Summary

Kanban workers on long hardware/firmware workflows can be reclaimed with stale_lock while still legitimately busy, because no kanban_heartbeat / heartbeat_claim can run during a single blocking MCP tool invocation (e.g. large amlogic_firmware_download). The default claim TTL is 15 minutes (DEFAULT_CLAIM_TTL_SECONDS), so any one tool call longer than that loses the claim even if the process is healthy.

Expected behavior

• Running tasks should not be reclaimed while the worker is actively executing work, or
• There should be a supported, config-driven way to extend the initial claim TTL / heartbeat extension window for profiles that routinely block on long MCP calls.

Actual behavior

• After ~15 minutes from claimed, release_stale_claims reclaims the task (reclaimed / stale_lock), dispatcher spawns a second worker for the same task_id.
• First session may have started burn/download pipeline; second session duplicates effort and confuses operators.

Environment

• hermes-agent revision: faa13e49f (v2026.5.7-26-gfaa13e49f)
• Profile: case_runner (dispatcher-spawned worker, HERMES_KANBAN_TASK set)
• Gateway: embedded kanban dispatcher (default)

Evidence (sanitized)

• Same task_id, two runs: first run ends with outcome=reclaimed, error=stale_lock=...; current_run_id increments.
• Session timeline: kanban_show → DUT acquire → amlogic_firmware_download / extract / amlogic_burn_firmware_start — no tool round available to call kanban_heartbeat inside a long-running single MCP response.

Related upstream issues / PRs

• [Bug]: kanban_heartbeat tool doesn't extend claim TTL — diligent workers reclaimed at 15min #21147 / PR fix(kanban): heartbeat + reclaim + zombie + retry-cap fixes (#21147, #21141, #21169, #20881) #21183 — kanban_heartbeat must extend claim_expires (already addressed for between-tool heartbeats).
• recurring stall loop: kanban workers repeatedly reclaimed with stale_lock, zero output across respawns #23025 — recurring stale_lock / respawn loops (may overlap symptoms).

This report is adjacent to #21147: even with heartbeat_claim wired correctly, heartbeats cannot be issued while blocked inside one tool call longer than the TTL.

Possible directions (for maintainers)

1. Config: kanban.claim_ttl_seconds (or env HERMES_KANBAN_CLAIM_TTL_SECONDS) honored by dispatch_once / heartbeat_claim default extension.
2. Task-level override on the task row (e.g. optional claim_ttl_seconds for long-running board tasks).
3. Document explicitly that any single blocking operation must finish within DEFAULT_CLAIM_TTL_SECONDS, or split work (e.g. pre-download artifacts outside the Kanban worker).

Thanks for considering.