fleet release intermittently returns Internal server error for live rostered remote agents

[khaliqgant]

Impact

Fleet cleanup is not reliable: exact-name release requests can fail with a generic internal error even while the target is reported as live on an online node. This strands agents and makes clean-room orchestration teardown non-deterministic.

Exact reproduction (2026-09-06)

Environment:

• caller CLI: agent-relay v11.10.2
• target node: finn-mini
• target node reported broker relay-broker/11.8.3, online/live with handlers live

Both targets were listed by agent-relay fleet agent list --all --json as remote live+roster on finn-mini immediately before and after release.

Commands:

agent-relay fleet release relaycast-374-retry-finn-0906 --reason "PR 375 handed off to supervised exact-head review and merge gate"
agent-relay fleet release relaycast-374-sonnet-0906 --reason "PR 375 handed off to supervised exact-head review and merge gate"

Both waited roughly 20 seconds and returned only:

Internal server error

Control: a release of relayfile-465-veto-sonnet-0906 on the same node shortly beforehand dispatched successfully. A second target without a roster entry returned the expected Agent not found, so the CLI can distinguish at least that path.

Definition of done

• Return the typed downstream failure with invocation/node/request correlation instead of a generic 500.
• Make release idempotent for a live or already-exited exact-name remote worker.
• Reconcile roster and node-local presence after release so a successful response proves the process and roster record are gone.
• Add deterministic tests for success, already released, node timeout, stale roster, and version-skewed node handlers.
• Add this case to the two-clean-Daytona Fleet RelayFlow and fail qualification if cleanup cannot prove absence.

[khaliqgant]

Fresh cleanup failure on 2026-09-10 after two task-owned sf-mini review spawns entered accepted-but-unconfirmed states. Exact-name fleet release of cloud3531-review-mimo-0910 returned workspace_busy (Workspace write capacity is busy; retry with backoff) and did not establish process or roster absence. This is better typed than the original generic 500, but cleanup is still non-deterministic: the command did not honor the server delay, reconcile an exact release intent, or prove absence. The targets remain excluded from successful Fleet evidence and need bounded exact-name/identity cleanup after production admission recovers.