[BUG] Munder Difflin v0.5.2 — condense subsystem kills the app during idle compaction + corrupts agent notebooks

[salem3qh]

What happened?

We are running Munder Difflin v0.5.2 (packaged, Windows 11 Pro, Electron 32.3.3) as a production multi-agent floor, and we have isolated two critical bugs to the same subsystem: the context condenser ("condense") that compacts idle agent conversations.

Bug 1: The app process dies silently during idle compaction

• The whole app process tree exits with zero OS trace: no crash dump, no Windows Error Reporting entry, no app-hang event, no reboot/sleep (desktop machine, continuously up), and the app writes no exit event of its own.
• We correlated 13 outages in hive/log.jsonl: the final minutes before nearly every death are bursts of "kind":"condense-abort" events (one every ~40-60s). In the latest death (2026-09-11 08:00:07Z) the last event ever written is a condense-abort; in an earlier one (2026-08-28 15:14:55Z) the last event is a successful condense.
• A 1-minute external process monitor pinned the latest death to 08:00:02–08:01:02Z — top of the hour, while the machine was idle and the user away. Compaction runs on idle, which matches.
• condense-abort volume is escalating: 11/day (Aug 27) → ~350/day (Sep 2), ~300/day since. Total 2,216 in our log.

Bug 2: Same subsystem corrupts agent notebook files (memory.md)

7 documented strikes on agents' memory.md. Byte-level signature is identical each time: [valid 3,352-byte header] + [stale old snapshot of the file] + [tail of the newest appended block] — i.e., the newest write is applied on top of an old snapshot of the file instead of the live file, and the file shrinks (68,859 → 41,342 across strikes). Strike 7 was stamped 2026-09-10 12:00:37 local while no session of that agent was running at all — only the harness's idle compactor was active.

Bug 3 (minor): child processes are not relaunched on app start

The app launches our Telegram bridge process once but never relaunches it after an app restart. We work around it with an external watchdog.

Evidence available on request

• hive/log.jsonl (condense / condense-abort timeline, app-start events)
• hive/bin/munder_monitor.log (external minute-by-minute liveness log capturing a death)
• Byte-level forensics of the notebook corruption (headers/tails/sizes per strike)

Workarounds in place on our side (external watchdogs to relaunch app and bridge), but the condenser itself needs a fix at the source.

Thanks,

Steps to reproduce

Non

Screenshot or screen recording

Non

Logs / stack trace

Operating system

Windows

OS version

11

Munder Difflin version

0.5.2

Node version

Lastest

Agent CLI and version (if relevant)

No response

Pre-flight

• I attached a screenshot or recording above.
• I'm on the latest release, or I've said above why I can't be.
• I re-ran npm install so node-pty is rebuilt for the current Electron ABI (source installs only).
• I searched existing issues and this isn't a duplicate.

[salem3qh]

Update: 12th notebook corruption strike (same byte signature, fixed 3,352-byte header now 5x in a row), 2026-09-11 ~21:15-21:57Z — no app restart inside the window, so the idle condenser corrupted the file while the app was alive. Also see #459 (condense payload fails JSON escaping) — likely the same root cause.