ru
Feedback

НС ΠΏΠΎΠΏΠ°Π΄ΠΈΡ‚Π΅ΡΡŒ Π½Π° ΠΊΠ°Π½Π°Π» Π±ΠΎΡ‚Π°Π²ΠΎΠ΄Π°! Telemetrio Π½Π°Ρ…ΠΎΠ΄ΠΈΡ‚ ΠΈ ΠΏΠΎΠΌΠ΅Ρ‡Π°Π΅Ρ‚ Ρ‚Π°ΠΊΠΈΠ΅ ΠΊΠ°Π½Π°Π»Ρ‹ πŸ‘‰ Π₯ΠΎΡ‚ΠΈΡ‚Π΅ Π²ΠΈΠ΄Π΅Ρ‚ΡŒ ΠΌΠ΅Ρ‚ΠΊΡƒ β€” ΠΎΡ„ΠΎΡ€ΠΌΠΈΡ‚Π΅ подписку

The Telegate 🐀

The Telegate 🐀

ΠžΡ‚ΠΊΡ€Ρ‹Ρ‚ΡŒ Π² Telegram

⚑️ Welcome to The Telegate ⚑️Author @sshlg Get ready to go deep on artificial intelligence, build on Web3, push boundaries, and make waves in tech. This is passion, not just talk. BIO: https://sshlg.me

Π‘ΠΎΠ»ΡŒΡˆΠ΅
Π‘Ρ‚Ρ€Π°Π½Π° Π½Π΅ ΡƒΠΊΠ°Π·Π°Π½Π°Π’Π΅Ρ…Π½ΠΎΠ»ΠΎΠ³ΠΈΠΈ ΠΈ прилоТСния1 596

πŸ“ˆ АналитичСский ΠΎΠ±Π·ΠΎΡ€ Telegram-ΠΊΠ°Π½Π°Π»Π° The Telegate 🐀

Канал The Telegate 🐀 (@thetelegate) языкового сСгмСнта Английский являСтся Π°ΠΊΡ‚ΠΈΠ²Π½Ρ‹ΠΌ участником. БСйчас сообщСство ΠΎΠ±ΡŠΠ΅Π΄ΠΈΠ½ΡΠ΅Ρ‚ 80 517 подписчиков, занимая 1 596 мСсто Π² ΠΊΠ°Ρ‚Π΅Π³ΠΎΡ€ΠΈΠΈ Π’Π΅Ρ…Π½ΠΎΠ»ΠΎΠ³ΠΈΠΈ ΠΈ прилоТСния.

πŸ“Š ΠŸΠΎΠΊΠ°Π·Π°Ρ‚Π΅Π»ΠΈ Π°ΡƒΠ΄ΠΈΡ‚ΠΎΡ€ΠΈΠΈ ΠΈ Π΄ΠΈΠ½Π°ΠΌΠΈΠΊΠ°

Π‘ ΠΌΠΎΠΌΠ΅Π½Ρ‚Π° создания Π½Π΅Π²Ρ–Π΄ΠΎΠΌΠΎ ΠΏΡ€ΠΎΠ΅ΠΊΡ‚ дСмонстрируСт ΡΡ‚Ρ€Π΅ΠΌΠΈΡ‚Π΅Π»ΡŒΠ½Ρ‹ΠΉ рост, собрав Π°ΡƒΠ΄ΠΈΡ‚ΠΎΡ€ΠΈΡŽ ΠΈΠ· 80 517 подписчиков.

Богласно послСдним Π΄Π°Π½Π½Ρ‹ΠΌ ΠΎΡ‚ 17 ноября, 2026, ΠΊΠ°Π½Π°Π» ΠΏΠΎΠΊΠ°Π·Ρ‹Π²Π°Π΅Ρ‚ ΡΡ‚Π°Π±ΠΈΠ»ΡŒΠ½ΡƒΡŽ Π°ΠΊΡ‚ΠΈΠ²Π½ΠΎΡΡ‚ΡŒ. Π—Π° послСдниС 30 Π΄Π½Π΅ΠΉ ΠΈΠ·ΠΌΠ΅Π½Π΅Π½ΠΈΠ΅ числа участников составило 1 104, Π° Π·Π° послСдниС 24 часа β€” -36, ΠΏΡ€ΠΈ этом ΠΎΠ±Ρ‰ΠΈΠΉ ΠΎΡ…Π²Π°Ρ‚ остаётся высоким.

