Hacker Newsnew | past | comments | ask | show | jobs | submit | adyavanapalli's commentslogin

This is actually brilliant. I've seen other approaches where the context is selectively pruned, but never in a recoverable way.

P.S. Love the "bonsai" name btw!


ive built a variation that spins the chunks out as if they had been done by a subagent, and so the summary is like what the subagent returns

i couldnt figure out a nice enough way to reload the conversation, and didnt want to have to worry about the cache rebuilds so its a bit languishing


Honest question, but since when were biographies ever _authorized_? I would assume you would simply inform the subject out of politeness if at all.

It's a very common thing. A very unscientific quick search in Google Books led me to the oldest example I could find in a minute or two being a mention in The Dublin Review in 1844 reviewing a book on the correspondence of Edmund Burke talking about the loss of material that would have been indispensable for an authorised biography. It's very unlikely to be remotely the oldest use of the term.

And that mention also gets to why authors often seek authorisation and why subjects often authorise them: Authors gain access to the subject and/or often their material, while the subject or their estates gains some control over how they are portrayed.


I know what you mean, but there is a reasonable distinction to be made especially for living persons. An "authorized" biography would be something like what Isaacson does, where they get exclusive access, anecdotes, and importantly confirmation or denial of certain facts or stories.

An "unauthorized" biography is not necessarily suspect but it is kind of a subgenre. Consider the difference between Isaacson's Musk biography and the work done by Kate and Ryan for Character Limit (https://www.amazon.com/Character-Limit-Elon-Destroyed-Twitte...) - not exactly a biography but the legwork is fundamentally different, reportorial or even adversarial where the narrative gets challenged, as opposed to accepted by the "authorized" Isaacson.


Authorized/unauthorized simply describes whether the subject works with the author or not. It's just the common terminology for it.

Technically, a wikipedia page of a [living] person is an unauthorized biography!

But we'd like to think the crowdsourced editing process gives it some more trustworthiness than a random person self-publishing a book


If a biography is authorized, then you can interview the subject, and in some cases those interviews provide a lot of the material. Walter Isaacson's biography of Steve Jobs was like this, for example.

In addition to the direct interview, there are other advantages of authorization that often allow the author a much deeper access to research the subject:

- Access to records, correspondence, and journals in the subject's possession

- explicit endorsement from the subject encourages their family, friends and colleagues to work with the author.

- Opportunities for the subject themselves to fact-check and contextualize the research.

That must be balanced against the fact that authorization also often provides the subject an opportunity to review the content, giving them a hand in crafting the narrative and/or suppressing unflattering details.


This is soooo incredibly cool. Beyond the Tom Riddle diary aspect, I love the idea of this as a new medium for interacting with an LLM. You could gift it to someone and they could just write naturally, their thoughts, questions, notes, and get responses back without typing or speaking. It feels less like a chatbot and more like a journal you can communicate through. You could give it a personality and all.


I mean no offense, but it's kind of crazy that people think that linking handwriting recognition, a technology that was first rolled out 24 years ago in Windows XP, with an LLM is "soooo incredibly cool". I don't get it.


What really irks me about it is that the palm pilot I had 20+ years ago actually had BETTER hand writing recognition than the software on my devices today!


If by hand writing recognition you mean force you to learn their "graffiti" system, and write one letter at a time sure. Lets not even pretend newton did any better.


All I know is I have dsygraphia and their system worked for me. I took tons of notes on mine in school.


Products are something other than the sum of their technical parts.


It's cool but it has nothing to do with the tom riddle application.

It's cool because LLMs are actually fucking amazing technology and people are already numb to it.


Yeah you could just run rsync and sshfs.


For the edit tool, you should consider implementing a hash-based approach where each line of code is hashed and referenced by it when doing replacements. You can read up on the approach here: https://blog.can.ac/2026/02/12/the-harness-problem/

I didn't do much benchmarking, but anecdotally, I found it to be making less edit errors. YMMV


Yup, I used this for a while and IME it may get you a few percentages more of useful context initially, so quality feels a bit higher, but things start breaking down in funnier ways when you do run out of that quality for any reason later, so definitely caveat emptor.


What you're talking about is so extremely rare that it's much more likely that the entire Earth is destroyed by an asteroid right this inst...


It is not quite as rare. I calculated it to be less common than being hit by a meteorite, and added a section about that and the Birthday Paradox to Wikipedia, to the article about UUIDs. It got removed / replaced a few years ago however. (If my source was correct, there was actually a woman hit by a meteorite, but she survived, with a leg injury.)

If you do have a UUID collision, chances are extremely high that it's either a software bug, or glitch in the computer. It could be a cosmic ray. Cosmic rays messing with the computer memory or CPU are actually relatively common.


About as rare as an asteroid typing an ellipsis and clicking the add comment button.


