Problem
When convert_session (session_to_subagent) is followed by SendMessage(to: <created_id>) in a session whose agent-teams resume runtime isn't functional, the harness returns:
{"success":false,"message":"Agent \"…\" has no transcript to resume. It may have
been cleaned up. (No transcript found for agent ID: …)"}
The file is written correctly and is still on disk — nothing was cleaned up. But the "may have been cleaned up" wording (harness-side) + the existence of our pristine-artifact reaper led a field agent to conclude a garbage collector was eating its artifacts. It re-converted three times, all "failed" the same way, then abandoned the resume path entirely and wasted many turns. Reproduced here in an agent-teams-enabled session: convert→SendMessage works first try (the artifact wakes and answers), confirming the field failure was the session's missing/partial agent-teams runtime, exactly as convert_session's docstring already warns.
What we could do on our side
We can't change the harness's error string, but we can keep agents from chasing the phantom:
- A lightweight preflight the skill/guidance recommends — convert a throwaway and
SendMessage it ONCE to confirm the resume channel works before building a multi-variant flow on it.
convert_session's response could surface a hint about whether SendMessage is likely to resolve (e.g. note that a "no transcript to resume" result means agent-teams isn't active, NOT that the artifact was deleted — the file is intact and delete_conversions/the reaper never touch in-use forks).
- Possibly a one-line note in the cc-explorer skill's conversion section distinguishing this from the reaper.
Breadcrumb
Field session bluetaka 22876bad; the convert→SendMessage detour preceded the rewind issues fixed in #44/#45.
Problem
When
convert_session(session_to_subagent) is followed bySendMessage(to: <created_id>)in a session whose agent-teams resume runtime isn't functional, the harness returns:The file is written correctly and is still on disk — nothing was cleaned up. But the "may have been cleaned up" wording (harness-side) + the existence of our pristine-artifact reaper led a field agent to conclude a garbage collector was eating its artifacts. It re-converted three times, all "failed" the same way, then abandoned the resume path entirely and wasted many turns. Reproduced here in an agent-teams-enabled session: convert→SendMessage works first try (the artifact wakes and answers), confirming the field failure was the session's missing/partial agent-teams runtime, exactly as
convert_session's docstring already warns.What we could do on our side
We can't change the harness's error string, but we can keep agents from chasing the phantom:
SendMessageit ONCE to confirm the resume channel works before building a multi-variant flow on it.convert_session's response could surface a hint about whetherSendMessageis likely to resolve (e.g. note that a "no transcript to resume" result means agent-teams isn't active, NOT that the artifact was deleted — the file is intact anddelete_conversions/the reaper never touch in-use forks).Breadcrumb
Field session bluetaka
22876bad; the convert→SendMessage detour preceded the rewind issues fixed in #44/#45.