Reported from use, 2026-07-19. Two problems in the same flow.
Pick a run, stop it, then type another message into it. Expected: it continues that run. Actual: a brand new run appears in the rail.
1. It is a new run by design, and that design reads wrong
While a run is live, a message is a message control entry the run drains between turns, so it continues in place. Once the run has ended, RunResumeChat instead spawns a fresh run with --resume-session <id> (#720). The agent conversation does continue, but the dashboard shows an unrelated new row, so the continuity is invisible.
Worth deciding: keep spawning a new run but make the relationship visible (label it as continuing run X, or group them), or hold the row and present it as one continued run.
2. The resume frequently fails outright
Observed repeatedly in test-project (four of eight recent runs):
driver error: No conversation found with session ID: cdffc594-e95e-4fa5-a072-d64df151431e
[framework] claude-code exited (1)
The captured session id outlives the conversation the agent CLI is willing to resume. The run then just shows FAILED, with the reason only in the event log.
Either check the session can be resumed before spawning and fall back to a fresh run with a note, or surface the failure as "that conversation is gone" rather than a bare FAILED.
Reported from use, 2026-07-19. Two problems in the same flow.
Pick a run, stop it, then type another message into it. Expected: it continues that run. Actual: a brand new run appears in the rail.
1. It is a new run by design, and that design reads wrong
While a run is live, a message is a
messagecontrol entry the run drains between turns, so it continues in place. Once the run has ended,RunResumeChatinstead spawns a fresh run with--resume-session <id>(#720). The agent conversation does continue, but the dashboard shows an unrelated new row, so the continuity is invisible.Worth deciding: keep spawning a new run but make the relationship visible (label it as continuing run X, or group them), or hold the row and present it as one continued run.
2. The resume frequently fails outright
Observed repeatedly in
test-project(four of eight recent runs):The captured session id outlives the conversation the agent CLI is willing to resume. The run then just shows FAILED, with the reason only in the event log.
Either check the session can be resumed before spawning and fall back to a fresh run with a note, or surface the failure as "that conversation is gone" rather than a bare FAILED.