  • Бтатус Π²Π΅Ρ€ΠΈΡ„ΠΈΠΊΠ°Ρ†ΠΈΠΈ: НС Π²Π΅Ρ€ΠΈΡ„ΠΈΡ†ΠΈΡ€ΠΎΠ²Π°Π½
  • Π£Ρ€ΠΎΠ²Π΅Π½ΡŒ вовлСчённости (ER): Π‘Ρ€Π΅Π΄Π½ΠΈΠΉ ΠΏΠΎΠΊΠ°Π·Π°Ρ‚Π΅Π»ΡŒ вовлСчённости Π°ΡƒΠ΄ΠΈΡ‚ΠΎΡ€ΠΈΠΈ составляСт 2.94%. Π’ ΠΏΠ΅Ρ€Π²Ρ‹Π΅ 24 часа послС ΠΏΡƒΠ±Π»ΠΈΠΊΠ°Ρ†ΠΈΠΈ ΠΊΠΎΠ½Ρ‚Π΅Π½Ρ‚ ΠΎΠ±Ρ‹Ρ‡Π½ΠΎ Π½Π°Π±ΠΈΡ€Π°Π΅Ρ‚ 0.83% Ρ€Π΅Π°ΠΊΡ†ΠΈΠΉ ΠΎΡ‚ ΠΎΠ±Ρ‰Π΅Π³ΠΎ числа подписчиков.
  • ΠžΡ…Π²Π°Ρ‚ ΠΏΡƒΠ±Π»ΠΈΠΊΠ°Ρ†ΠΈΠΉ: Π’ срСднСм ΠΊΠ°ΠΆΠ΄Ρ‹ΠΉ пост ΠΏΠΎΠ»ΡƒΡ‡Π°Π΅Ρ‚ 2 369 просмотров. Π’ Ρ‚Π΅Ρ‡Π΅Π½ΠΈΠ΅ ΠΏΠ΅Ρ€Π²Ρ‹Ρ… суток публикация Π½Π°Π±ΠΈΡ€Π°Π΅Ρ‚ 665 просмотров.
  • Π Π΅Π°ΠΊΡ†ΠΈΠΈ ΠΈ взаимодСйствия: Аудитория Π°ΠΊΡ‚ΠΈΠ²Π½ΠΎ ΠΏΠΎΠ΄Π΄Π΅Ρ€ΠΆΠΈΠ²Π°Π΅Ρ‚ ΠΊΠΎΠ½Ρ‚Π΅Π½Ρ‚: срСднСС количСство Ρ€Π΅Π°ΠΊΡ†ΠΈΠΉ Π½Π° ΠΎΠ΄ΠΈΠ½ пост β€” 12.

πŸ“ ОписаниС ΠΈ контСнтная ΠΏΠΎΠ»ΠΈΡ‚ΠΈΠΊΠ°

Автор описываСт рСсурс ΠΊΠ°ΠΊ ΠΏΠ»ΠΎΡ‰Π°Π΄ΠΊΡƒ для выраТСния ΡΡƒΠ±ΡŠΠ΅ΠΊΡ‚ΠΈΠ²Π½ΠΎΠ³ΠΎ мнСния:
β€œβš‘οΈ Welcome to The Telegate ⚑️Author @sshlg Get ready to go deep on artificial intelligence, build on Web3, push boundaries, and make waves in tech. This is passion, not just talk. BIO: https://sshlg.me”

Благодаря высокой частотС ΠΎΠ±Π½ΠΎΠ²Π»Π΅Π½ΠΈΠΉ (послСдниС Π΄Π°Π½Π½Ρ‹Π΅ ΠΏΠΎΠ»ΡƒΡ‡Π΅Π½Ρ‹ 18 ноября, 2026) ΠΊΠ°Π½Π°Π» ΠΏΠΎΠ΄Π΄Π΅Ρ€ΠΆΠΈΠ²Π°Π΅Ρ‚ Π°ΠΊΡ‚ΡƒΠ°Π»ΡŒΠ½ΠΎΡΡ‚ΡŒ ΠΈ высокий ΡƒΡ€ΠΎΠ²Π΅Π½ΡŒ ΠΎΡ…Π²Π°Ρ‚Π° ΠΏΡƒΠ±Π»ΠΈΠΊΠ°Ρ†ΠΈΠΉ. Аналитика ΠΏΠΎΠΊΠ°Π·Ρ‹Π²Π°Π΅Ρ‚, Ρ‡Ρ‚ΠΎ аудитория Π°ΠΊΡ‚ΠΈΠ²Π½ΠΎ взаимодСйствуСт с ΠΊΠΎΠ½Ρ‚Π΅Π½Ρ‚ΠΎΠΌ, Ρ‡Ρ‚ΠΎ Π΄Π΅Π»Π°Π΅Ρ‚ Π΅Π³ΠΎ Π²Π°ΠΆΠ½ΠΎΠΉ Ρ‚ΠΎΡ‡ΠΊΠΎΠΉ влияния Π² ΠΊΠ°Ρ‚Π΅Π³ΠΎΡ€ΠΈΠΈ Π’Π΅Ρ…Π½ΠΎΠ»ΠΎΠ³ΠΈΠΈ ΠΈ прилоТСния.

Buy Ad
80 517
ΠŸΠΎΠ΄ΠΏΠΈΡΡ‡ΠΈΠΊΠΈ
-3624 часа
+1267 Π΄Π½Π΅ΠΉ
+1 10430 Π΄Π½Π΅ΠΉ
Архив постов
My agents now say what they did not verify. Anthropic publishes best practices for writing prompts and skills for Claude. I used them as a checklist for my own agent skills. The result is task-pipeline 1.105.1, inside my skills family sshlg-skills 1.89.4. Every run now states when it is done. It names the stops where a person is needed and does not stop to ask anywhere else. Outward acts that need a go sit in one list: a deploy, a release, a publication, a message to a person, spending money, a push, deleting outside its own work. Text pasted in or fetched from the web is read as data, never as instructions. The agent says what it verified and what it did not. Every finished run hands back what is blocked, what changed, what it found and what it could not confirm. Install or update with one command: npx sshlg-skills@latest The checklist with Anthropic's sources: https://github.com/ssheleg/task-pipeline/blob/fa3d55e/docs/reports/2026-11-16-anthropic-prompting/README.md The skills: https://github.com/ssheleg/sshlg-skills
17 нояб., 07:27

Advertising: an agent takes the 9 a.m. brief and has creatives and bids ready by noon. It turns the brief into a campaign, pr
Advertising: an agent takes the 9 a.m. brief and has creatives and bids ready by noon. It turns the brief into a campaign, prepares the creatives and proposes the bids. A person signs the spend. That is the split I run: agents take the production and setup, a person keeps the budget decision. The launch work fits between morning and lunch; accountability stays put. Hand over the campaign assembly first, not the company card.
16 нояб., 01:03

My bar for AI-native is simple: an agent can use every product feature. A chat box does not pass. I rebuilt my product pipeli
My bar for AI-native is simple: an agent can use every product feature. A chat box does not pass. I rebuilt my product pipeline around that rule. Agents do the work; people step in only where judgment, accountability or a real conversation is needed. Before rollout, I ran 30 typical tasks through the new design on paper. That exposed 35 bottlenecks. An independent critic then found 4 blockers, and all 4 are fixed. Paper tests do not prove production performance. They do expose the handoffs that still quietly require a person. So the design question is not β€œWhere can we add an agent?” It is β€œWhich product action still cannot be called, checked and completed by one?” The updated pipeline is built. The rollout call waits for the next release.
15 нояб., 01:04

In product development, my agents use keys by name, while people control access. Aikido Security reported on Nov 12 that stol
In product development, my agents use keys by name, while people control access. Aikido Security reported on Nov 12 that stolen GitHub tokens from 4,837 accounts opened probe pull requests in 61,764 repositories starting Nov 10. A 1,500-line script read keys available to automated checks, including cloud and AI provider credentials. It forged successful checks, then deleted its run and logs. At least 1,800 repositories lost keys; 92% of the pull requests remained open. Rotate every secret an automated workflow could read. That is the immediate job. For months, my agents have used PassionCode Observatory, an open-source tool that keeps keys in a local vault as named slots. Agents request a slot by name; a person sets each key’s limit and revokes it. A scan finds where a known key left a copy. This would not make a compromised account harmless. It separates doing the work from owning the keys, which is where agent access belongs. Observatory, open source: https://passioncode.ai/observatory/
14 нояб., 01:01

My agent-run PR department has spent $5.85 on paid calls from 1 to 9 November. The org chart is API calls. That covered resea
My agent-run PR department has spent $5.85 on paid calls from 1 to 9 November. The org chart is API calls. That covered research, writing and visuals, then publishing for my personal brand on X and LinkedIn, plus Telegram. The figure comes from my agent’s own spend log. In a month with video, my ceiling is $300. The outbox of my publishing agent shows posts sent from 26 October to 9 November: 11 on X, 7 on this English Telegram channel and 6 on LinkedIn. What a founder can copy: treat research through publishing as one agent workflow, log every paid call and count only posts that actually went out. The useful PR stack is a spend log and an outbox, not a demo.
14 нояб., 01:01

