PH热榜 | 2026-07-31

@zaczuo Asad's re-render consistency point is the real test, not the first output. anyone can nail one clean take, the actual claim to check is whether the same layout survives when you swap one word for a longer one. that's the difference between a demo and something a branding team could actually rely on for repeat work.

curious whether commercial users are mostly generating final assets directly, or still treating this as a first-pass that gets touched up in an editor before it ships.

@zaczuo Congratulations on the launch!

What caught my attention isn't just the video quality, it's the ability to combine multiple references into one coherent output. Creative work rarely starts from a blank page. It starts from scattered ideas, assets, and inspiration that need to come together naturally.

I can see this becoming a huge time saver for marketing teams, agencies, and content creators who iterate constantly.

I'm curious, what's been the most unexpected use case you've seen so far? Did someone use H3 in a way your team never anticipated? Wishing you an amazing launch!

@ridhwikvinod Ridhwik, the honest answer on brand fidelity is that you should not be asking a generative model for it, and the good news is that you do not have to.

Exact hex and one specific typeface are the two things this class of model is structurally worst at, because both are discrete and everything inside the model is continuous. Asad is right that consistency across re-renders is the real test, and it is the test these models will keep half-failing for a while yet.

The pipelines that actually ship brand work do not fight that. They let the model make the motion, the atmosphere and the camera, then composite the brand layer on top deterministically. Logo, wordmark, exact colour, legal line, all placed by something that cannot hallucinate.

Which turns your question into a far more answerable one. Not can it render my typeface, but can I get a clean plate to composite onto.

So the thing I would ask MiniMax directly: does the API return an alpha channel, or a background plate with no burned-in text? That one feature decides whether this belongs in a brand pipeline or stays a concepting tool.

@shivvv hi help you

@shivvv Shivarchan, the distinction worth pressing on is whether the modalities actually condition each other, or just arrive in the same call.

Accepting text, image, video and audio at one endpoint is a convenience. Genuinely joint generation means the supplied audio drives the cut timing, the reference video drives camera motion, and the image drives palette, all at once and influencing one another. Those are very different capabilities sitting behind an identical looking API.

There is a cheap test. Feed it audio with an obvious beat and see whether the cuts land on it. If the edit comes back rhythm aware without being asked, the modalities are genuinely fused. If the cuts fall wherever they like and the audio is simply laid underneath, you have parallel pipelines sharing one endpoint.

Worth establishing before you build a workflow on it, because the second kind falls apart the moment you need the inputs to agree with each other.

@paul_from_dentro Paul, these are not really in the same category, and that is itself the useful answer.

Flux and Seedream are image models. H3 is video with audio generated in the same pass. If what you need is one still, an image model will almost always beat a frame pulled out of a video model, because video models spend capacity on temporal coherence that a still generator puts straight into per-frame fidelity. That tradeoff is structural rather than a quality gap somebody closes next month.

So the question that actually decides it is whether your output moves. If it does not, stay on the image models. If it does, the comparison you want is against other video models, not these two.

@levon_adamyan From the buyer side, the hard part of prospecting was never finding leads, it was trusting the enrichment. I've paid for lists where half the emails bounced or the decision maker had left months ago. How fresh is the data you enrich against, and do you flag confidence so I know which contacts are worth outreach? Bad data burns sender reputation before you notice.

@levon_adamyan For a founder-led sales team that’s currently juggling 2–3 tools, what’s the one workflow you’d recommend we rebuild first in Cleanlist to feel that 80% time-save fastest? And is there a tiny “success signal” where teams usually go, “okay, this actually changed our week”?

@levon_adamyan Congratulations on the launch, Levon! 🚀

I think most people assume prospecting is a data problem. From my experience, it's actually a workflow problem. Finding leads is easy. Cleaning them, enriching them, verifying them, and getting them into the CRM without breaking the flow is where the real time disappears.

That's why I like the direction you've taken. Instead of adding another point solution, you're removing the handoffs between tools. Those small context switches add up more than most teams realize.

Reaching 500+ paying customers also tells me you've solved a pain people are willing to pay to eliminate, not just a feature they think is nice to have.

I'm curious, once customers start using Cleanlist, what's the feature they end up relying on the most? Is it the AI Copilot, the 15-provider waterfall, or the CRM automation that quietly saves them hours every week? Wishing you and the team a fantastic launch! 🚀

@davemadeit Thank you. The CRM sync came straight out of customer conversations, so it is good to hear that combination lands. Appreciate you taking a look.

