Known limitations

Enforcement gaps

• No mid-call abort. A plugin cannot cancel an in-flight model call. The response that is already generating completes and is billed. The tool gate (pre_tool_call) stops subsequent work at the next tool boundary.
• Runaway text-only sessions. A session that generates text without calling any tool never reaches the tool gate, so budgets cannot stop it.
• Estimated data softens hard verdicts by default. Under on_estimated.mode: warn_only, a hard verdict that rests on estimated data is degraded to soft. Set mode: enforce if your provider reports reliable usage. See Budgets.

Subagent attribution

• Child agents (delegate_task) run as their own sessions and fire their own hooks. Their tokens are captured independently and included in global totals.
• per_cron_job budgets and reports attribute delegated subagent spend back to the cron job that started the delegation chain. global remains the catch-all cap, including for any child whose link to its parent could not be resolved.
• A detached async child still self-enforces only against the global scope while it runs. Pausing a runaway child mid-flight through its parent’s cron budget is not implemented.

Mixture of Agents (MoA)

• Selecting a MoA preset as your model (/model <preset> --provider moa) runs several reference models plus an aggregator on every iteration. Only the aggregator fires a hook. The plugin re-attributes it to the aggregator’s real provider and model and tags the call with the preset name (moa_calls and moa_preset, shown in /stats and the dashboard). Reference-model tokens are never captured, because they run through a Hermes auxiliary call path that fires no hooks. A MoA session’s recorded cost is therefore a lower bound.
• The one-shot /moa <prompt> command is entirely invisible to telemetry, by Hermes’s design. It runs both the references and the aggregator through the auxiliary path and injects their synthesis into the real main-model call. Only that final call is recorded, so MoA cost, aggregator included, is not captured, and moa_calls and moa_preset are not set. Select the preset as your model if you want MoA usage tracked.

See Mixture of Agents for the details.

Pricing

• Automatic refresh covers OpenRouter models (through its API) and Google AI Studio (through a built-in table). Direct Anthropic and OpenAI prices need manual pricing.yaml entries, a new PricingSource, or a core pricing snapshot.
• The Google table is constant and drifts without manual updates.
• Gemini models with tiered pricing use the tier for contexts up to 200k tokens. Usage above that is undercounted.
• A private OpenAI-compatible endpoint with no cached model metadata resolves to no core snapshot, so it falls back to pricing.yaml.

Profile attribution

Per-profile tags are accurate only when each profile runs in its own process. A single multiplexed gateway may tag some runs with the wrong profile. See Profiles and shared home.

Database retention

telemetry.db grows without bound, and no automatic purge exists. For high-frequency cron jobs or more than about 100,000 rows, consider manual cleanup of old rows. A retention policy is not implemented yet.

Gateway restart required

Enabling the plugin takes effect only after a gateway restart. Cron runs that started before the restart have no telemetry.

For the engineering view of these limits, see Known limitations in the Internals section.