Skip to content
Agentic Workflow 02 — Agentic Workflow & Tech Stack

Upstream comment posted on anthropics/claude-code#25032

S5 · Instance ✓ Stable 2026-04-23

Posted: anthropics/claude-code#25032 — comment 4304663073 on 2026-04-23.

This file is kept as the durable local record. The body below matches what’s on GitHub verbatim.

Stale sessions-index.json — 14 projects drifting up to 93 days on one machine, hypothesized mechanism

Section titled “Stale sessions-index.json — 14 projects drifting up to 93 days on one machine, hypothesized mechanism”

Hit this on WSL2, v2.1.118, single-OS (not the cross-platform fragmentation from #17682 / #9668 / #9306). Posting because the cumulative-drift evidence and the mechanism hypothesis tying graceful-shutdown dependency to overwrite-of-cached-state don’t appear to be in the thread yet.

Environment

Section titled “Environment”
  • WSL2 on Windows (Linux 6.6.87.2-microsoft-standard-WSL2)
  • Claude Code v2.1.118 (Bun-bundled binary at ~/.local/share/claude/versions/2.1.118)
  • Single ~/.claude/projects/ bucket per project — not cross-OS-dual-encoded for the projects showing drift

Symptom

Section titled “Symptom”

/resume picker reports one of my active sessions as a 9-message stub last modified 2026-01-09, with summary “OpenCode permissions setup and Terminal Workspaces bug fix.”

Actual state of the .jsonl:

  • Size: 59 MB
  • Last .jsonl append: 2026-04-23 02:22 UTC (the current session I was in when I opened the picker)
  • Continuous append activity from 2025-12-10 through 2026-04-23 — about four months of real resume-ed work

The .jsonl content is clean and picker-loadable via claude -r <uuid> (bypass the picker). The index is the only broken artifact.

Cumulative drift — 14 projects on this machine

Section titled “Cumulative drift — 14 projects on this machine”

Scan of ~/.claude/projects/*/ on 2026-04-23, flagging projects where sessions-index.json mtime lags the newest non-subagent .jsonl mtime by more than 7 days:

Project (tail of encoded name)Lag
-mnt-d-MEDIA93d
home-assistant-projects90d
…tasknotes…obsidian-plugins-tasknotes (duplicate-encoded)84d
…tasknotes-enhancements80d
cynario79d
retake-studio66d
4-VAULTS65d
…terminal-workspaces57d
b-g-33---Home-Lab--Home-Server57d
b-g-15---Family-Planning--Parenthood46d
3d-printing44d
b-g-vault-b-g, …tasknotes…plugins-tasknotes (dot-prefix variant)12d each
DFD-Excalidraw-System11d

This rules out one-time WSL crash. It’s consistent with every ungraceful close dropping an index update, accumulating across months of normal WSL lifecycle (window-close without typing /exit, wsl --shutdown, machine suspend, OOM-kill).

Mechanism hypothesis — graceful-shutdown is necessary but not sufficient

Section titled “Mechanism hypothesis — graceful-shutdown is necessary but not sufficient”

  1. Ungraceful-shutdown path clearly breaks. SIGKILL / forced WSL shutdown / suspend skip application-side signal handlers and atexit — index update is lost, .jsonl survives because it’s append-on-write.

  2. Clean-exit path also loses entries per #41946. So the bug isn’t purely “SIGKILL skips flush.”

  3. A Zenn writeup (tjst_t) notes the index “rolled back two weeks” after a recovery attempt — suggests the rewrite path re-reads cached in-memory state and overwrites the on-disk index, rather than merging with the current filesystem state. Combined with racy writes on any abnormal timing, that would produce the cumulative drift pattern I see on this machine.

  4. @agatho’s #24729 comment points to a case-sensitivity bug in the multi-worktree code path (function xa in the minified cli.js). Other commenters confirm the bug triggers without worktrees too, which is what I observed.

So the fuller mechanism is likely: the write path is both racy against forced-kill AND regresses state on some graceful closures, because it re-reads a cached snapshot rather than merging with reality.

sessions-index.json format (as observed — this is not documented publicly)

Section titled “sessions-index.json format (as observed — this is not documented publicly)”
{
  "version": 1,
  "entries": [
    {
      "sessionId": "uuid-36-chars",
      "fullPath": "/abs/path/to/<sessionId>.jsonl",
      "fileMtime": 1769531155388,
      "firstPrompt": "first 200 chars of first user message",
      "customTitle": "…",          // optional, from type:custom-title records
      "summary": "…",              // optional, LLM-generated on graceful exit?
      "messageCount": 9,
      "created": "2026-01-09T15:38:59.766Z",
      "modified": "2026-01-09T15:48:27.815Z",
      "gitBranch": "",
      "projectPath": "/abs/project/path",
      "isSidechain": false
    }
  ],
  "originalPath": "/abs/project/path"
}

Fully undocumented in docs.anthropic.com / docs.claude.com. It would help third parties considerably if the schema were published, even as a footnote.

Workaround landscape (no Anthropic fix required to use these, but none should need to exist)

Section titled “Workaround landscape (no Anthropic fix required to use these, but none should need to exist)”

Third-party tools converging on two shapes:

  • Rebuild the index. tirufege’s Python gist is the original; I shipped a Rust port as pconv rebuild-index with atomic writes, dated backup, and a doctor counterpart for read-only detection. Round-trip tested against the schema above — pconv rebuild-index --all followed by pconv doctor reports zero stale projects.
  • Replace the picker. KirillPuljavin/cres, riii111/claude-resume — both read .jsonls directly and ignore the index.

Asks for Anthropic

Section titled “Asks for Anthropic”

Prioritized:

  1. Rebuild-on-startup-if-stale. On every launch, if sessions-index.json is missing OR lags the newest .jsonl in the same dir by >24h, regenerate it from the .jsonls before showing the picker. Cheap (<100 ms for the largest project I have; likely faster than what the picker already does).
  2. Switch write path from overwrite-with-cached-snapshot to merge-with-filesystem. Addresses the “rolled back two weeks” regressions even on graceful exit.
  3. Document the sessions-index.json format publicly. Footnote-length; enables the third-party ecosystem to operate on stable ground rather than reverse-engineering per-version.
  4. Consider emitting a warning from /resume when picker results disagree with .jsonl mtimes by more than N days. Even without an auto-fix, a hint like “the picker may be stale; consider running claude --rebuild-index or pconv rebuild-index” would close the feedback loop for users.

Reproducibility hints for whoever picks this up

Section titled “Reproducibility hints for whoever picks this up”

strace -fe trace=openat,rename,write,fsync -p $(pgrep -f claude) through a session, then diff index state across /exit vs wsl --shutdown vs SIGKILL, would pin the write cadence. I haven’t done this myself because claude -r <uuid> and pconv rebuild-index unblock me today; but the drift pattern I report above is strongly suggestive of the hypothesis above.

References: