Completed background subagents unresumable after background-session fork: "No transcript found for agent ID" (transcript exists in pre-fork session dir)

Environment

• Claude Code v2.1.216, macOS (darwin 25.5.0)
• Background session (job) started via claude as a background job, with subagents dispatched in background

Summary

After a background session forks (SessionStart:fork fired mid-conversation, session UUID rotated), SendMessage to a subagent that had completed before the fork fails with:

{"success":false,"message":"Agent <id> could not be resumed: No transcript found for agent ID: <id>"}

The transcript is not lost — it still exists on disk, in the pre-fork session's subagents/ directory. The fork carry-over mechanism symlinks running agents' transcripts into the new session dir, but does nothing for completed agents, and the resume lookup only searches the current session's subagents/. Resuming completed agents from transcript is a documented feature ("names keep working after an agent completes — a send resumes it from its transcript"), so this looks like a path-resolution gap in the fork, not GC.

Timeline (single occurrence, fully reconstructed from transcripts)

Local time; old session bc552010-…, new session 29a1e714-… (= the job id).

1. ~11:55 — subagent A dispatched in background. Completes ~12:01, completion notification received with full result.
2. 12:02:20 — SessionStart:fork hook runs mid-conversation; all MCP servers disconnect/reconnect. Old session JSONL ends 12:02:19, new session JSONL starts 12:02:26 (first entry: queue-operation — the fork coincided with delivery of a queued user message to the job; the job's state.json carries respawnFlags incl. --reply-on-resume).
3. Subagent B, dispatched seconds before the fork, fails with the system notification "was checkpointed for the background fork but could not be resumed (resume failed)".
4. Subagent C, dispatched fresh after the fork, works; its outputs now live under the new session UUID.
5. ~12:25 — SendMessage to completed pre-fork agent A → No transcript found for agent ID.
6. ~12:55 — control: SendMessage to post-fork agent D, completed ~25 min earlier (comparable delay) → success, "had no active task; resumed from transcript". The only variable separating success from failure is pre- vs post-fork, ruling out a completion-age timeout.

Filesystem evidence

Under ~/.claude/projects/<project-slug>/:

bc552010-…/subagents/
  agent-<A>.jsonl        (294 KB — completed pre-fork builder, INTACT)
  agent-<A>.meta.json
  agent-<B>.jsonl        (1.3 KB — the checkpoint-failed agent, barely started)
  agent-<E>.jsonl        (real file; agent E was RUNNING at fork time)

29a1e714-…/subagents/
  agent-<E>.jsonl -> ../../bc552010-…/subagents/agent-<E>.jsonl   ← symlink: running agent carried over
  agent-<E>.meta.json    (fresh)
  agent-<C>.jsonl        (post-fork agents: real files)
  … no entry at all for completed agent <A>, nor for <B>

So: running agents are carried across via symlink (and keep working — E resumed fine); completed agents are stranded in the old dir, and the resume lookup (scoped to the current session's subagents/) can't see them.

Expected

Either the fork carries over all subagent transcripts (completed included, symlink or copy), or the resume lookup falls back to the parent/pre-fork session directory when an agent id isn't found locally.

Workaround

Manually symlinking agent-<id>.jsonl (+ a meta.json) from the old session's subagents/ into the new one restores resumability.

Repro sketch

1. Start a background session (job) and dispatch a background subagent.
2. Wait for the subagent to complete (notification received).
3. Cause the session to fork — in our case it happened on delivery of a queued message to the job (SessionStart:fork in the transcript, session UUID rotates).
4. SendMessage to the completed agent id → No transcript found for agent ID, while <old-session>/subagents/agent-<id>.jsonl exists on disk.

Related (searched, none covers this case)

#38443 (cross-session resume limitation), #23821 (subagent files lost after compaction), #59248 (orphaned subagents dirs), #23707 (Web sandbox recycling) — none describes the running-symlinked / completed-stranded asymmetry on background fork.