Hermes proposes task-level cost records that survive Kanban retries

A September 10 Hermes feature request brings fleet spending down to the unit an operator actually assigns: the task. The user describes a fleet of builder, tester, judge, security and verifier profiles, with session telemetry insufficient to explain the cost of one board item across multiple attempts. PR #107765 responds with immutable usage captured on closed Kanban runs and aggregation across retries, segmented by profile. The request and implementation proposal remained open when inspected.

The distinction is not simply another token counter in a dashboard. A task can have several runs, sessions and models before it is completed or blocked. If the board shows only the last session, an expensive failed attempt can disappear from the apparent cost of the final result. The proposal makes the durable run the accounting record and the task the aggregation boundary, exposing that information through existing operator surfaces.

The facts

- Issue #107744 asks for task-to-run-to-session accounting rather than only session-level telemetry. - The fleet use case explicitly names builder, tester, judge, security and verifier profiles. - Requested fields include input, output, cache, reasoning and turn counts together with model, provider, profile and timestamps. - The request distinguishes estimated_cost_usd from actual_cost_usd rather than treating an estimate as a provider bill. - PR #107765 proposes immutable token, cost, model, provider, session and turn usage on each closed Kanban run. - The proposed read surfaces are the Kanban CLI, worker tool and dashboard API, with task totals accumulated across retries.

Why it matters

Retries are a normal part of agent work, but they are also where a per-session cost view can understate the price of a deliverable. Profile segmentation lets an operator ask whether testing, review or repeated construction dominates a task's bill without concluding that the cheapest individual model is the cheapest workflow. The issue also asks that transcript deletion privacy not regress. Durable accounting need not mean retaining every conversation indefinitely.

Current

Inspected on 2026-09-11. The Hermes Agent release baseline is v2026.9.7 published 2026-09-07T22:17:01Z. The main source was open (not merged) when captured. Release metadata bounds availability; proposals, merged mainline changes and packaged releases are not interchangeable.

Evidence

Primary evidence: NousResearch/hermes-agent PR #107765 (https://github.com/NousResearch/hermes-agent/pull/107765). Supporting context: NousResearch/hermes-agent issue #107744 (https://github.com/NousResearch/hermes-agent/issues/107744); Hermes — multi-agent Kanban board documentation (https://hermes-agent.nousresearch.com/docs/user-guide/features/kanban/). These pages were inspected directly. Linked documentation establishes the product contract; it does not independently reproduce contributor or author experiments.

Operator take

Review accounting with a task that fails once, changes model or profile, then completes. The total should include both attempts without double counting, and unknown or estimated cost should remain visibly different from measured billing. Verify the records after process restart and after an allowed transcript deletion. These are proposed acceptance tests, not results from this publication; the PR's test checklist was still unchecked when captured.

Caveat

No green test result or shipped availability is established by the inspected PR. Broader Cron, webhook and bot-routine accounting is motivation in the issue, not demonstrated implementation in this Kanban proposal. Process-local delegation is explicitly outside the request's durable-unit scope.