Bot Harness Gateway transport
Slack and Webex independently opt into detached Harness Engine execution through
the existing UI/BFF. Both switches default to false; Admin UI controls are not
part of this setting.
| Bot | Environment variable | Helm value |
|---|---|---|
| Slack | SLACK_HARNESS_GATEWAY_ENABLED | slack-bot.config.SLACK_HARNESS_GATEWAY_ENABLED |
| Webex | WEBEX_HARNESS_GATEWAY_ENABLED | webex-bot.config.WEBEX_HARNESS_GATEWAY_ENABLED |
slack-bot:
config:
SLACK_HARNESS_GATEWAY_ENABLED: "true"
webex-bot:
config:
WEBEX_HARNESS_GATEWAY_ENABLED: "false"
- For Compose, set either variable in
.envand recreate that bot service. - Configure Harness Engine on the BFF and enable the agent's runtime first.
Deep Agents additionally requires
HARNESS_ENGINE_DEEP_AGENTS_ENABLED=trueon the BFF. Existing native conversations retain their runtime binding. CAIPE_API_URLremains the BFF URL. Bots do not connect directly to workers, Engine storage, or MCP services.
Request and replay
- Admission transport failures and HTTP 502/503/504 retry at most five times, retaining one idempotency key. Once admitted, only the subscription reconnects.
- Slack message timestamps (scoped by workspace/channel) and Webex message IDs produce stable admission keys for redelivery of the same platform message. Calls without a delivery ID receive a fresh key for that invocation.
- If admission conflicts after redelivery rebuilds the thread context, the bot looks up the original run through the BFF and replays it. Lookup requires current agent-use and conversation-write access and is scoped to the caller, agent, conversation epoch and delivery key. It cannot start a replacement run or change the original input fingerprint. A missing run remains a conflict.
- Reconnects use numeric event/subframe cursors; duplicate events are discarded. Incomplete SSE frames replay from the previous complete frame.
- Empty/failed subscriptions retry at most five times, resetting the budget when a new event arrives. Exhaustion reports that the server run may remain active. Closing a subscription never sends a provider cancellation.
- Only explicit
HARNESS_RUNTIME_MISMATCHorAPP_STREAM_REQUIREDresponses before admission use the existing native/app stream. Other failures stop. - Turning a switch off restores the previous bot transport on subsequent calls; it does not cancel already-admitted runs. HITL uses the existing resume route.
Identity and persistence boundaries
- Admission and replay use the same existing OBO/service-account token selection. Each attach rechecks access; token expiration or permission denial stops replay.
- No sharing grants, tool permissions, or user credentials change. Webex's native app fallback retains its signed room/thread context.
- This transport stores neither bearer tokens nor delivery cursors on disk.
- Engine run persistence survives subscriber loss. Automatic bot delivery catch-up after a bot process restart is not implemented. There is no durable notification outbox or saved renderer state. Platform redelivery reuses the admission key but may replay previously delivered text into a new bot reply.
- Reconnects cannot resume provider work lost to worker failure; provider/Engine recovery guarantees remain runtime-specific.