Session resume crashes on large multi-day session (12MB transcript, 207 subagents)

Summary

claude --resume <id> freezes/crashes when resuming a large multi-day proactive-mode session with a deep subagent history. The session originally froze mid-execution on Mar 2, and attempting to resume it today crashes again before reaching an interactive prompt.

Environment

• Claude Code version: 2.1.65-dev.20260303.t044818.sha296b760
• Platform: macOS (Darwin 25.3.0, arm64)
• Shell: zsh
• Reporter: Alexey Komissarouk (@AlexeyMK, Anthropic Growth Infra)

Session details

• Session ID: 78def59a-608d-41c7-8820-57f165580fc9
• Project path: /Users/alexey/code/apps
• Mode: proactive/autonomous, multi-agent (heavy Task-tool subagent spawning)
• Duration: Feb 26 → Mar 3 (multi-day, continuous)
• Final transcript timestamp: 2026-03-03T03:26:59.982Z — last entry is a system message. Session died mid-tool-use (a Task spawn was in flight checking a PR).

Size characteristics (likely the repro key)

| Artifact | Size |
|---|---|
| Main transcript (78def59a-....jsonl) | 12 MB (2,768 lines) |
| Subagent transcripts directory | 107 MB total |
| Subagent count | 207 files |
| Largest single subagent (agent-a95932feed5ebae72.jsonl) | 15 MB |
| Second largest (agent-a6f979a2ef7ce9d25.jsonl) | 14 MB — heavy Playwright/screenshot work: 77 Bash calls, 22 navigate, 17 screenshot |

Four subagent transcripts are individually in the 13–15 MB range (all screenshot-heavy).

$ ls -laS .../78def59a.../subagents/ | head -5
14830487  agent-a95932feed5ebae72.jsonl
14254913  agent-a6f979a2ef7ce9d25.jsonl
14006967  agent-a3c21c1d76457db18.jsonl
13667678  agent-a2c9e58fabbdf255a.jsonl

Reproduction

cd /Users/alexey/code/apps
claude --resume 78def59a-608d-41c7-8820-57f165580fc9
# → freezes / crashes, never reaches prompt

What's different from #21022 / #25417

• #21022 is about grepping large session files during a running session. This is --resume itself failing to bootstrap.
• #25417 is keypress lag after a successful resume. This never gets that far.
• Neither mentions the subagent dimension — the main transcript here is only 12 MB (well under #21022's 50 MB threshold), but there are 207 subagent files totaling 107 MB. If resume walks/loads the subagents directory, that's the likely culprit.

Hypothesis

Resume may be eagerly loading (or at minimum stat-ing + parsing headers of) all subagent .jsonl files, and the 4× ~14 MB screenshot-laden subagents (base64 image data in tool results) blow out memory or a synchronous parse loop.

Impact

• Can't recover work from a 5-day proactive run (9 wired-up GrowthBook experiments, branch state, etc.)
• Had to reconstruct state manually from memory files
• Proactive/agent mode naturally produces lots of subagents over a multi-day run — this will hit anyone using that mode for real work

Workaround not yet attempted

Moving the subagents/ directory aside before resume — happy to try if useful for confirming the hypothesis.

---
Filed on behalf of Alexey Komissarouk (@AlexeyMK / slack @alexey)