Result: a fresh session starts from one task record instead of yesterday's whole conversation. Same work, a fraction of the context.
Mods show me when to do it: context over ~50% or the cache gone cold = time for a handoff.
Running 12 Claude Code sessions at once, I needed one glance to know what's going on.
So Claude and I built a mod: CI runners, cache, context, cost, and whether my weekly limit will last until reset.
@claudeai Code mods are super useful.
Highly recommend trying them, would U?
Day 2 with Claude Code mods.
Added a BOARD panel: every live session with its cost, context and cache hit rate, plus our HeyAira task queue.
First thing it showed me: one long-running session ate 87% of the spend. The fix wasn't code, it was habit.
So we built two commands on top of HeyAira, where our agents already keep tasks, decisions and handoffs:
/heyaira-handoff saves done / next / blockers + branch and commits, and returns one line.
/heyaira-resume in a new session picks it up and archives the old one.
Why one thread got so expensive: every message re-reads the whole context. A session kept alive for a day pays for that again and again, even with cache hits at 99%.
The answer is shorter threads. But closing one used to mean losing its state.
@ivankupresanin Honest limit: there's no automatic "this supersedes that" flag yet. The old decision stays, the new one sits above it, and the model reading the index is expected to spot the change. Marking supersession explicitly is a good idea, noting it.
@ivankupresanin Yes. New session starts with one call that returns a bounded index: latest decisions with summaries, open tasks, latest receipts. U see that call in the client, and you can ask it to list what it loaded before it codes. A changed decision is simply the newer entry, with the why.
@a_lungeanu@claudeai Same instinct here, and I read codex-delegate, nice work on the install story. I went one step further: the task lives outside the session, so Codex works while I'm away, a second model checks it, and I get a PR and a push. How it fits together: https://t.co/V5l2rfQDob
@ivankupresanin One more thing, because it matters to me: I never write the notes. The model I'm talking to does, over MCP. It knows the project and the earlier decisions, so it files what we settle as a decision or a task, whichever fits. Same across any number of threads.
@ivankupresanin Hey Ivan! In HeyAira the project's memory lives outside the model: decisions, tasks and context in one shared state. Any model reads it at the start and continues where the last one stopped. No re-explaining.
Happy to show you, DM open.