Observed in production (MaKlaude local serve, 2026-08-09 ~20:31Z):
session ended: error_max_turns turns=41 cost=$6.47 559s
hit max turns; resuming f39692d6 (continuation 1, $6.47 spent) - work landed in the repo
WARNING: session ran no tools at all - check that the agent profile is authenticated
session ended: success turns=0 cost=$0.00 0s
Loki confirms: session-outcome subtype=success turns=0 cost_usd=0.0 duration_s=0 tool_calls=0 at 20:31:19, six seconds after the continuation-decision continue event.
Two problems:
-
claude --resume produced a session that did nothing and reported success, which ends the chain (_died_mid_task sees success). The interrupted session's wrap-up work only continued because the repo-fingerprint follow-up pass (commit 526fd93) launched a fresh session afterwards. Without that feature the task would have stranded with $6.47 spent - the exact shape the continuation mechanism exists to prevent. A zero-tool, zero-cost, zero-duration "success" immediately after a resume looks like the resume failing to load the session rather than a real outcome, and should probably be treated as a failed resume (retry once, or fall through to the follow-up path explicitly) rather than a chain-ending success.
-
The warning text says "check that the agent profile is authenticated" - the agent claude profile isolation was removed in commit 4776803, so the message points at a mechanism that no longer exists.
The net outcome in this instance was fine (follow-up session picked the work up within seconds), so this is a robustness gap plus a stale message, not an active stall.
Observed in production (MaKlaude local serve, 2026-08-09 ~20:31Z):
Loki confirms:
session-outcome subtype=success turns=0 cost_usd=0.0 duration_s=0 tool_calls=0at 20:31:19, six seconds after the continuation-decisioncontinueevent.Two problems:
claude --resumeproduced a session that did nothing and reportedsuccess, which ends the chain (_died_mid_tasksees success). The interrupted session's wrap-up work only continued because the repo-fingerprint follow-up pass (commit 526fd93) launched a fresh session afterwards. Without that feature the task would have stranded with $6.47 spent - the exact shape the continuation mechanism exists to prevent. A zero-tool, zero-cost, zero-duration "success" immediately after a resume looks like the resume failing to load the session rather than a real outcome, and should probably be treated as a failed resume (retry once, or fall through to the follow-up path explicitly) rather than a chain-ending success.The warning text says "check that the agent profile is authenticated" - the agent claude profile isolation was removed in commit 4776803, so the message points at a mechanism that no longer exists.
The net outcome in this instance was fine (follow-up session picked the work up within seconds), so this is a robustness gap plus a stale message, not an active stall.