@odeth_negapatan1 Thank you. Babysitting CSVs is the thing we were most sick of ourselves, which is where a lot of this came from. We just want teams to focus on revenue generating tasks

@dalemooney Every enriched field records which provider it came from, and that's stored against the record. We can pull the source for any contact after the fact. It doesn't get discarded when the match lands.

Caveat worth being straight about: it's not surfaced in the UI today. You'd ask us and we'd pull it. Fine for a subject access request, less fine if you want to self-serve it, and that's on the list.

Lawful basis is a longer answer and I'd rather give you that in writing. Email support@cleanlist.ai and I'll send the DPA along with how we handle UK and EU records.

@dalemooney Victor's answer makes this more interesting, not less. The provenance exists. It is just not surfaced.

That is fine for the case Dale is describing, because a subject access request comes with 30 days and a human in the loop. It is not fine for the case that shows up sooner.

A 15-provider waterfall does not produce one quality of email. It produces fifteen. A verified match from the first provider and a pattern guess from the fourteenth arrive in the same column looking identical, and sending to both at the same confidence is how a sending domain gets burned.

That decision happens at send time, not over email with support. Which means the field you are already storing is the input to send-or-skip, and surfacing it turns a compliance obligation into a deliverability feature.

Is per-field source available on the export or the API today, even though the UI does not show it?

@doganakbulut That's the use case I'd pick too, old exports are usually where the most value is sitting and nobody wants to touch them.

On HubSpot it matches on email and updates the existing contact, so you're not getting duplicate records. Default is fill-empty-only, so it leaves alone anything your team has already put in, and you can set it per property if you'd rather it overwrite or append.

I'd run a small batch through first so you can see exactly what it's writing before you point it at the whole export.

@suryansh_tiwari2 Hey, just checked! You don't have an account with us and are actually one of the people/agencies who reached out promising upvotes which we declined

As this is fully false, please kindly remove your comment, thanks!

@suryansh_tiwari2 I believe your exact words after 2 times of telling you “NO” was:

“Going to do mass bot attack as well as Going to report to PH team”

————————

@product can we please take action against this user!

@sm0r3ll Clay is a good product, it's just built for teams who have a GTM engineer sitting there, and most people asking this question don't. We come in cheaper and you don't need to hire anyone to make it work, which is where the real money goes with Clay anyway.

Start with actually finding people, which Claude Code can't do at all. You search by title, seniority, location, headcount, industry and you get a real list back, not whatever the model manages to scrape off a page. Same on the company side.

Then enrichment sits on top of that, 15+ providers waterfalling until you get a verified email or a direct dial.

Then Smart Agents, which is the bit I actually like. You ask a question in plain English and it runs against every row, so stuff like which CRM are they on, are they hiring SDRs right now, does the site mention SOC 2, how well do they fit our ICP. It goes and reads the site and the job posts and fills the cell in.

All of that's on the API too, and there's an MCP server, so if you want to drive the whole thing from Claude Code you can, then push the result straight into HubSpot or Salesforce.

Claude Code is a good researcher, it just has no database underneath it and nowhere to put the answer at the end.

@sm0r3ll Steven, the useful way to tell these apart is not the feature list, it is who does the composing.

Clay is a workbench. You assemble the waterfall, wire the AI columns, decide the order, and it is genuinely powerful once you know what you are doing. The cost is that building the pipeline is itself a skill, and plenty of teams who buy it end up using a fraction of it because nobody on the team actually owns that job.

The pitch here is the opinionated version. The waterfall is already composed, you ask in plain language, you get a list. Less control, far fewer decisions to get wrong.

So the question is not which is better. It is whether you want a workbench or an appliance. If nobody on your team is going to own the pipeline, the appliance wins by default, because a workbench you do not operate is worth nothing.

@andrei_constantin_alexandru Salut!

Thanks for testing it properly and writing this up, genuinely useful.

For filtering, People Search has manual filters as well. You can set title there directly and get exact matches on QA Manager.

For the rest, hit us on live chat in the portal or support@cleanlist.ai and we'll get you sorted.

@zvonimir_sabljic1 Thanks, appreciate you checking it out. If you end up giving it a go, tell us what you think :)

@alex_watson2110 It stops at the first confident match. They're ordered by hit rate and cost, so most records resolve in the first one or two. The hard ones go deeper.

