Repository navigation
Replies: 1 comment 2 replies
|
@YishenTu — a follow-up, because 2.2.5 turned out to matter more to this than a version bump usually does. Short version: the per-device session metadata in 2.2.5 broke my companion plugin silently, I've adapted to it, and the adaptation surfaced a design question I think is worth putting to you directly — partly because if you'd seen this project before 2.2.5 you might have shaped that change differently, or told me to stop building it. What brokeNew conversations file their record at The part that isn't a compatibility fixThe device key comes from a That's a coherent design and I can see what it buys: exactly one writer per record, which is a real problem to have solved. But it means "the same conversation, on both my machines" is now unreachable from outside Claudian, and that is the entire purpose of the thing I built. I've shipped what I could reach without touching your invariants: an opt-in setting that publishes a copy of this device's records into the flat layer, since your precedence is It works, and it has two costs I have to put in the settings text: the owning machine keeps reading its own copy, so metadata edited on the other machine doesn't come back; and on the other machine the conversation shows "Assign to this device", which if pressed hides it on the machine that created it. The questionIs there a supported way to have one conversation visible on several of a user's own devices? Options I can see, in the order I'd prefer them:
If any of that is interesting as a change to Claudian, I'm happy to do the work as a PR against your repo rather than keep it in mine. And the standing offerThe original question is still open too: whether you'd want this plugin inside Claudian. Nothing has changed about the honest part — it would move maintenance onto you, and this month is a fair sample of what that costs: 2.2.5 needed a compatibility fix, a design decision, and two rounds of measurement to get right. A "no" remains completely fine. Repo, if it's useful: https://github.com/powoct/claudian-session-sync — the design record (68 decision records with the alternatives and why they were rejected) and the real-machine findings are in |
Uh oh!
There was an error while loading. Please reload this page.
Hi — I built a companion plugin for Claudian and I'd like to ask whether you'd want it
in Claudian itself, before I keep developing it separately. A "no" is genuinely fine;
I'm asking because it seems rude to build something this tightly coupled to your plugin
without checking.
The problem it solves
I use Claudian on two machines, a Mac and a Windows box. The conversations themselves
live in the agent CLIs' own session files, so a conversation I start on one machine
can't be resumed on the other — the vault syncs, but
~/.claude,~/.codexand~/.grokdon't.What it does
A desktop-only Obsidian plugin that carries those session files through a local folder
you point at Dropbox / iCloud / Syncthing / whatever you already use. It deliberately
does not understand conversations — only the append-only structure of the files and
where each machine expects them to live. It doesn't turn sessions into notes, doesn't
touch credentials or CLI config, and talks to no cloud API.
Merging is prefix-safe: a file is only ever fast-forwarded to a version that provably
contains it byte for byte. Timestamps are never trusted. A genuine fork is detected by
content, quarantined on both sides, and handed to the user — nothing is overwritten
without a backup first.
Why I'm asking rather than just shipping
It only makes sense as a companion to Claudian, and the coupling is real: which
conversations belong to a vault is answered by your
.claudian/sessions/records, notby anything my plugin invents. That's a dependency I'd rather not maintain across a
project boundary if you'd prefer it inside.
The honest part
Merging would move maintenance off me and onto you, and that is a real cost, not a
gift. It tracks the on-disk session formats of several CLIs, and those formats change —
I've already had to re-measure Grok twice this month. Concretely, today:
fork / kill, on both platforms; two-machine resume verified.
they're written, so it's tiered per file rather than per provider.
no per-session file, no export path.
measure.
If that maintenance burden reads as "no thanks", that's a completely reasonable answer
and I won't take it as anything else.
If you'd like to look before answering
The parts most worth your time, in order:
docs/zh-CN/architecture.md— the design record: 61 decision records, each with thealternatives that were considered and why they were rejected.
docs/zh-CN/findings/— 12 real-machine measurement reports. Every provider's supportlevel traces to one of these; nothing is claimed that wasn't measured.
The design docs are in Chinese — that's a genuine obstacle if you'd want to own
them, and I'm happy to translate any part, or all of it, on request. MIT-licensed, same
as Claudian, so there's no licensing friction either way.
Happy to answer anything, and equally happy to hear it doesn't fit.
All reactions