GitHub Repo
Hermes Agent Community Proposes Curator Fix After Skill Curation Is Marked Complete Despite Truncation
A report posted on October 13 says Curator may approach the model context limit while consolidating a large skill library, then mark a truncated response as complete. Two draft fixes propose propagating failure states and reserving output space, but neither has been merged.
Hermes Agent’s skill library maintenance feature has received a report about reliability during long-running tasks. An issue published on October 13 says that, after Curator’s model-based consolidation flow is enabled, the review agent may keep accumulating context until its response is truncated. The system still updates the run count and scheduled time, making unfinished work appear complete. The issue remains open, and the related fixes are still in draft form. Community issue
The reporter’s environment contained 44 agent-created skills, and the model’s context limit was 262,144 tokens. One curation run took 6.9 hours and made 98 tool calls. Its final request reached 261,643 tokens, leaving only 501 tokens of headroom. The author’s analysis suggests that full skill reads and rewritten content repeatedly entered the recent conversation, where they were protected from compaction, making the context difficult to shrink. This is an observation from a single environment; no smaller, deterministic reproduction case has been reported. Reproduction conditions and analysis
The official documentation says Curator’s deterministic idle archiving and model-based consolidation are separate stages. The model flow, which reads, patches, and merges skills, is disabled by default and must be enabled with curator.consolidate: true or invoked manually. A snapshot of the skill tree is created before consolidation and can be used for rollback. The report therefore directly concerns usage where consolidation has been explicitly enabled. Curator documentation
Draft PR #138076 proposes reserving up to 4,096 tokens for output when checking whether a request fits, with a cap of 10% for smaller context windows. It would also mark truncated results as incomplete and write pending skills to run.json. Another draft, #138012, focuses on propagating structured failure reasons so status queries and reports can show when a run ended early. Output reservation proposal, status propagation proposal
For long-running agents, this case shows that completion status needs to account for the model’s termination reason. Engineering teams should watch whether the fixes are merged, whether pending work can be handed off for resumption, and how the changes perform with large skill libraries. The public test results so far come from the proposal authors and do not establish that the issue has been resolved in an official release.