You pay one credit for a verified email whether it took one provider or nine. We absorb the rest. And we only charge when something comes back, so an unfindable lead costs you nothing.

@faruk_kamcici Really like the Compact Health idea - quiet until something's actually wrong is exactly right.

Quick question: your description mentions "temperature and fans" alongside CPU/memory/battery/etc., but the menu bar picker only has 6 modules (CPU, Memory, Battery, Network, Disk, GPU) - no temp or fan module. I see CPU temp does show up as an alert rule, but fans don't appear anywhere.

Are those planned, or just described a bit early?

@faruk_kamcici For people coming from iStat Menus or Stats, what’s one alert rule you’d recommend setting up first so they feel that trust without getting notification fatigue?

@faruk_kamcici Congratulations on the launch, Faruk!

The feature that immediately stood out to me wasn't the monitoring itself. It was Compact Health. Most monitoring tools overwhelm you with data, but what we actually want is confidence that everything is fine until it isn't. A quiet system is often the best system.

I also appreciate your privacy-first approach. Reading everything locally and leaving network activity entirely under the user's control is becoming increasingly rare, and I think many developers will value that as much as the monitoring itself.

I'm curious, after people started using mectrics, what metric surprised you the most? Was there a system signal that users consistently overlooked before Compact Health surfaced it? Wishing you a fantastic launch and plenty of happy users!

@dalemooney That’s a really good use case. There’s no CLI or JSON output for alerts right now, but we can build it. I’ll work on adding it in the next release. Thanks for the idea, Dale!

@doganakbulut Thanks, Dogan! Compact Health takes up only the space of a single icon. It shows the overall state at a glance, and clicking it opens a popover with the details. It works especially well on notched MacBooks where menu bar space is limited.

@galdayan Thanks, Gal! That tradeoff was one of my main concerns too. The main app currently settles around 60 to 65 MB of memory. Lightweight metrics update every second on AC and every two seconds on battery, while GPU, temperature, and fan readings run every third cycle.

The planned read-only CLI for the next release should have a lower footprint on headless setups because it skips the UI and samples only what each command needs. If you try Mectrics and see a very different footprint, let me know and I’ll dig into it.

@jacob_sherwood Thanks Jacob! If iStat already works well for you, I would not claim Mectrics has more depth.

The main difference is a quieter approach: Compact Health instead of watching several numbers, sustained alerts that ignore brief spikes, zero telemetry, and free open source code.

The next release also adds a read-only CLI, JSON snapshots, and alert events for headless setups.

It is less about replacing every iStat feature and more about offering a simpler, focused, and private monitoring workflow.

@omri_ben_shoham1 Right now it's the raw number, so you'd catch a hot sensor but not the

throttling. You're right that the throttling is the part that costs you.

The app already reads ProcessInfo's thermal state, but only to back off its own

sampling. It never surfaces it.

Going into the next release, good contribution. Thanks!

@titus_gan Thanks. It's off by default too, so it only shows up if you decide you want it.

@irahimiam Hey, thanks for the comment!

That's exactly why I built it. I had six numbers up there and had stopped

looking at any of them.

It only speaks up if something stays off for a while, not on a quick spike.

Otherwise it just sits there being boring.

Let me know if it ever nags you when it shouldn't.

@mojmir_horvath The version of this I've actually needed is one place that remembers context so success and sales stop asking customers the same questions. In practice the blocker is always integrations. Which systems do you pull from on day one, and how do you handle it when the CRM and support tool disagree about the same account?

@mojmir_horvath For teams just starting out with Poth: what’s the minimum set of data sources you’d recommend connecting first to get a meaningful ‘Ask Poth’ answer within the first week? And what’s a realistic example question a small team could ask at that stage?

@hazy0 Not out of the box today, current sources are more product analytics, calls, and CRM/support. But Poth doesn't care where activity comes from, it's all behavioral signal. We build integrations customer-by-customer during onboarding, and Discord + GitHub activity for devtools is exactly the kind we'd build first. Totally agree that's where dev churn shows up before anyone cancels.

Good to know the integration layer is open ended. If Discord and GitHub are on the near term roadmap for devtools, the question becomes the latency on that data. Community signals go stale fast. Is the ingestion close to real time, or is it a daily batch?

@rabnoor_s On silence: Poth ties the graph to behavior, not just ingested artifacts, so an account going quiet is visible rather than invisible. What we don't do yet is the learned version: "this quiet pattern preceded churn before," but that's where this is headed.

