Claude Code Session Recovery (stale index + cross-OS fragmentation)

S2 · Pattern 🔬 Research 2026-04-23

The promise

When /resume shows the wrong sessions, wrong summaries, or loses your active session entirely, you have three recovery primitives ordered by cost. Pick one in under a minute. No data lost.

The two failure modes (same surface, different mechanisms)

Both surface as “/resume is lying” — but the fix is different for each.

1. Stale `sessions-index.json`

Claude Code caches session summaries in ~/.claude/projects/<encoded-cwd>/sessions-index.json. The /resume picker reads it. The write path only runs on graceful shutdown. Ungraceful closures skip it:

wsl --shutdown
WSL terminal window closed without typing /exit
Machine suspend / sleep
OOM-kill, SIGKILL, force-quit
Upstream bug in the rewrite path itself (#41946 — clean exits can also lose entries)

The .jsonl session content keeps accumulating (append-on-write), but the index lags. Scan your machine:

for d in ~/.claude/projects/*/; do
  idx=$(stat -c %Y "$d/sessions-index.json" 2>/dev/null)
  jsonl=$(ls -t "$d"/*.jsonl 2>/dev/null | head -1)
  [ -n "$jsonl" ] && jsonl_mt=$(stat -c %Y "$jsonl")
  if [ -n "$idx" ] && [ -n "$jsonl_mt" ] && [ $((jsonl_mt - idx)) -gt 86400 ]; then
    echo "$((  (jsonl_mt - idx) / 86400 ))d lag: $(basename "$d")"
  fi
done

On one machine: 14 projects lagged up to 93 days. Upstream canonical issue: #25032.

2. Cross-OS content poisoning

Different bug, different fix. When you launch Claude Code from both WSL (/mnt/c/…/project) and Windows (C:\…\project), you get two encoded dirs under ~/.claude/projects/:

-mnt-c-Users-...-project/
C--Users-...-project/

/resume from one doesn’t see the other’s history. Worse: the .jsonl content carries OS-specific absolute paths (tool_use inputs, cwd fields, prose) — so even if you merge the storage, a resumed session tries to Read /mnt/c/… on Windows and fails.

Captured in challenge 02. Resolved via portaconv’s extract-and-paste design.

The three recovery primitives

Ranked by cost. Start at the top of the table and stop at the first one that fits.

Primitive	Cost	When	Limits
`claude -r <uuid>`	Zero	You know the session UUID; picker is unreliable	Requires knowing the UUID
`pconv rebuild-index`	One command	Stale index; want native picker to work	Requires the session to be in the right encoded bucket
`pconv doctor --dump-stale`	Paste into fresh session	Index rebuild didn’t help OR the `.jsonl` is too big to resume practically OR cross-OS content poisoning	Abandons the stale session; you’re pasting a window, not continuing

A. `claude -r <UUID>` by id

Skips the picker entirely. Reads the .jsonl directly.

# Find the UUID — newest jsonl by mtime in your project's encoded dir.
ls -t ~/.claude/projects/-mnt-c-...your-project.../*.jsonl | head -1
# => …/97d7b58b-09f5-41ea-a59f-a12f230083b0.jsonl

claude -r 97d7b58b-09f5-41ea-a59f-a12f230083b0

Works now, no tool installation, immune to index staleness. The obvious move when you know which session you want.

B. `pconv rebuild-index`

Rewrites sessions-index.json from the actual .jsonls. Native /resume picker works correctly afterwards.

# Check what's stale first (read-only).
pconv doctor

# Single project:
pconv rebuild-index --project ~/.claude/projects/-mnt-c-...your-project

# All projects on this machine (skips fresh ones with --lag-threshold-hours):
pconv rebuild-index --all --lag-threshold-hours 24

# Dry-run to see what would change:
pconv rebuild-index --all --dry-run

Atomic write + dated .bak-YYYY-MM-DD backup. Portaconv v0.1.0+.

Install:

cargo install --git https://github.com/cybersader/portaconv

Routine maintenance (weekly cron):

0 9 * * 1 pconv rebuild-index --all --lag-threshold-hours 168

Runs Monday 09:00, rebuilds any project whose index lags by more than a week.

C. `pconv doctor --dump-stale`

Detects stale projects AND emits paste-ready markdown for the newest session in each. Start a fresh claude session, paste, continue.

pconv doctor                    # list stale projects, read-only
pconv doctor --dump-stale       # + paste-ready markdown for each
pconv doctor --format json      # machine-readable for scripting

This is the escape hatch — use it when rebuild doesn’t apply (cross-OS case) or when the .jsonl is so large that continuing in place is more painful than starting fresh. The paste inherits enough context for the new session to pick up where the old one left off; you abandon the stale session.

Decision tree

Pick the cheapest primitive that fits — read top to bottom, stop at the first YES.

flowchart TD
    Start["/resume shows wrong sessions"] --> Know{"Know the<br/>session UUID?"}
    Know -- yes --> A["<b>A.</b> claude -r &lt;uuid&gt;"]
    Know -- no --> Cross{"Crossing<br/>WSL ↔ Windows?"}
    Cross -- yes --> X["pconv dump --rewrite<br/>(see challenge 02)"]
    Cross -- no --> Picker{"Want native<br/>/resume picker fixed?"}
    Picker -- yes --> B["<b>B.</b> pconv rebuild-index"]
    Picker -- "rebuild didn't help<br/>or jsonl too big" --> C["<b>C.</b> pconv doctor --dump-stale<br/>paste into fresh session"]

MCP integration

If Claude Code agents need to self-heal, pconv mcp serve exposes doctor alongside list_conversations and get_conversation:

// ~/.claude.json or project .mcp.json
{
  "mcpServers": {
    "portaconv": { "command": "pconv", "args": ["mcp", "serve"] }
  }
}

Agents call doctor via tools/call → JSON result with project_dir, lag_hours, missing, newest_session_id, newest_jsonl_mtime_ms. rebuild-index is deliberately CLI-only — write operations via MCP risk unintended fan-out.

Prevention

The honest answer: you can’t fully prevent this from inside Claude Code. The rewrite path is ungraceful-shutdown-vulnerable by design, and shell-level signal traps don’t survive wsl --shutdown or process-group-kill.

What you can do:

Habit: type /exit before closing WSL terminals. Catches the graceful-shutdown subset.
Cron: weekly pconv rebuild-index --all --lag-threshold-hours 168 as above.
Awareness: when the picker looks wrong, it usually is wrong — don’t assume you misremembered.

Why it matters

/resume is one of the most-used Claude Code commands. When it silently shows the wrong list:

Users lose confidence in the tool’s memory
Accidental session restarts instead of continuations (lost context)
Wasted tokens reloading context the agent already had
Unaudited work — “did we actually resume that yesterday?”

The symptom (“picker wrong”) is indistinguishable between the two failure modes, but the fix is different. This pattern separates them so you pick the right one in under a minute.