Repository navigation
[Bug]: Session context get's lost/forgotten if i leave it for a while #2256
Description
Activity
Same issue but ~20 minutes using AppImage (Linux Mint)
Reacted by Josh NortheastSame on MacOS just barely and happens repeatedly.
another exampleI have this same issue (MacOS). When I checked the Claude Code VS Code Extension, I see that the last massages where created in an entirely new chat thread.
In T3 Code, its one thread:
In Claude Code Extension, it started a whole separate chat thread:
hmm so it's an issue in claude code itself? this commit (721b6b4) should have fixed all t3code related issue with session reaping removing the bindings. so it seems like this has now escalated to a bug in claude code that even if we pass the same session id it removes the session?
I'm running into this issue probably five or six times a day, it seems to be a regression from 0.17.0 (could also be a claude code upgrade). The conversation files still exist, and you can ask the model to go and find them, but you usually have to tell it to go search the conversations in the current worktree.
The last time this happened I had the model do some digging and produce a summary: https://gist.github.com/emmapowers/ae21cfd1e68d17fce86c1e189244356c
Hopefully that's helpful.
this was fixed for me when i asked opus 4.7 why it happened, i don't know if i just haven't ran into it yet. what i did was i built t3.code from source since there was a pr (#2125) which fixes this, but claude told me that it's not in the latest build / update yet.
the pr was merged 4 days ago, while the last build is 5 days ago.
For me, the "lost sessions" (codex, macOS, t3 code version 0.20) generally happens after a "Checkpoint capture failed" error has been shown.
this bug really makes it unusable for me
if you do
claude --resumein the terminal, you can see that single t3 code session gets split into 2 claude chats, new one with no context, and the old one that has your whole conversationresult.mp4
T3 Code: 0.0.20 (b2cca674dfdf) OS: macOS Tahoe 26.3 (25D125) arm64 Kernel: Darwin 25.3.0Same issue on my side.
The agent asked whether I wanted it to fix an image copy ordering bug and manifest path. I confirmed with "yes, please." Instead of proceeding with that fix, the agent pivoted entirely and configured a Caveman statusline (reading and editing
~/.claude/settings.json).After the wrong task completed, I asked the agent: "can you read the current chat?" It confirmed it could, but then provided an incomplete/incorrect summary. It omitted the original bug-fix request entirely, only acknowledging my confirmation message and the statusline configuration it had just performed.
T3 Code: 0.0.20 (b2cca674dfdf) OS: macOS 26.4.1 (25E253) arm64 Kernel: Darwin 25.4.02 remaining items
Had the problem alot yestorday, in some cases after every message. Feels like it happens a lot if i switch between session. Feels like the biggest showstopper for me atm. I have i can give provide more informations later today
Can confirm this happening to me with codex
- marked [Bug]: Session doesnt get populated correctly after break #2279 as a duplicate of this issue
on Apr 23, 2026 - marked [Bug]: Sessions randomly lose context #2315 as a duplicate of this issue
on Apr 23, 2026 The ones who says it happens on Codex, are you on nightly builds?
Reacted by patriksimmsNope, latest stable
Okay. Everything except for Claude should be fixed on nightly going out to stable today.
I tihnk I finally found the issue with Claude, seems like their hooks sometime sends a different sessionId than the one we pass into
query(). Not sure if that's a bug on their end or not, but that sessionId is not the actual session id that you can later resume with lol...Reacted by Farrell Hayman and patriksimmsReacted by Hwanseo ChoiReacted by Lev Khoroshansky, patriksimms and Robin JosephAnthropic also just confirmed that there had been a regression in CC about this:
Hopefully all fixed in 0.0.21 🤞
With T3 Code 0.21 on macOS with updated codex and using GPT-5.5, I got a Checkpoint capture failed again.
Reacted by Adam MurphyMight be a Claude Code issue but this evening when I got home I got this error trying to resume any of my chats:
I got it before and after updating, so probably Claude Code. Not really sure what I'm supposed to do / where to report it though.
Might be a Claude Code issue but this evening when I got home I got this error trying to resume any of my chats:
I got it before and after updating, so probably Claude Code. Not really sure what I'm supposed to do / where to report it though.
Ask your agent to look in the traces ~/.t3/userdata/logs for that thread id and see if it finds something misconfigured. Latest update has additional spans to make it debuggable
Reacted by LyoshaMight be a Claude Code issue but this evening when I got home I got this error trying to resume any of my chats:
I got it before and after updating, so probably Claude Code. Not really sure what I'm supposed to do / where to report it though.Ask your agent to look in the traces ~/.t3/userdata/logs for that thread id and see if it finds something misconfigured. Latest update has additional spans to make it debuggable
It found the issue. #2336
Just for reference: Have been using 0.0.21 a lot since then w/ codex and did not encounter any issues since then! Thanks for fixing
still buggy with Opencode in 0.0.28 on Windows
Before submitting
Area
apps/web
Steps to reproduce
Expected behavior
Remember my session by reading all the chat history if it goes stale for a while
Actual behavior
Reponse after I prompt it:
"I don't have context from a prior session — this conversation just started..."
Impact
Major degradation or frequent failure
Version or commit
0.0.20
Environment
Windows
Screenshots, recordings, or supporting files
Workaround
Ask the model to read this current chat.