On the disagreement, I mentioned in my reply to Artem but we surface both supporting and contradicting evidence leaving the human user to make the judgement call!

@alex_watson2110 Agreed, nobody announces they've gone quiet. That's why our graph tracks behavior, not just documents. A champion leaving looks like silence in the activity data, and silence is a signal if you're watching for it.

@aliberkuyanik We lost track of AI spend the moment every team started expensing their own API keys and seat licenses. Finance saw one number, eng saw another, nobody could explain the gap. Are you pulling from provider billing directly or from usage logs, and can you attribute spend down to a team or project?

@aliberkuyanik Congrats on the launch!

This is a real gap for us too. We run our own product on top of Anthropic's API, and tracking what each feature actually costs per run, especially after switching between plans, has been a spreadsheet exercise so far.

Curious how granular this gets, can you trace spend down to a specific feature or session, or is it more of a monthly aggregate view right now? We currently do that by hand and it's the part that doesn't scale.

@aliberkuyanik Really like your approach of only showing numbers that can actually be verified instead of estimating them. One thing I'm curious about: how does DepthData handle companies where teams use their own API keys or personal subscriptions outside of company-managed accounts? Can it still surface those blind spots somehow?

@irahimiam Exactly. Most spend trackers ask for way more access than the problem needs. Depthdata is read only and metadata only: seats, usage counts, spend. The endpoints we connect to don't carry prompt content at all, so conversations never enter our system, there's nothing to leak. Narrow scope is the architecture, not a promise, and that's what security teams actually check.

@dalemooney That AWS experience is exactly the thesis. The total never changes behavior, a name next to the line does.

On shadow spend, straight answer: no AI vendor API will ever show a personal subscription someone quietly expenses. That money lives in your expense system, not the vendor's org account. So Depthdata shows it as a gap, not a number. Closing it properly means pulling from the expense side, which is a different connector class. Until then, same rule as everything else: no API to back it, we show the gap instead of a guess.

@raffay_sajjad That six months later failure is exactly the one we designed against. The label isn't a UI decoration, it travels with the number, exports and reports carry MEASURED or ALLOCATED on every figure. So even when nobody remembers which number was which, the number remembers. Once an estimate loses its label it becomes a fact, and that's how dashboards quietly go wrong.

@galdayan Not out of scope at all. The overlap itself is measurable from data we already pull: same people holding seats in two same category tools, both active, and what the double coverage costs. So Depthdata can name it: 40 people pay for both X and Y, here's what the second one costs.

What we won't do is pick the winner. Whether it's really the same job is your call. We put a name and a number on the overlap so somebody has to defend it, the judgment stays with you.

@asadmalik901 You're right, the label is the product. And agreed, idle seats only matter on seat priced tools. On usage pricing the money concentrates, one heavy user can outspend a team, and a seat view hides that. So Depthdata shows cost per person, not just seats. Those numbers come straight from vendor APIs, Anthropic and OpenAI and Cursor all expose per user spend in their docs. Demo runs on sample data today, but nothing on screen an endpoint can't back.

The gaps, Gemini bundles AI into Workspace so per user cost doesn't exist, Replit pools credits with no per member API, Vercel needs tagging first. We show those as gaps instead of estimating over them.

@neo_tiangratanakul great stuff! Incredible and timely tool for stopping deepfakes and scams before they strike!!

@neo_tiangratanakul the real time catch is genuinely the right problem to attack, the Arup story only works as a warning if the detection happens during the call. the thing i'd push on is the on-device part specifically though: no cloud inference means the model shipped in the app is what you're stuck with until the next app update, while the generators on the other side (the ones actually producing the deepfakes) get better continuously and don't need anyone's permission to ship. a cloud model can get retrained and pushed same day a new generation technique shows up. an on-device one is gated by your release cycle and by how fast users actually update the app. how are you thinking about that gap, is there some kind of lightweight signature/model update channel separate from full app releases, or does "on-device" mean you're accepting some lag against the newest generation methods as the tradeoff for the privacy story?

@neo_tiangratanakul Congratulations on the launch, Neo! 🚀

What stood out to me most wasn't the deepfake detection itself—it was the decision to make Halo work during the call instead of after it. In security, timing is everything. A perfect report after the money has been wired isn't nearly as valuable as a warning before someone clicks "Send."

I also appreciate the on-device approach. When the purpose is to build trust, keeping sensitive conversations off the cloud feels like the right design choice.

