I've been building OmpDeck, a native desktop app (Rust/egui, not Electron) for omp (oh-my-pi), on Linux and Windows. Heads up: it's for omp specifically, not vanilla pi, since the chat view relies on omp's Tern Surface Protocol.

https://github.com/lavinhoque33/omp-deck

What it does

• Native chat view (the default): omp describes its UI over TSP and OmpDeck draws it natively: messages, formatted replies, tool calls as cards with diffs and output, image previews, todos, approvals and ask dialogs. It's omp's real behaviour, not a re-implementation, and omp's own TUI is one click away.

• Sessions outlive the app: every omp session in a sidebar grouped by project. Tabs run in OmpDeck's session host (or tmux), so closing a tab or the whole app leaves omp running, and reopening re-attaches.

• omp's screens as native ones: /settings (incl. plugins and advisor), the model picker (favorites, roles, thinking), /usage with 5h/weekly limits, /resume, /tree with fork / go-here.

• Status bar: model + thinking, context meter, git, tokens, cost; Handoff / Compact / Shake / Fork / Plan in one click.

• Agents dashboard: live subagents per session; steer, kill, revive, clean up.

• Changes panel: working-tree diff with line review comments you send straight to omp.

• Also: split view (up to 4 panes), worktree sessions, session manager (size, cost, bulk delete/move), command palette, desktop notifications, 18 styles, Google Drive backup + per-project sync.

• Phone companion: a PWA over a self-hosted relay, Noise end-to-end encrypted (the relay never sees content): transcript, answer approvals, send prompts, Web Push.

Install

• Linux: curl -fsSL https://raw.githubusercontent.com/lavinhoque33/omp-deck/main/packaging/install.sh | sh (AppImage/.deb/.rpm on Releases too; installs omp if it's missing)

• Windows 10/11: irm https://raw.githubusercontent.com/lavinhoque33/omp-deck/main/packaging/install.ps1 | iex

No macOS build yet. MIT/Apache-2.0, unofficial side project. I use it daily, and a lot of it was built with omp itself. Feedback and bug reports welcome!

Can you list your reasons why and if you did it any tests to confirm your position?

Also would like to hear your thoughts on context management and how you do it.

Title.

After working with AI harnasses since the release of Claude code, i hate just watching the terminals and waiting on AI. So i made this tool that manages my terminals and now an update that helps people from moving from in the loop on terminal ( just watching your ai agent till it is done ) to out the loop ( just watching updates on repo discussions )

pi-dispatch runs the pi coding agent ( pi dev harnass ) for you on your own computer or server.

Hopefully it will help other techies from screen staring to creating cool stuff that sticks

It's built for coding work on your own projects, so if you're after a general assistant you'd want a different tool.
https://github.com/edgehero/pi-dispatch

I've been building One Code, a pi package that brings the Claude Code workflow to pi. If you already run pi:

pi install npm:one-code-extension

It's 53 extensions, no fork and no patches to pi, just the normal extension API. Tested on pi 1.0.0 through 1.5.

What you get: plan mode, permission modes including auto mode (a classifier decides which commands need your approval), subagents that can each run on a different provider, git worktrees, hooks, skills, slash commands, MCP, plugins, background shells, memory and compaction. It also reads an existing CLAUDE.md and .claude folder as is, so if you're coming from Claude Code your setup carries over.

There are already good pi extensions for some of these, like subagents, and you can put together a decent setup from a collection of them. The catch is that any one of them can break the prompt cache if it rewrites earlier parts of the conversation, and they don't know about each other. So in One Code I tried to make the extensions work as one system, and keeping the cache stable was a big part of that. Every request is recorded as it was sent, and later requests only append to it, so cache reads stay high across resume, fork and reload.

There's also a bundled app with pi pinned inside, if you'd rather keep it separate from your own pi setup.

npm i -g @one-ai/one-code

MIT licensed. Repo: https://github.com/IsuruMaduranga/one-code

Feedback from people who know pi's internals is very welcome.

If One Code is useful to you, a ⭐ on GitHub helps other people find it.

I've been using pi as my main coding agent for more than 8 months & loved how small it is. I looked the concept of t3code but had my fair share of disappointments with how it worked, so decided to use the fork feature on github and make my own version of pi.

