Opus 5.5 with a 1M window in Hermes Agent is brilliant, until every turn re-reads 800K tokens of old chat and your Claude limit is gone by lunch.
Here's the compression setup I landed on. It keeps Opus sharp and takes the summary work off your Claude plan completely.
What I learned:
- The 1M window costs nothing by itself. What burns allowance is how much context each turn re-reads. Compress early and Opus stays fast and focused.
- Set the trigger per model, not globally. Opus compresses at 25% (about 250K). GPT models at 85%: about 231K on the 272K route, about 765K on the -900k models.
- The summariser doesn't have to be Claude. I use GPT-6 Luna on high through my ChatGPT plan, so compression never touches my Claude limit.
- Use gpt-6-luna-900k, never plain gpt-6-luna. Hermes won't let a summariser read more than its own window, so plain Luna's 272K quietly pulls every bigger trigger down to 272K.
- Keep lean mode. One summary call, your messages kept word for word, and anything summarised away stays searchable.
No ChatGPT plan? Put Sonnet 5 [1m] in as the summariser instead. Same idea, it just comes out of your Claude allowance.
Copy and paste into ~/.hermes/config.yaml:
compression:
enabled: true
threshold: 0.50
threshold_tokens: null
tail_mode: lean
in_place: true
model_thresholds:
"claude-opus": 0.25
"gpt-6": 0.85
"gpt-5.6": 0.85
auxiliary:
compression:
provider: openai-codex
model: gpt-6-luna-900k
reasoning_effort: high
timeout: 300
Why Hermes has to own this: the Claude subscription plugin turns off Claude Code's own auto-compact, so Hermes' trigger is the only compaction that runs.
Docs and the plugin:
Docs:
Hermes compression: https://t.co/1BhrEZuWZl
Hermes config: https://t.co/7XppP9bFIo
Claude subscription plugin: https://t.co/5ig7SbmIds
Claude Code auto-compact window: https://t.co/Q35dj8uSSo
Claude Code costs: https://t.co/xxJA5OqbgQ
Which is really a shame, that we have to make these choices. Maybe as prices fall and some companies (ahem) are less constrained for compute, we can purely base these decisions on who has the best model for our needs right now.
@LatentOperator That's exactly what I ended up doing and just added a $20 Claude code to dabble a bit!
GPT rocks. I've been using it for so long now. I've built tools for it. I love Codex.
Only made sense to go that route. And I have 3 banked resets to break it in!
@tonysimons_ If I can only buy one, I'd go for ChatGPT. I think it's the better overall value and the resets and lack of a 5-hour window help a lot. I do both and can justify it with billings, and I think that's the real sweet spot. Have both and leverage each.
I did it and it's genuinely been nice to work with Opus 5.5 on some coding and project planning stuff. But it's not the night/day difference between it and Sol some people say. And I've had "I should be straight with you" type incidents at least twice.
Against my better judgement, I’m thinking about it.
This poll will make my decision for me.
Do I eat crow and resub to Claude now that it’s back in Hermes Agent?
I should elaborate that most posts from my son's school come in the form of canva style graphics🤷♂️ without any raw text. Consistently amazed at how my Hermes agent can pull dates, times, event info from images and keep it all straight.
@tonysimons_ I have a Hermes agent that just monitors communication from my kid's school and turns it into calendar items and tasks and shares them with me and my wife.
Welcome back to Hermes Agent, Claude
New official plugin that uses Claude SDK without the tradeoffs to enable Claude Code subscriptions to work in Hermes Agent again!
Check it out and install it here: https://t.co/r5YghWTwiY
I threw it on a little Surface Go and was blown away. Didn't even need the Surface kernel. Everything worked: keyboard, touchscreen, trackpad, wifi. All things that haven't in the past on Surface.
@MichaelGannotti Can't wait to see what it shall feel like on the Microsoft Surface I refused to buy precisely because I feared it would suck at Linux.
@OmarchyLinux's totally reshaping the computing landscape accross architecturesK
Who knows if there won't be a PDP-11 version ...
Obsidian doesn't get enough love for keeping systems / agents in sync. I use @obsdmd to keep detailed configs and setup guides so agents keep five different machines in sync across OS, apps, and prefs.
@tonysimons_ I have a Hermes agent that just monitors communication from my kid's school and turns it into calendar items and tasks and shares them with me and my wife.
So Omarchy + @NousResearch is a gift. Have a custom plugin in Hermes for a daily GDT-style "punch list" of tasks populated from current projects (agents manage). Now I have a plugin for the punch list with pomodoro timer directly in Omarchy. Beautiful.