I'm curious, as deepfakes become more realistic every month, do you see the future being purely AI detecting AI, or will we eventually need identity verification and cryptographic proof to truly know who's on the other side of a call? Wishing you and the team an incredible launch! 🔒🚀

Thanks Vikram 🙌 It's neither, actually we use WGC to grab the rendered meeting window directly. That way our model sees and takes in the same screen the user is seeing!

Thanks so much Teagan!!!

Thanks Aleksandar!

Halo is generalizable to do alright against new deepfake methods but to make sure it keeps up to date we have an agentic system that constantly scrapes the web for new deepfake methods, collects data, and fine tunes our model towards these new deepfakes.

For your second question: Halo runs a quality check on the input itself before it ever makes a call. Bad lighting, shaky footage, heavy compression, or a face too far from the camera can introduce noise, so we filter that out first. On top of that, if the model's ever on the fence and not fully confident, it pulls in more signals before deciding, so we don't throw false alarms during calls!

@byalexai Aleksandar has asked the question the whole category rests on, and I think the honest answer is an uncomfortable one.

Detecting whether something was generated is a losing race by construction. You train on the generators that exist, a new one ships, and the detector degrades quietly. Nobody gets a notification on the day it stops working.

The version that does not decay is a different question altogether. Not is this synthetic, but is this the same person as the reference I already trust. Verification against an enrolled identity does not require recognising the generator at all, so a new model does not break it.

Neo, which of those two is Halo actually doing underneath? The answer decides whether this needs retraining every time something new comes out, or whether it holds.

Around 1 in 7 digitial onboarding fraud attempts involve deepfakes and there has been a 16x increase in deepfake attacks in the last 2 years. From our enterprise customers we're seeing these same statistics parroted. I think we're lucky that right now it isn't too wide spread but we believe this to be more common place in the next year. Which is why we built Halo to get ahead of the game!

🚀🚀🚀

Thanks so much for the support!

Thanks Rohit and I totally agree, after-the-fact detection doesn't help much for situations like a wire transfer. Hence why we're so proud to have made a product like Halo!

Halo actually uses WGC to see the rendered meeting window directly, the same way whether it's Zoom desktop or Meet in a browser tab. Which allows us to cover every video conferencing app and browser!

Thanks so much! Honestly, this "anti-solution" space has always been super inspiring to us, there's something exciting about getting to keep building things that protect people as fast as this space moves. Though we see some legitimate use for AI video generation, we also see how much harm it can cause when it's used maliciously. But honestly, stopping bad actors from abusing AI is a big reason why we started this journey in the first place!

As for user base: we've actually already got some enterprise customers using our detection tech, and Halo is really us expanding that out to the consumer side now 🙂

Really good question! Honestly one we thought about a lot. We made it so that Halo checks the quality of what it's seeing: bad lighting, shaky video, heavy compression, or if a face is too far away from the camera. If the conditions are bad it gather more data or waits for the lighting or image size to get better later in the call, and only makes a decision once it actually has good enough data to be confident. So a bad frame on its own isn't going to set off a false alarm. Since, just like you said, getting that wrong costs a real relationship.

@raffay_sajjad The failure direction Raffay raised has a second-order version that worries me more.

A detector people trust changes their behaviour. Before Halo, that Arup employee at least had the option of being suspicious. After Halo, no warning reads as an endorsement, and the natural response to a clear screen is to scrutinise less. So a false negative after deployment costs more than the same miss would have cost before the product existed.

Which argues for never showing a positive verdict. Flag, or say not enough signal. Never say this person is real, because that is the sentence somebody eventually wires money on.

And the uncomfortable half: the Arup money moved because a wire got approved on the strength of a video call. That is a process gap. Halo makes the call safer without touching the reason the call was load bearing in the first place.

So do you position it as a control finance can rely on, or as one signal that still requires an out of band confirmation?

Thanks so much! Really appreciate you taking the time to check it out!

@asadmalik901 Thanks for the detailed use case. Two things here, since witr treats them differently. Zombies are handled, a defunct process shows a [zombie] marker and a warning. But a live process that lost its parent isn't correctly handled. It prints shell as source which is inaccurate. I've filed a bug in the repo to handle this correctly. Thanks again.

@leon_ostrez Thank you! Really appreciate the kind words and support. Hope you enjoy using it.