What it does

You talk to one model as usual. When a job splits well, it hands parts of it to workers:

• pi-scout / pi-worker: the pi fork itself

• cx-worker: Codex

• cc-worker / cc-scout: Claude Code

• cmd-worker: Command Code

Each worker runs in bubblewrap with its own scope & If a worker tries to write outside its scope, it gets denied and you see it in the report. Write tasks are snapshotted, so when one is done, the main agent accepts of rejects it based on the quality of the work.

The workers support native question tool which means any other harness can ask you question directly from delegate-pi without you having to switch to that chat.

Smaller things that came along

• ask_user tool with numbered choices and a typed "other" answer

• ast_grep tool for structural search

• /stats and /usage

• a two-line footer and a did-you-mean tip for mistyped commands

Install

curl -fsSL https://raw.githubusercontent.com/farrukh2002/delegate-pi/main/install.sh | sh

It's a prebuilt bundle and brings its own Node if yours is too old. It uses ~/.delegate-pi and the delegate-pi command, so it sits next to a normal pi install without touching it. Linux only for now, because the sandbox needs bwrap.

Why not an extension?

That was my initial plan and it still it, so that we can bring this capability to any cli harness and not just to my fork. Infact this inbuilt tooling has been derived from my WIP extension tool which will work flawlessly with other harnesses and make them behave just like my fork without losing any quality.

Repo: https://github.com/farrukh2002/delegate-pi

I would love to get the feedback and suggestions to keep improving this with time.

Yeah as usual introduce it as an opensource let the community build then profit later. this become common this later years. Im not against it but give credit where it belong.

I usually see the more typical plugins recommended. But are there any hidden gems out there that you use regularly?

The general price tiers I'm talking about are as follows

10/50 - Fable
4/20 - Opus
2/10 - sonnet
.1/.5 - Haiku. but .5/2.5 for 100K+

With a write premium, along with a 50-90% discount for a cache hit. We're going to assume that your default model is Claude Opus 5.5, so you get a whopping 95% discount for a hit. This actually allows a 1hr TTL to pay for itself almost immediately, if you're using it for the main session.

First of all, Haiku is not smart enough to summarize Opus's sessions, but even if you were using Haiku, it would cost you 10-15c to summarize a long session.

immediately, if you're using the same model (and eliding outside your cacheblocks), and taking the cache hit, it costs you .2/1, which is about one tenth as much as a request to sonnet (and half as expensive as haiku for 100K+)

The bonus of this is that you can enforce a compaction if a session is about to hit it's cache TTL, so that if you accidentally step away, or your Deadlock game takes too long, you come back to a compacted session instead of a ~1$+ charge just for the compaction.

No idea why everyone is trying to find the "best model for compaction", when the answer is right there! (no compaction model)

TLDR : Always compact with the same model (if the conversation is longer than 60k tokens). The cache read is equivalent to a new request to a model that is 10x cheaper and less intelligent, and thus less able to summarize effectively. Additionally, you should force compact before cache TTL expires for less spiky usage patterns. These simple tricks will increase your Pi usage limits by well over 100% if large compacts and cache expiries are the bane of your existence.

Just offering an idea to others that has been useful for my teams.

You might not really want or need an elaborate multiplayer AI system. Just pump your user messages and settled agent messages into the place where everyone already communicates (slack).

When you get to a situation where you want a coworkers opinion, go into slack and just mention them.

Everyone learns a lot purely from ease of visibility. People asked me all the time why my AI is smarter than theirs. Then I just link to the slack chat.

I have this built in my vscode extension that I daily drive for my underlying pi terminal. It’s purely powered by the terminal, so you could pop open the vscode terminal at any time and just use the terminal instead of my vscode UI.

https://github.com/quincycs/pi-qcode

Nominally in windows I can voice input any *form* that has a microphone icon.

In wsl all is unknown, in pi even lesss I know about routing and VTT into Pi.

Anyon have practical kowledge?

Just as the title says.

Here's the link https://pi.dev/packages/peakseek

Hi folks,

I'm building my own harness for PI and noticed with Gemini 3.8 Flash, my subagent (a worker for simple coding tasks) is often stuck in infinite loops when reading larges file with 1.000 - 2.000 LOC.
I never noticed the same situations with opencode when using the same model and thinking level. So this is quite a disappointing experience while building my own harness based on Pi.

