Downstream production report: 27 defects still present on main (each verified against fb2db8ee1, with fixes and tests)

[Dmatut7]

Hi — we maintain an independent downstream line of prime-agent (started from 0.9.1, merged your main three times, now 325 commits ahead of our merge base at d74a75f, while you have moved on to 0.9.4). We run it in production every day, and along the way we fixed a fair number of defects.

We went through everything we fixed and checked each item against your current main (fb2db8ee1), line by line. This is the result: 27 defects we can still see on main, each with your current file:line, a root-cause chain, and the fix + regression test we run on our side. We deliberately excluded anything we could only grep but not reach, and the doc lists what we checked that does not apply to your tree (Appendix A), so the signal stays clean.

The full report: https://github.com/Dmatut7/prime-agent-rlm/blob/merge/repl-kernel/docs/fork/upstream-feedback-20260911.md

A few examples of what is in there: a failed client catch-up is never retried or requeued (permanently stale views); one stray unhandledRejection exits the daemon and every hosted session; startup readiness waits on adopting all workers including 24h-timeout creates; several unbounded waits on the agent-message path; transcripts written with the default umask. (We first tried issue #2238, which the vouch bot auto-closed; the three items from it are restated in §2.13 so they are not lost.)

For every item we already have a commit, a red-first regression test, and a changelog fragment on our side, so patches are cheap to produce. Per your CONTRIBUTING we are not opening PRs uninvited — happy to proceed however you prefer: a discussion per cluster, patches as diffs, or PRs if you decide to vouch us.

Nothing here is urgent on our side; we run our own line and our users are not blocked. Sharing because re-finding these costs time and we already spent it.

[kevinjosethomas]

Hey, just saw your comment on the cleanup PR stack. Thanks for contributing all the work you've done upstream! We have a long list of improvements and reliability work and new features we're adding to Prime Agent, so hopefully you won't need to maintain that soon :)

Seems like you have a lot of improvements here and we'd love to slowly add them in. I'm pretty swamped with our current workload, but would appreciate if you can create separate discussions for the most important bug fixes and reliability improvements that you have noticed so far, and then we can tackle them. Thanks so much for contributing back to prime-agent!

I can add you to the vouched list after you make a few of the initial discussions, and would love to get more of your contributions!

[Dmatut7]

And one thing that is not a bug report, from the person who runs this downstream line every day: after adopting Prime Agent he stopped using Claude Code and Codex entirely. The persistent-REPL + subagents + daemon-resume design is why — it is the first agent setup where long multi-agent work actually survives real life.

We file all these reports because the base is worth building on. Thanks for moving fast and for being open about it.

[Dmatut7]

Quick follow-up — the focused discussions you asked for are up: #2261 (supervisor: catch-up retry / adoption off the readiness path / unhandledRejection), #2262 (agent-message delivery: unbounded waits + idempotency), #2263 (kernel: brick-on-death + watchdog liveness), #2264 (long-run resource growth cluster). Each item has the current file:line on main plus the fix and regression test we run downstream.

And nice catch on the stack: we noticed the waitForIdle fix landed in #2249's tip (3e96f9e). That closes the one hole we flagged.

On vouching — you mentioned adding us after a few discussions. No rush at all; just tell us what else you'd like to see first. The moment we're vouched we can start sending focused PRs instead of prose.

[Dmatut7]

Second round, same method (every item re-verified against main at fb2db8ee1, file:line included, nothing carried over from our notes without a fresh read):

32 more defects we can still reproduce on your tree, in 16 clusters — highlights: the compaction budget math overflows the provider's real input cap on CJK-heavy sessions (chars/4 undercounts by ~1.6×; production session hit the 400 wall 14 turns in a row), the /share secret preflight scans base64-encoded bytes so plaintext key patterns can never match, temp files land world-readable (0644), the orphan-process journal has no compaction, and one test fixture hard-links the signed node binary (on macOS that kills the inode's code signature for the whole machine — every node exec afterwards is SIGKILLed; took us a day to find).

Full report: https://github.com/Dmatut7/prime-agent-rlm/blob/merge/repl-kernel/docs/fork/upstream-feedback-20260913.md (fixes + regression tests for each, on our line).

Also Appendix A lists what we checked that does NOT apply to your tree (so the list stays honest), and §3 has four defects neither of us has fixed.