@kritishpuri Yes, it goes down to individual sockets and file descriptors (in --verbose mode). Sample output section:

  Sockets     : 127.0.0.1:6010 (TCP | LISTENING)
                127.0.0.1:22008 (TCP | ESTABLISHED)
  Open Files  : 9 of 10240 (0%)
  File Descriptors: 9/10240
    0 -> /dev/pts/2 (deleted)
    3 -> pipe:[23241]
    4 -> socket:[26671]
    5 -> /dev/ptmx

It also works in reverse, which sounds closer to what you're actually hunting. --file finds which process is holding a file and gives you its full ancestry chain, and --port does the same for a socket. The TUI has a locks and ports tab for browsing interactively.

Thanks for the question, and I hope you try it out.

@dalemooney @dale_mooney Real results from Windows + WSL2 with Docker Desktop running.

Docker Desktop containers are picked up. Container detection goes through the Docker API rather than the local process table, so it works even though Docker Desktop runs containers in its own VM where those PIDs don't exist in your distro at all:

  Container   : witr-dd-test (id d0a9b15ffb6f)
  Image       : alpine
  Why It Exists :
    docker → witr-dd-test
  Source      : docker
  Note        : The owning process is not visible in this environment.

That last line is deliberate. When the processes live in the Docker Desktop VM it tells you, rather than inventing a chain.

On systemd, I ran it inside the docker-desktop distro, which has no systemd at all and PID 1 is init, and it resolved the source from the cgroup:

  Container   : docker (d0a9b15ffb6f)
  Source      : docker (container)
  Why It Exists :
    init (pid 1) → SessionLeader (pid 8) → Relay (pid 10) → wsl-bootstrap (pid 12)
      → unshare (pid 42) → init (pid 43) → containerd-shim-runc-v2 (pid 2657) → sleep (pid 2680)

Thanks for the question, and I hope you try it out.

@mrbrjan A daemon reparented to PID 1 still resolves correctly, because the owning unit is read from the process cgroup rather than the parent walk:

  Service     : cron.service
  Why It Exists :
    systemd (pid 1) → cron (pid 11300)
  Source      : cron.service (systemd)
  Unit File   : /usr/lib/systemd/system/cron.service

The chain there is degraded to systemd → cron, but the real cause is still recovered. Containers work the same way, the cgroup identifies it through the runtime layers.

Where it falls down is when a bare orphan with no supervisor behind it. Then it walks ppid, lands on PID 1, and reports whatever is sitting there as the source, which is inaccurate. The issue to handle this exists.

Thank you for your comment and hope you personally try out witr.

@phardinghy4670 Thank you for the awesome comment. Hope you try it out and like it.

@talha_ramzan3 Thank you for your appreciation and trying out playgound. To answer your question, it traces through to the host. Example output pointed at a containerized process from the host:

  Container   : docker: witr-hosttest
  Why It Exists :
    systemd (pid 1) → containerd-shim-runc-v2 (pid 49883) → sleep (pid 49904)
  Source      : containerd (container)

So you get the real host-side chain through the shim, plus the container name resolved rather than just an ID.

@talha_ramzan3 The manual chain has one property worth not losing on the way to a single answer.

When you run lsof then ps then systemctl yourself, you see the raw evidence at every hop, so you can tell when a step is lying to you. A synthesised "why" is faster and quietly takes that away.

Which is why I would want the evidence behind the chain available on demand, not just the conclusion. Same output, plus what it read to get there.

You save the four commands. You keep the ability to notice when the answer is wrong.

@raffay_sajjad All the actions in TUI are properly guarded, it compares process start time before a kill or renice, so a signal can't land on a recycled PID. The ancestry walk could be improved here, it walks ppid with loop protection but no start-time validation. So the risk does exists for the parent PID. I've filed an issue to handle this appropriately, and thank you for your comment.

@rabnoor_s Thank you for the kind words and a very fair challenge.

The name came from frustration rather than a claim. The moment it's built for is "why the hell is this port occupied", "why is this even running", and the ancestry chain happened to answer that version of why well enough that the two blended and thus I gave this name.

To answer your other questions, today witr stops at the chain. I made an orphan to check, and it just reports Source: process_api (init) with no warning. But the data to do better is largely already collected: session id and tty sit in the /proc/<pid>/stat slice I parse and currently skip, so checking whether the session leader still exists would give you the dead-shell case directly. systemd may need additional logic, I'll investigate more on this and file issue to enhance the logic.

Thanks again for in-depth technical challenge.

