
},
{
"scenario": "rollback",
@@ -1089,11 +1089,11 @@
"threshold_tokens": 100000,
"context_length": 200000
},
- "compress_seconds": 0.194,
+ "compress_seconds": 0.144,
"warm_last": {
"path": "warm",
"reason": "accepted",
- "elapsed_s": 0.015,
+ "elapsed_s": 0.016,
"prompt_tokens": 33845,
"cached_tokens": 16922
},
@@ -1272,10 +1272,10 @@
"run",
"rollback"
],
- "process_seconds": 52.328
+ "process_seconds": 48.135
}
],
- "finished_utc": "2026-10-07T03:48:32Z",
+ "finished_utc": "2026-10-07T04:04:08Z",
"passed": true,
"run_folder": "outside repository"
}
diff --git a/tests/test_wc_engine.py b/tests/test_wc_engine.py
index 6c03fee..0b2b037 100644
--- a/tests/test_wc_engine.py
+++ b/tests/test_wc_engine.py
@@ -4,6 +4,7 @@ import json
import threading
import unittest
from types import SimpleNamespace
+from unittest.mock import patch
import wc_hermes_stub
from warm_compaction.fallback import END_LINE
@@ -864,21 +865,34 @@ class EngineTest(unittest.TestCase):
self.commit(engine)
self.assertEqual(engine._warm_failures, 1)
- def test_a_late_worker_cannot_change_the_current_candidate_or_counter(self):
+ def test_a_late_worker_cannot_change_the_current_result_or_counter(self):
+ for wait_at in ("fallback", "build"):
+ for committed in (False, True):
+ with self.subTest(wait_at=wait_at, committed=committed):
+ self.check_late_worker_publication(wait_at, committed)
+
+ def check_late_worker_publication(self, wait_at, committed):
+ from warm_compaction import layout
+
entered, release = threading.Event(), threading.Event()
results, errors = [], []
engine = self.make()
history = [*old_turns(), assistant("done")]
- complete = self.llm.complete
+ complete, build = self.llm.complete, layout.build
- def delayed_complete(*args, **kwargs):
- if threading.current_thread().name == "old-compaction":
+ def wait_for_replacement(phase):
+ if phase == wait_at and threading.current_thread().name == "old-compaction":
entered.set()
if not release.wait(5):
raise TimeoutError("test worker was not released")
+
+ def delayed_complete(*args, **kwargs):
+ wait_for_replacement("fallback")
return complete(*args, **kwargs)
- self.llm.complete = delayed_complete
+ def delayed_build(*args, **kwargs):
+ wait_for_replacement("build")
+ return build(*args, **kwargs)
def old_attempt():
try:
@@ -886,23 +900,40 @@ class EngineTest(unittest.TestCase):
except BaseException as error:
errors.append(error)
- worker = threading.Thread(target=old_attempt, name="old-compaction", daemon=True)
- worker.start()
- try:
- self.assertTrue(entered.wait(5), "old worker did not enter the fallback")
- engine.compress(history)
+ with (patch.object(self.llm, "complete", side_effect=delayed_complete),
+ patch.object(layout, "build", side_effect=delayed_build),
+ self.assertLogs("warm_compaction.engine", level="INFO") as logs):
+ worker = threading.Thread(target=old_attempt, name="old-compaction", daemon=True)
+ worker.start()
+ try:
+ self.assertTrue(entered.wait(5), "old worker did not reach the wait point")
+ engine.compress(history)
+ if committed:
+ self.commit(engine)
+ current = dict(engine.warm_last)
+ current_status = engine.get_status()
+ current_pending = engine._pending_warm_result
+ metadata = [line for line in logs.output if "warm_compaction: path=" in line]
+ self.assertEqual(len(metadata), 1)
+ finally:
+ release.set()
+ worker.join(5)
+ self.assertFalse(worker.is_alive())
+ self.assertEqual(errors, [])
+ self.assertEqual(len(results), 1)
+ self.assertTrue(results[0] is history, "old worker returned a candidate")
self.assertEqual(engine.compression_count, 1)
- finally:
- release.set()
- worker.join(5)
- self.assertFalse(worker.is_alive())
- self.assertEqual(errors, [])
- self.assertEqual(len(results), 1)
- self.assertTrue(results[0] is history, "old worker returned a candidate")
- self.assertEqual(engine.compression_count, 1)
- self.assertEqual(engine._warm_failures, 0)
- self.commit(engine)
- self.assertEqual(engine._warm_failures, 1)
+ self.assertEqual(engine._pending_warm_result, current_pending)
+ self.assertEqual(engine._warm_failures, int(committed))
+ with self.subTest(publication="warm_last"):
+ self.assertEqual(engine.warm_last, current)
