AI CLI Disk Full And Session Not Saved

When Gemini CLI, Claude Code, Codex, or another local agent hits ENOSPC, the worst failure mode is not a loud crash. It is continuing the session while logs, transcript files, task output, or recorder state are no longer being persisted.

No mail app or GitHub login? Send this directly from any inbox.

liuminsheng3@gmail.com - SafeDisk AI CLI Disk Full Pilot Audit Invoice

Silent data loss guard

Check recording state before trusting a long agent session after disk pressure.

Use read-only checks to prove whether the filesystem is still writable, whether the recorder file is active, and which cache or worktree path should be cleaned first.

df -h . "$HOME" /tmp 2>/dev/null; df -i . "$HOME" /tmp 2>/dev/null; touch "$HOME/.safedisk-write-test" && rm -f "$HOME/.safedisk-write-test"

Ask for safe next step Email $29 invoice Request $29 review Request $99 team policy Open runbook Mac agent storage guide Open GitHub invoice

No mail app or GitHub login? Send this directly from any inbox.

liuminsheng3@gmail.com - SafeDisk AI CLI Disk Full Pilot Audit Invoice

Free first-pass triage

Get the safest next AI CLI persistence check by email.

Email is enough for the first pass. Add the CLI name or warning only if you have it handy; we reply with the next read-only check for transcript, task output, or recorder state.

First Response Runbook

If a CLI disables recording or claims it can resume a session after an ENOSPC event, verify persistence before continuing. A working chat UI does not prove the transcript, recorder, or task output is still durable.

df -h . "$HOME" /tmp 2>/dev/null
df -i . "$HOME" /tmp 2>/dev/null
touch "$HOME/.safedisk-write-test" && rm -f "$HOME/.safedisk-write-test"
du -sh "$HOME/.gemini" "$HOME/.claude" "$HOME/.codex" 2>/dev/null
find "$HOME/.gemini" "$HOME/.claude" "$HOME/.codex" -maxdepth 4 -type f -size +100M -print0 2>/dev/null | xargs -0 du -h 2>/dev/null | sort -hr | head -30

What To Guard

Recorder files or conversation JSONL that can be set to null or abandoned after ENOSPC.
Task output capture files that can report disk full while normal writes still work.
Per-agent worktrees, build outputs, browser profiles, model caches, and package caches.
Exit summaries or resume commands that imply a session is restorable when no transcript was saved.

Copy-ready issue reply

Use this when a CLI keeps running after ENOSPC.

This keeps the thread focused on recorder durability and safe cleanup order instead of broad cache deletion.

One thing I would verify before continuing the session: after ENOSPC, prove the recorder/transcript is still durable, not just that the interactive CLI is still responding.

Read-only checks:

df -h . "$HOME" /tmp 2>/dev/null
df -i . "$HOME" /tmp 2>/dev/null
touch "$HOME/.safedisk-write-test" && rm -f "$HOME/.safedisk-write-test"
du -sh "$HOME/.gemini" "$HOME/.claude" "$HOME/.codex" 2>/dev/null
find "$HOME/.gemini" "$HOME/.claude" "$HOME/.codex" -maxdepth 4 -type f -size +100M -print0 2>/dev/null | xargs -0 du -h 2>/dev/null | sort -hr | head -30

If the recorder file was disabled or abandoned, the UI should surface that immediately and avoid printing a misleading resume command. Cleanup should start with rebuildable caches/worktrees, while active transcripts, credentials, plugins, and session state stay review-first.

Request $29 review GitHub $29 invoice

Paid Triage Scope

The $29 pilot audit reviews one short scan or summary and returns a safe first cleanup order plus a persistence checklist. The $99 team pilot turns one incident into rules for when to pause agents, what to clean first, and what must be preserved.