@hi_i_am_mimo Thanks, good question. I tested this on your exact setup before answering.

Advisory, both kinds. On Linux the Locks tab reads /proc/locks, so fcntl/POSIX byte-range locks and BSD flock locks both appear, and the Type column tells them apart. No separate mandatory category. witr shows locks that are held, not workers blocked waiting.

Read vs write per worker: yes. Two Node workers on one SQLite file, one in a read txn, one in a write txn:

PID      Process    Type      Mode      Path
 8964     node       POSIX     READ      .../app.db
 8972     node       POSIX     WRITE     .../app.db

@kocanaoglu80506 Thank you! That was exactly our goal.

We wanted AI to feel like a natural part of building campaigns, not another tool you have to switch into or learn separately.

The idea is to help with the repetitive work, like segmentation, audience selection, and workflow creation, so marketers can spend more time on strategy and creating great customer experiences.

Really appreciate you checking it out.

@rutefig Is this a competitor to screenpipe?

@rutefig Really like that Screencap focuses on capturing how work actually happens instead of relying on people to manually document their workflows. The emphasis on consent, built-in privacy protections, and reviewing sensitive data before it leaves a device makes the approach especially interesting for teams looking to build automation and AI systems responsibly. Good luck with the launch!

@rutefig Small thing: "Browse the dataset" and "Read the anonymization spec" both link to the same URL (screencap.sh/dataset) the spec link looks like it's pointing to the wrong place.

Cool concept overall!

@doganakbulut Hey! Yeah you can definetely use it to document how processes happen as well. The exported data is structured steps yes and window context

@kritishpuri thanks a lot! Labeling is not done at this layer as of course but my colleague @aayushtheg can give you more insights there

@dalemooney yes it is also in terms of content, anything that looks password, emails, names, etc is masked, review is mostly just another safeguard for people that want to be really sure

@german_merlo1 thank you :)

@justin2025  As a self learnt Developer and former Nurse - I am completely convinced that Robots esp. in Medical - Area are already “ IN”! The future will be - decentralized, self sustaining and - fantastic! Imagine this Robot with a Doctors And Nurses brain !!? Imagine sickness “ repair “ in hours instead of months! Gemini Robotics 2 is - fascinating for me!!

@jernej_jan_kocica yes, it can carry the grounding context and tool results, not just call timing, but we make that explicit and configurable because production AI traces can contain sensitive customer data.

@jernej_jan_kocica Jernej has named the only question on that list that matters, and the answer to it creates a second problem worth planning for.

Grounding capture is configurable because production traces hold customer data. Reasonable. But it means it gets switched off exactly where a silent wrong answer is most expensive, and whoever switches it off has to make that call before they know they will need it.

The version that survives both constraints is capturing identity instead of content. Retrieval document ids, a hash of each chunk, tool arguments and a hash of the result. You can then prove which context a turn actually used, and whether it changed between two runs, without storing a line of customer text.

Not the same as replaying the answer. But it turns "the worst incident of the year looks like a healthy span" into something you can diff.

Could the span schema carry a content-free grounding fingerprint today?

@dalemooney Yes. TraceLLM has a first-class tool span kind, so tool calls do not have to live as a JSON blob on the model span.

In a typical agent/RAG flow you can emit separate spans for agent.plan, tool.crm.lookup, retrieval.docs.search, and openai.chat.complete. Each one gets its own timing, status, metadata, optional input/output, and errors.

On OTLP export, those TraceLLM spans are mapped to real OpenTelemetry spans, so in SigNoz/Tempo/Honeycomb the tool latency can be inspected separately from the model latency.

The current caveat is that this is SDK-instrumented today: you create the tool spans from your app code. Automatic MCP tool-call capture is on the roadmap, so MCP calls can become first-class spans without as much manual wiring.

@dalemooney Building on Dale's point, the thing that bit us is that a tool span is usually not a leaf.

An MCP tool call often fans out to two or three services of its own. If the whole tool is one span you get a total, and you still cannot say which downstream call is the slow one. That is the same JSON blob problem, one level down.

The other half is cost. If half the latency is tools, a real share of the spend is too, and most tracing attributes cost only to model calls. Cost per turn is only honest if tool cost lands on the turn that caused it.

So two things: does a tool span nest, and can it carry a cost attribute?

@alexmubert nice launch.. congrats qq what’s the average response time when requesting real-time stem swaps over the API?

