Post

--resume

Claude Code gives you a command to re-start a given conversation, I'll save that into a bash script named restart_claude.sh or something.

Not sure if other harnesses have a similar option but I assume they all will at some point. This is pretty important

Yeah let me second herdr, it's quite nifty

Give the sessions memorable names and remember that you can just exit and resume later and it will be indistinguishable Unless this is more of a "those are my emotional support browser tabs" situation

check this shit out github.com/david-crespo...

I think this is the right approach. Context should be ephemeral. The stored history is documentation, git log, ticket summaries, and the code.

i made a gui that's ~claude desktop but for claude code sessions, in order to avoid that problem

simply create an eighth window responsible for ordering the other seven windows

For "research" stuff I just use the Gemini web interface. For "development" each project has a devcontainer that runs on one of my Linux boxes with an antigravity CLI in it that I access via a terminal in VSCode.

Now I do, because lots of ppl informed me :)

Ask your agent for a "takeaways" file for anything you might revisit, as if you're going to resume on a different thread, and then save it with an appropriate name and date.

not automatically; I believe there's a hash that I could use to restore, if I knew where it lived and had an index — hmm I could probably create instructions to this effect in CLAUDE-dot-md ...

each task has a session, and those sessions can get picked back up whenever by re-activating the card

I use kandev for an agent integrated kanban board. It works ok but subagent/subtask management is not flexible.

humanlayer stans rise up

coming to an uncomfortable realization myself that the better coding harness are kanban-shaped I can just paste a link to a PR and the harness sets up a new session with the context for doing review and writing feedback comments, then that session goes away

i built an obsidian clone where agents have their own journal space and per-project notes. markdown backing with a sqlite index for hybrid search. i know a lot of people working from something similar

Ah. CC has in the past given me a hash code and said "you can use this with /resume," but I never thought to try /resume just cold -- so never got the list. This will help.

Forgetful github.com/ScottRBK/for... for atomic memories related to projects and entities, plus plans and tasks with acceptance criteria. Then plain old markdown for day logs and ideas and docs it can load back up in a future session.

cmux plus every session writes their plans into markdown in version control. Coordination goes through a CRDT-based issue tracker, kind of like the kanban board mentioned sidethread.

I tweaked my Pi prompt to show an auto-updating convo summary so I can whack-a-mole more easily, then I just treat it like a messy desk system and prune stuff if it gets too cluttered. Also, cmux. Tabs are cool.

Paseo.sh

Do note on default config it deletes after 30 days and you should change it so it doesn't.

or IDE plugin

Here's what it looks like in another repo I use it in: github.com/gitronald/ci...

Starting a new session/conversation on $topic (like marketing, this past client project, building that spite fork of that thing...) <resume command> <project pet name> And tell the agent "Hey, orient from handoff / notes / last steps, tail the transcript..."

If I have a conversation I really want to save insight from, at the conclusion I ask the agent to create a copyable markdown formatted block so it's easy to export the essence of what was discussed; that usually works well

Obviously I don't. But also I get them to write notes in my obsidian and I often resume sessions.

in theory my agent system can load these back up but i've never actually bothered to do that

I just archive all the sessions to my NAS if I ever need to reference or search for something, and the recent ones I can just resume when needed directly from their project folder

Use one codex session to rule them all. Those sessions and agents should only communicate with the coordinator session.

Sometimes I open a new terminal window and tell myself this one will be different.

heck yea! a lot of harnesses have this baked in somehow (afaik claude code has a search if you /resume or --resume) the solution here can have nice knock on effects

Domain-specific folders, each with a MEMORY.md file

You can always resume a long-closed session and pick right back up where it left off — context and all. It tempting to always leave them open like browser tabs though

at this point, Claude Code takes so many notes into memory that I don't think any sessions as hard boundaries anymore. I can load up something new and be like "okay where were we?"

i don't

I'm liking paseo lately

i use zed.dev agent panel. so anything that's a come-back-to i leave in the thread sidebar. can be named to remind me wtf it was about too (e.g., have an older onw rn that's just called "fix logseq sync") one thing that'd be cool (cc @zed.dev) is being able to move it btwn workspaces though.

I start a new session and say "what have we been working on?" and it dives into previous sessions with memory tools with sql and polars. Also session/load to talk directly to past agents, but having everything in sql really makes persistence trivial

I think with the brain in my head

How do you do it without?

(we do the same thing with slack, linear, notion, etc and let the agent use tool calls to search them all)

there’s both a fully agentic “ask” workflow with a well-tuned prompt that takes natural language questions via slack/cli, and the MCP server exposes raw retrieve/expand calls if you want your agent to drive it

My point was more like, imagine Alice talks to the model about something. Alice explores solution A. Then Bob later talks to the model. Just coz Alice had already talked to it, Bob doesn't start with a clean slate, his design is informed by better famliarity with solution A

This is probably the same as having dozens of browser tabs open for days: it's partly about bad tools and partly the endowment effect. True enlightenment comes from knowing the tabs you have open are no more valuable than the ones you have closed.

Handoffs & gh issues. Chats are always disposable and state is persisted on request

They're nice guys haha

I like it more than just -c/-r bc I can get a fresh window but keep the details