In a reported Hermes deployment, background process wakes competed in the same bounded queue as user-originated inputs. The operator recorded extended delays and queue-cap warnings while scheduled and delegated work continued. This matters for agent safety because a human stop, correction or new instruction can become less timely than routine autonomous…
Inspect the ClaimsCase · DiggingBeagle record
Hermes internal wake traffic can starve a human interruption
A single shared wake queue gives internal jobs enough capacity to delay or displace incoming human messages.
One Hermes Agent implementation and operator trace on October 20, 2026; proposed change not verified as released.
Assessment profile
Assessment method- Setting
- production
- Exploitation
- observed live
- Evidence
- Grade C
- AI role
- AGAINST AI
Basis and provenance
Wake backlog and cap events were observed; the logged dropped entries were internal wakes, and human-message loss is a demonstrated code-path risk rather than a counted observed outcome. Impact band is left unassigned because no calibrated damage threshold or representative victim impact is supported by the available record.
Assessed Oct 20, 2026 using diggingbeagle.assessment/1.
Evidence basis: CLM-HERMES-WAKE · CLM-HERMES-WAKE-OBSERVED · Hermes Agent issue 141822: human messages starve under wake queue · Hermes Agent issue 141824: workload and undelivered delegation
At a glance
User-originated interrupts require delivery and priority semantics independent of routine self-generated wake events.
Read the implicationsCross-deployment prevalence and actual time to execute an emergency stop are untested.
Limits and uncertaintyFull account
In a reported Hermes deployment, background process wakes competed in the same bounded queue as user-originated inputs. The operator recorded extended delays and queue-cap warnings while scheduled and delegated work continued. This matters for agent safety because a human stop, correction or new instruction can become less timely than routine autonomous triggers.
The issue describes a proposed reservation of queue slots for human traffic. Such a patch can reduce dropping under bursts but does not alone prove fair priority, bound delays or verified acceptance of user corrections. These are observations from one installation, not a measured fleet-wide defect rate.
The reporter observed 212 internal wake events over about 70 minutes, a full shared queue and repeated drop warnings. The report states that the observed dropped events were wake events; human follow-ups were vulnerable to the same shared cap and scheduling path, not counted as a proven set of discarded human messages. A reservation of eight slots for human traffic reduces one capacity collision, but does not itself prove bounded interrupt latency or durable acknowledgment.
Timeline
- Oct 19, 2026disclosure
Public disclosure
Claims & evidence
CLM-HERMES-WAKEThe issue documents human messages delayed behind internal wake events, including a bounded queue of 32 slots and repeated cap warnings.supported
Basis: reported finding
- supportsHermes Agent issue 141822: human messages starve under wake queueprimary disclosure
Issue #141822: trace counts, queue implementation and proposed correction
CLM-HERMES-WAKE-OBSERVEDThe issue records 212 internal wakes in roughly 70 minutes and 140 full-queue warnings; it specifies that the observed dropped events were wakes, while human traffic was exposed to the same shared limit.supported
Basis: reported finding
- supportsHermes Agent issue 141822: human messages starve under wake queueprimary disclosure
Issue #141822, paragraph (c), shared _BUSY_QUEUE_MAX_PENDING = 32 cap
Implications
User-originated interrupts require delivery and priority semantics independent of routine self-generated wake events.
Unknowns and contradictions
- Cross-deployment prevalence and actual time to execute an emergency stop are untested.
Sources and citation
Material revision history
- Oct 20, 2026 · Canonical change recorded · new in release · revision 177