@asadmalik901 That's a great point. Once music becomes editable instead of something you regenerate from scratch, it fits much better into real creative workflows. I also agree that consistency is critical, especially for brands that revisit campaigns weeks or even months later.

@asadmalik901 Asad has named the thing that decides whether this is an API or a toy, and I would push it one step harder.

Seed plus prompt determinism is not really the requirement. The requirement is determinism across model versions, and no vendor can promise that while also shipping a better engine. If the November model is genuinely better, the same seed hands you a better track, which is exactly wrong for a brand video that needs the one from September.

So the two things I would want written into the contract are a pinnable engine version, and generated tracks treated as immutable stored assets rather than something you regenerate. Re-fetch the artifact, do not re-roll it.

Vadim, is a generated track permanently addressable once created, and can a caller pin the engine version it was made with?

@kirill_kiryanov the stem-swap-without-full-reroll thing is the right fix, Asad's point about it turning from a slot machine into an editable asset is exactly it. question tied to the "copyright-protected" framing on the tin though: once someone edits a track, swaps a stem, and ships that as part of a paid product, who actually holds the copyright on that specific edited output, the user or Mubert? asking because "copyright-protected" usually means "you won't get DMCA'd for using it," but it's a separate question from "you own this specific derivative enough to enforce it if someone else rips your final mix." those are different guarantees and only one of them matters if you're building a commercial product on top of the API rather than just using the output as background music.

@evokoa always struggling with context window and optimize cost as my db grows large, and i started moving into using polygres. i was very suprised to see how my agent can queries everything a lot more efficiently in term of memory and cost after that like literally. still working on trying out the product more as the knowledge based become larger, but this looks exciting!!!

Congrats the team on the launch

@evokoa I like the idea of reducing unnecessary complexity

@dylan_bryan2 That's a great idea! Happy to help you build this out you can dm me on my X.

@lucapiekarski The full database remains available as external memory. For each request, Polygres retrieves and ranks the most relevant rows, relationships, and search results, then fits that information into the model’s actual token limit.

As the conversation continues, the system can fetch additional context on demand instead of trying to load the entire database into one prompt. So the database is effectively unbounded context.

We've dubbed this the infinite context window!

@1nk Hey inky! so we have a cli and skill that you can connect directly to the agent :D https://docs.evokoa.com/polygres/agent-skills

@istp Glad to hear that! our goal is to make it incredibly easy to deploy a production grade AI agent :D Build it on sunday, be enterprise ready by monday :D

@b4boom Yeah! For code specifically we are releasing a feature called Polygres Pocket soon, stay tuned!

@james_yang7 Loops are just shitty graphs. I wrote a whole piece about this :D https://x.com/daleverett/status/2078969402046009374

If you have any questions or feedback I would love to hear it and will try to answer as quickly as possible thank you again for being here

@ngotomek Super clean landing page! How do notifications work once a vulnerability is found? Can we route alerts to Slack or Email...

@leon_ostrez Thank you, appreciate the words of support. Any suggestions for improvement?

@kisekiya Thanks 🙏 That surge is what we are going after, more people shipping faster with less ability to verify it. It's free, so give it a run and let me know what you think

@chirag_chamoli Thanks means a lot, and you named the exact reason we built it. The "small team shipping AI code with no security function" is precisely who gets left out. Real security tooling is priced for companies with a security budget and those are the people who need it just as much. Appreciate you taking the time 🙏

@istiakahmad Thanks a lot. That's exactly what we're going for, letting people build as much as they want, free for now, with the guarantee that whatever you vibe-code gets an independent verification. So you don't have to worry as much about what your AI is producing. Thank you again for this comment and if you have any suggestions on improvements please let us know

@garmdotcom Really interesting approach. I'm curious—how does Nommer handle recipes where one step finishes earlier than expected? Does it dynamically rebalance tasks so one person isn't left waiting?

@jamesjustgains awesome!! Let me know how it goes :-)

@omri_ben_shoham1 It’s simple. Meta for example focuses on max profit and annual mass layoffs. Eventually any project needs at least enough to pay for web hosting. Beyond that the owner(s) must make a simple decision: Max profit at all costs, or just enough to pay for infrastructure and some employees?

And a better version of the iOS app is coming. Apple is taking their sweet time approving. Grrr. To the moon, Apple, to the moon.

@zvonimir_sabljic1 Thank you so much 🤗

@zvonimir_sabljic1 Thanks, Zvonimir! Are you in the franchise space? If so, I'd love to connect!