Well, this joke dates back to (at least) the dial-up days where {#`%${%&`+'${`%& NO CARRIER


That’s just a result of jounce from localized gravity effects and atmospheric pressure disturbances in the moments before impact.

Think the ultrasonic typing hacking scene in Pantheon combined with the keyboard bouncing due to rumbling.


Well it would be statistically even rarer for that UUID collision to happen and the earth to be destroyed by an asteroid.


It's very common if you improperly seed, as others in the thread brought up! Or in your framing, as rare as earth getting hit if it were surrounded by a sci-fi density asteroid field.


For a single database using UUIDs, yes, it's astronomically rare. But it's quite a different thing to say that no computer system on Earth has ever experienced a UUID collision. The number of systems out there is also astronomical.


>The number of systems out there is also astronomical.

Not even close


Sure it does. Planets are astronomical, and we only have 8 of those in our solar system.


What are the numbers you are getting?


Initial drop, as people learn to use the tools, and while they keep babysitting their harnesses. Then significant boost once people start getting used to running the agents in the background, especially once they start running multiple sessions in parallel. I'd say you need a ~6 month push of getting people trained if they are not used to this way of working, and to customised setups etc. for your organisation, and then you start seeing significant payoff.


Was there previously huge backlog of work to do, or are you just building tons features for the hell of it because you can?


I've never once in 30+ years worked on a project where there weren't a lot of work we didn't get to because of lack of resources. So, yes, there's always been huge backlogs.


I had a chance to look at this and noticed you were sending telemetry to an endpoint you control: https://dirac.run/v1/event. It doesn't seem like you're sending anything obviously sensitive or doing anything in bad faith (though, I do see api errors being sent, which could potentially leak sensitive info), but you gotta admit that that's scary seeing you as the sole dev for this. Plus, it's opt out too. Sorry, it's no go for me.


Here's all the telemetry:

1. Telemetry to dirac.run/v1/event — Sends machine ID, token usage, model info, events, errors (first 500 chars), and platform info. Hardcoded API key. Defaults to opt-in (setting is "unset", not "disabled").

2. Feature flags from dirac.run/v1/event/decide — Polls every 60 minutes with your machine ID. Always enabled, independent of telemetry opt-out. No way to disable without code changes.

3. Web tools route through api.dirac.run — Web search and web fetch tools proxy through Dirac's own API server, sending your request content plus system headers (platform, version, machine ID).

4. Model list fetches — Calls OpenRouter, HuggingFace, Groq, etc. for model listings even when using the Anthropic provider.


> Web tools route through api.dirac.run

This is something that needs to be deprecated entirely. The web fetch tool no longer is used or works. There is nothing even listening at api.dirac.run. This was the result of me stretching my capacity too thin and bulk renaming cline.bot to dirac.run

UPDATE (+1h): both Web search and web fetch tools are now nuked.


Thanks! Since it is a Cline fork, the telemetry mechanism is inherited. I left it as it might help debug issues. There is no evil purpose behind it nor does it create or store any PII


I haven't tried it, but I'm curious why you decided to implement a whole new harness over just writing extensions in pi. From whatever I've done with pi so far, the extension api is quite extensive. Hash anchored edits, for example, can definitely be implemented in pi. Anyhow, thank you for showing us your project and will be checking it out later. Cheers!


A few months ago one afternoon I was very frustrated with how slow Cline was being so decided to look under the hood. Decided to make a couple of changes. Got sucked in. About 70k lines of change, another 40k lines of deletions and two months later, here we are.


The best kind of project. I'm trying this today. I've been happily using OpenCode so far.


I've been looking into local LLMs and new harnesses recently, how good is Pi compared to OpenCode, I'm seeing that it's a lot better? What are the best models and customizations for it to fully utilize it?


It definitely _could be_ an agent harness issue. For example, this is the logic opencode uses:

1. Agent is "plan" -> inject PROMPT_PLAN

2. Agent is "build" AND a previous assistant message was from "plan" -> inject BUILD_SWITCH

3. Otherwise -> nothing injected

And these are the prompts used for the above.

PROMPT_PLAN: https://github.com/anomalyco/opencode/blob/dev/packages/open...

BUILD_SWITCH: https://github.com/anomalyco/opencode/blob/dev/packages/open...

Specifically, it has the following lines:

> You are permitted to make file changes, run shell commands, and utilize your arsenal of tools as needed.

I feel like that's probably enough to cause an LLM to change it's behavior.


I saw this a couple of years ago and felt that it might help you https://youtu.be/k7X7sZzSXYs?si=d1ibZfR9uKbuXpCd. Best of luck OP.


Consider applying for YC's Fall 2026 batch! Applications are open till July 27.

Guidelines | FAQ | Lists | API | Security | Legal | Apply to YC | Contact

Search: