rdmsm4x — replicantDB resume/handoff state and auto-resume after reboot

2026-08-31 16:30 EDT · rdmsm4x · claude@rdmsm4x · ticket TASK-20260831-12

Rich asked this session to be planned/documented for resume and handoff, and to auto-resume after a reboot. Both are done. No application code changed; this is state, documentation and one login agent.

What changed on this host

~/dev/fleet/ops/resume/ is untracked in git and was left that way: peer sessions were writing it during this change, and committing the directory would have swept their in-flight files into my commit. The file inventory above is the reconstruction record.

Commands run

plutil -lint <both plists>
DRY_RUN=1 zsh resume_pinned_on_login.zsh replicantdb
launchctl bootstrap gui/501 <DRYRUN twin plist> ; launchctl bootout gui/501/<twin>
launchctl print gui/501/com.eastcoastscience.replicantDB.daemon | grep runatload
sqlite3 "$HOME/Library/Application Support/replicantDB/replicantdb.sqlite" \
  "SELECT (SELECT count(*) FROM files), (SELECT count(*) FROM file_hashes), ..."

Verification evidence

• Daemon restarts itself: launchctl print reports properties = runatload for com.eastcoastscience.replicantDB.daemon, pid 91373, state = running, up 1h51m. Read from launchd's view of the loaded job, not from the plist file.
• Resume agent's launchd wiring: verified with a throwaway twin of the plist — label suffixed -DRYRUN, EnvironmentVariables DRY_RUN=1. Bootstrapped; resume-replicantdb.log gained exactly the two expected lines at 16:23:19 under launchd's own execution; twin booted out and confirmed absent from launchctl list. This is why no duplicate Terminal opened. A peer session that bootstrapped its real agent got a live second claude --resume on its own still-running transcript. The twin costs nothing and avoids that.
• Transcript resolution, with negative controls: DRY_RUN=1 resolves the 37 MB transcript for cd4dc58c…; a bogus session id logs STALE PIN and exits 0; a missing pin file logs a different message and exits 0; no argument exits non-zero. The check can fail, so its passing is informative.
• Corpus, 16:26 EDT: 2,025,107 files · 8,153 hashed · 229,145 purposed · 330,960 classified · 1,694,147 CANDIDATE. Database 5.8 GB. Sampled twice 45 s apart: purposes +1,500, categories +731, file count static. The moving counters are the proof the daemon works — a static file count alone is indistinguishable from a stall.

Not verified, stated as such

The osascript → Terminal dispatch has not run for this agent, because testing it means either rebooting or opening a duplicate session on the live transcript. It is the same code path the Tyrell and rdreceipt agents have executed for real, which is evidence about the script and not about this agent. Recorded as ASSUMED in SESSION-STATE.md §4.

Found while measuring, filed not fixed

ISSUE-20260831-14 — only 8,153 of 2,025,107 files are hashed and the number is static while classification advances. Dedup depends entirely on hashing. Three candidate causes recorded and undistinguished; not diagnosed. No dedup or reclaimable-space figure is quotable until it is.

How to undo

launchctl bootout gui/501/com.eastcoastscience.claude-resume-replicantdb 2>/dev/null
rm ~/Library/LaunchAgents/com.eastcoastscience.claude-resume-replicantdb.plist
rm ~/dev/fleet/ops/resume/PINNED-SESSION-replicantdb

The daemon agent predates this change and should be left alone.

Outstanding, Rich's

• ISSUE-20260831-10 — CloudKit container has never existed; needs the Apple Developer account.
• DEC-20260831-04 — 39% staging debris inflating reclaimable-space figures.
• Credential rotation. Time Machine holds pre-remediation state; deletion is not remediation.