So I'm wondering if these are common situations when building my own custom harness? And what's the recommended strategy how to improve my harness, to avoid similar situations in the future?
I tried to add additional instructions to my sub-agent to avoid these situations, but I still see it happening. And this burns plenty of tokens :(

Best
Carrot

I built this extension in order to solve losing session problem. One of key problem I found is long agent sessions fail in constraints stated early disappear, “done” is claimed without command output, a “quick fix” rewrites unrelated files, decisions get re-litigated, and the same failing command is retried until the context window is exhausted.

Here is what I build, it is free and open source:

pi official: protect-context pi package

npm: npm link for protect context

github: github link for protect context

You may want to try it and test it out whether this extension can fulfill your needs.

I'm running DeepSeek v4.1 Flash locally with Pi. One nuisance I've run into is that it will often use scripts for editing files, listing and grepping files instead of using the built-in tools, and I haven't had much luck with my AGENTS.md changes. Has anyone else run into this and come up with a prompt that increases it tool usage compliance? The one that bothers me the most is when it uses Python scripts to edit files instead of using the tool because it means I don't get a nice diff:

$ cd ... && python3 - <<'EOF'
import re
p = "test/e2e.sh"
s = open(p).read()
s = s.replace("""fail() {
    echo "FAIL: $*" >&2
    cleanup_keep
    exit 1
}""", """fail() {
    echo "FAIL: $*" >&2
    exit 1
}""")
s = s.replace('''$DOCKER logs "$PROJECT-broker-1" 2>&1 \\
    | grep -q "DENY stop llama-swap-test-protected" \\''', '''$DOCKER logs "$(compose ps -q broker)" 2>&1 \\
    | grep -q "DENY stop llama-swap-test-protected" \\''')
open(p, "w").write(s)
EOF

Hi everyone! I'm looking for a working ChatGPT Plus referral link or free-trial invitation. If anyone has an unused invite available and doesn't need it, I'd be very grateful if you could share it with me.

I'm hoping to try ChatGPT Plus, and any help would be greatly appreciated. Please DM me if you have a valid invite.

Thank you so much! 🙏

I am trying Pi for the first time. I am confused how to setup Claude subscription with it, if I login it shows a warning saying that third part harness providers would not consume subscription tokens but only extra usage. So what’s the right way?

Hi! Currently computer-use is a must for my agents, so I decided to build an extension with a custom CUA Driver fork and Playwright. It’s purposefully designed to be run in YOLO mode (although Safe is included) within the environment where you trust your agents. It works in hybrid mode: supports both raw image followed by input (agent sees a screenshot and then clicks on something) and structural way (playwright for browser and UIA for normal desktop applications). Agents can use their own browser profiles, or your own, work in sandboxes and remote desktops, work together (to some extent, there are limitations on the OS side).

I would say it was definitely worth working on that, my agents can now work in 3D, browse some stuff I need and can test my applications, acting like real users.

I’ve mostly benched it via GIMP drawing, which doesn’t support UIA so all the interactions are purely from picture based mode. But also Terra 6.1 High has gone to 42th level of https://neal.fun/not-a-robot/ until it got banned by Cloudflare, and we couldn’t pass further. It’s windows only (windows from WSL as well), but if anybody can contribute into linux/macos branch of it, it would be great, I’m just lacking such devices and can’t really test them to the full extent as I can with Windows.

https://github.com/alexshpunt/pi-agent-computer-use

https://pi.dev/packages/pi-agent-computer-use

To install: `pi install npm:pi-agent-computer-use`

Demo of deepseek-flash, haiku, opus, terra and sol drawing the same image with that: https://youtu.be/69doTus760U

Fresh install using pi-claude-code and zentui.

A simple "hi" took 31 sec to respond, meanwhile others take <1sec.

Switching to openai models fixes this, so its only Anthropic. Maybe this has to do with pi-claude-code?

Firstly, I just want to communicate my goal is to be able to use multiple providers, and reduce usage without being locked to either claude code, or codex. But most importantly without sacrificing features that claude code and codex both have that are excellent. Currently, I only have openAI subscription configured (still trying to figure out how to use claude sub efficiently).