+ with self.subTest(publication="status"):
+ self.assertEqual(engine.get_status(), current_status)
+ with self.subTest(publication="metadata_log"):
+ self.assertEqual([line for line in logs.output if "warm_compaction: path=" in line],
+ metadata)
+ self.commit(engine)
+ self.assertEqual(engine._warm_failures, 1)
def test_an_unusable_capture_does_not_set_the_reasoning_rule(self):
# The capture is of this route, but its rows are not the stored rows: it does not show what the next
diff --git a/warm_compaction/engine.py b/warm_compaction/engine.py
index 9fd5aa6..13c0f40 100644
--- a/warm_compaction/engine.py
+++ b/warm_compaction/engine.py
@@ -161,7 +161,7 @@ class WarmCompactionEngine(ContextEngine):
self.compression_count = 0
self.warm_last: dict[str, Any] | None = None
self._wc_session_id = ""
- self._wc_result_lock = threading.Lock()
+ self._wc_result_lock = threading.RLock()
self._wc_attempt_serial = 0
self._pending_warm_result: tuple[str, int, dict[str, Any]] | None = None
self._wc_route: tuple = (None, None, None)
@@ -273,9 +273,10 @@ class WarmCompactionEngine(ContextEngine):
self.awaiting_real_usage_after_compression = False
def get_status(self) -> dict[str, Any]:
- status = super().get_status()
- status["warm_last"] = dict(self.warm_last) if self.warm_last else None
- return status
+ with self._wc_result_lock:
+ status = super().get_status()
+ status["warm_last"] = dict(self.warm_last) if self.warm_last else None
+ return status
def compress(self, messages: list, current_tokens: int | None = None, focus_topic: str | None = None,
force: bool = False, memory_context: str = "") -> list:
@@ -334,13 +335,13 @@ class WarmCompactionEngine(ContextEngine):
record["path"] = "fallback"
if self._cancelled():
record.update(path="cancelled", reason=record["reason"] or "cancelled")
- self._finish(record, started)
+ self._finish(record, started, attempt[-1])
return messages
if self._attempt() != attempt:
# A model or session switch during the attempt (while a request was on the network, for example): the
# summary is of the old route or session. Hermes keeps the history.
record.update(path="cancelled", reason="route_changed")
- self._finish(record, started)
+ self._finish(record, started, attempt[-1])
return messages
# The fixed summary takes at most the reserve, and at most the room.
fixed_budget = self._fixed_budget(overhead, reserve)
@@ -417,10 +418,10 @@ class WarmCompactionEngine(ContextEngine):
with self._wc_result_lock:
if self._cancelled() or self._attempt() != attempt:
record.update(path="cancelled", reason="cancelled" if self._cancelled() else "route_changed")
- self._finish(record, started)
+ self._finish(record, started, attempt[-1])
return messages
self.compression_count += 1
- self._finish(record, started)
+ self._finish(record, started, attempt[-1])
# The host can still reject this candidate. Only its successful boundary updates the failure streak.
self._pending_warm_result = (attempt[3], self.compression_count, dict(record))
return new
@@ -670,12 +671,17 @@ class WarmCompactionEngine(ContextEngine):
except Exception:
return False
- def _finish(self, record: dict[str, Any], started: float) -> None:
- record["elapsed_s"] = round(self._clock() - started, 3)
- self.warm_last = record
- logger.info("warm_compaction: path=%s reason=%s elapsed_s=%s prompt_tokens=%s cached_tokens=%s",
- record["path"], record["reason"], record["elapsed_s"], record["prompt_tokens"],
- record["cached_tokens"])
+ def _finish(self, record: dict[str, Any], started: float, attempt_serial: int) -> None:
+ """Publish metadata only while this local attempt owns the engine."""
+ # The final candidate already holds this reentrant lock. All other exits use the same publication guard.
+ with self._wc_result_lock:
+ if attempt_serial != self._wc_attempt_serial:
+ return
+ record["elapsed_s"] = round(self._clock() - started, 3)
+ self.warm_last = record
+ logger.info("warm_compaction: path=%s reason=%s elapsed_s=%s prompt_tokens=%s cached_tokens=%s",
+ record["path"], record["reason"], record["elapsed_s"], record["prompt_tokens"],
+ record["cached_tokens"])
def _clear_warm_failures(self) -> None:
self._warm_failures, self._warm_fixed = 0, 0