[Bug] 2.5.2: completed subagent cards have no transcript after resuming a pre-2.5 session #4740

Environment

• Kimi Code CLI 2.5.2, upgraded from 2.4.1
• Session created under 2.4.1, resumed with kimi --session <id>
• Reproduced on a copy of the session, both binaries run side by side

What happens

Completed subagent cards render with no body and no tool calls. Same session, same subagent, same card:

2.4.1:

✗ agent stopped (<model> · high · <task description>)

2.5.2:

✗ SubAgent Failed (<task description>) · 0 tools
  · 0 tools
  │ Background agent killed

Subagents that are still running do render content. Historical ones are always empty. Ctrl-O does not expand them either: computeHiddenContent() returns false early for isSingleSubagentView().

Same on the web/desktop client: the subagent panel opens empty. The frontend already has openAgentPanel, agentPanelTurnsOf, agentPanelHasMore, loadOlderAgentMessages (paginated history), but there is nothing to page through.

Data is still there

The session has 9 subagent directories. The main wire holds 220 subagent.spawned links, and every agents/<id>/wire.jsonl still contains the full conversation. Reading those files by hand is currently the only way to see the history.

Steps to reproduce

1. On 2.4.1, run a session that spawns a subagent and let it finish.
2. Upgrade to 2.5.2.
3. kimi --session <id>, or open the session in web/desktop.
4. The subagent card shows · 0 tools and no body.

Likely cause

The TUI resume path calls:

harness.resumeSession({ id, replayTurnLimit: 11 })

with no includeSubagents. The session layer only materializes subagent replay behind:

if (replay?.includeSubagents === true) { /* walk agents/*, create + resumedAgentState */ }

so historical subagent state is never loaded. 2.4.1 makes the same call, but its card did not display detail. 2.5.0 switched to the SubAgent card (changelog: "优化长时间运行的 subagent 卡片，输出持续增长时保持响应"), which surfaces the gap.

Suggestion

Load subagent detail on demand instead of eagerly:

• keep resume as it is — summary line only, no eager folding;
• when an agent is opened, fold that agent's agents/<id>/wire.jsonl and page it by turns (e.g. 10 turns per page), reusing agentPanelHasMore / loadOlderAgentMessages, shared by TUI and web.

Default rendering cost stays the same; only the opened agent pays the fold, so old sessions with hundreds of subagents are not slowed down. A one-line stopgap is to pass includeSubagents: true on resume, but folding every subagent of an old session gets slow, so it only works short-term.

Expected

Resuming a pre-2.5 session and opening any completed subagent shows its tool calls and messages, with older turns loadable, without slowing the default view.