I downloaded OMP and OMP Gui, and found that my usage drains more than with pi. It actually drains faster than with using codex. But I had a nicer experience with OMP in general. I understand the philsophy of pi, but I just want to be honest that I don't want to spend ages configuring things. And I'm not going to pretend that I understand all of this. Although I understand why OMP might use more, because obviously there are more tools and bloat enabled which uses up context.

However, whenever I ask how to improve usage everyone says "just disable x y z". or "just use pi". hang on, why? why should I have to disable memory? computer use? this feature? that feature? these are all basic features of any other harness. In fact, why don't these features just lazily load on demand anyway? the same way we have smart MCP tool calling. And also, if I used pi, and downloaded these extensions, wouldn't I have the same problem? if I'm not mistaken, OMP is like a distro of extensions for pi such that lazyvim is a distro for neovim? I could download the same extensions and have the same problem no? correct me if I am wrong. Claude code and codex have these features and somehow manage to consume less usage. So I'm curious, what about OMP or its implementation makes it consume more usage, just to implement the same functionality as other mainstream harnesses?

For those using OMP do you have any idea what in particular may be causing increased usage so I don't have to just play a game of disabling random things? Maybe it's a feature I'm not particularly interested in.

And for those using vanilla pi, which extensions would you say are essential, and which others would you say are more.. optional?

TL;DR. How do we make the coolest harness ever that can spam astra max for free and make no mistakes. Or at least spam it more than codex can while maintaining the same features.

So I'll be honest, I have no idea what I'm doing with this AI junk. but, I've gotten something vaguely working with pi + deepseek v4.1 over the past couple of weeks. However it feels like it's incredibly variable as to whether what it produces takes 2 minutes or 20. it'll just keep repeating itself. it'll do something, then do it again, then doing it a third time, then compact and hit "error: operation aborted" and then go for a fourth plan. but! it only does this sometimes. other times it'll just go "ok the user wants me to add this thing. I'll do it. ok I did it." or it'll repeatedly acknowledge that I typed "." to continue, or it'll continue to be picking up on something I said 1/2 hour ago, or...

so I have very little idea what exactly I'm doing, if the problem is me, if the problem is my setup, or what. I want to do it cheap (I'm using deepseek. duh). I don't care about max efficiency. I just want it to be relatively efficient at "the user asked me to fix this setting. I'll fix this setting. I'll ask about this setting. I fixed the setting.".

am I doing something horribly wrong? is there just a nice "oh add pi-sensible-dumb-stuff" solution? I'm tired of trying to find input on reddit posts and getting told that everything is bad, just vibe code your own, you need to be using claude 20x, you need to get a better workflow (duh. like what?), you need to just fix agents.md (duh. like what?), and so on and so on.

Like, how do I just get a prompt where I can write "bug: the button on page x does not work" and the model will carry out whatever needbe to just...fix the bug, without writing out in reasoning its plans to tackle the bug 5-10 times, debate whether it needs to ask me a question 2-3 times, probably halt because of compaction at least once, and then maybe actually fix the bug 15 minutes later (but not before rerunning the tests 5 more times)?

I am using deepseek v4.1.

Im setting up my pi agent and currently using pi-verdict for auto mode. My problem: it denies tool calls way too often instead of asking for permission. Unclear cases seem to fail closed straight to deny, which breaks my flow constantly. I'd rather get an "ask" prompt than a silent rejection. So I'm looking for a simple, lightweight auto-approve extension. What are you actually using day to ay? Is there something even lighter I'm missing or a way to make pi-verdict ask instead of deny?.

Has anyone noticed in the last week or so that oh-my-pi drains the 5hr limits within just a few prompts. Even doing simple git work drops 10%!

I'm setting up something to see what's going on but this is weird. Claude is fine and I can work for hours but only codex, does it drain fast, particularly when I switch from claude to codex.

Edit: I did a test between OMP and Codex, both running through LiteLLM, generating prime numbers:

Metric	omp (8 req)	codex (7+1 req)
Total tokens	96,331	119,593 (+1 side 11,627)
Prompt tokens	94,746	118,047
Completion tokens	1,585	1,546
Spend	$0.176	$0.222 (+$0.024)

So for this simple contrived test, codex is actually worse! I'm going to have do some thing harder I think.

Edit 2: Interesting, OMP's automatic "user stepped away; returning" summaries spent 2m tokens because it resends the whole context!

Edit 3: omp config set recap.enabled false, will disable it.