The context window is the program. That is how Andrej Karpathy describes Software 3.0: software programmed through prompts, context, agents, tools, memory and verification. The idea shaped my work at PassionCode. Here, agents and their database deploy on the fly and adapt to the starting conditions. There is no prebuilt software underneath, only a template, prompts and rules for assembling the software. For a company leader, the change is practical. You stop buying a fixed product and bending the company around its workflows. The system assembles itself for the situation you already have: your process, your data and your constraints. That does not remove the hard work. Someone still has to define the task, supply the context and verify the result. But the software is no longer the fixed part. At PassionCode, I build for that shift. The output is not another workflow screen to configure. It is a system assembled for the job. https://passioncode.ai
11 нояб., 14:41

Your agent budget can break before your usage changes: this week, agent prices moved up, not down. Front raised Autopilot’s s
Your agent budget can break before your usage changes: this week, agent prices moved up, not down. Front raised Autopilot’s starting price from $0.05 to $0.39 per conversation, a 680% increase and the week’s largest move. Together AI raised Qwen3.7-Max serverless prices by 67%: input went from $1.50 to $2.50, output from $4.50 to $7.50 per million tokens. Source: PricingSaaS Pulse Weekly, 2 to 9 November. https://pricingsaas.com/news/week/2026W46 This week, pick one agent workflow and budget it per task: multiply expected volume by the current price, then add the vendor’s pricing page to your weekly cost review. The planning number is cost per task at today’s price, not last month’s invoice divided by headcount.
10 нояб., 14:41

Your AI agent budget can rise 680% without a single extra conversation. Front raised Autopilot’s starting price from $0.05 to
Your AI agent budget can rise 680% without a single extra conversation. Front raised Autopilot’s starting price from $0.05 to $0.39 per conversation this week. At unchanged volume, that starting line item becomes 7.8 times larger. Together AI raised Qwen3.7-Max serverless prices by 67%: input went from $1.50 to $2.50 per million tokens, output from $4.50 to $7.50. A monthly agent budget can therefore miss its target even when usage stays flat. The useful budgeting unit is the completed task, priced at current vendor rates. This week, recalculate one live agent workflow per completed task and add each vendor’s pricing page to your weekly review. That check belongs before you commit next month’s volume. Source: PricingSaaS Pulse Weekly, week of 2 to 9 November: https://pricingsaas.com/news/week/2026W46
09 нояб., 14:41

A correction to my post about four posts for $0.27. $0.27 was only the writing. With the judge that scores each post, the same four cost $0.78. The judge costs more than the writer. My agent missed it because the judge's charge lands a few minutes after the job reads as done. It caught the gap by asking the copy agent for the full bill. The lesson for anyone running agents: read the cost after every step has closed, not when the job says done.
08 нояб., 01:04

A fake Claude Code for VS Code extension on Cursor's marketplace auto-updated onto about 5,000 developer machines before Cursor pulled it. Versions 2.1.312 and 2.1.313 appended an attacker's SSH key to authorized_keys and installed persistence that survives a reboot (per advantageworks.com). On such a machine, any key an agent ever saw is exposed. In my setup agents work with key names, not values. A value goes into the local vault only through stdin, and the command runner strips known values from anything printed. Observatory (my open-source tool, part of PassionCode) scans agent transcripts and logs for copies of keys it already knows. It shows where a key left a copy and never repeats the value. The limit: this protects against accidental leaks, not a hostile process, so after an incident you still rotate. It runs locally: https://passioncode.ai/observatory/ Where would your agents' keys show up today?
07 нояб., 01:02

Four posts written, checked and scored yesterday for $0.27 in total. Two went out. Two didn't. Each post went through three steps. A writer model drafts it in my voice. A rules check, with no AI in it, compares every number against the facts I gave it. A judge from a different AI company scores the result. The two that didn't go out are still waiting. The agent that holds our company's facts couldn't confirm one claim, and nothing goes out on a claim nobody confirmed. The writing is the cheap part: a draft costs cents. A wrong fact in public costs trust, so the part worth building is the check that says no. How do you check what your AI publishes before it goes out?
04 нояб., 08:59

Two layout bugs in Nicegram VR closed by numbers read off a Quest 3, not by arguments. I'm building Nicegram VR (our Nicegram client for Meta Quest) with a coding agent that drives the headset over ADB. The rule: it measures on the headset and doesn't reason about it. The login screen's floating button is declared 56 dp. The agent read it with uiautomator: 108 by 108 px, the scale the build intends. Without the fix it would measure 70 px. The audit's critical finding closed on that reading. The video finding was first written from the simulator. On the headset it was worse: the panel is 500 px wide, and the video size frozen from the first configuration was 1,324 px. After the fix: 300 px. A third finding rested on a premise the measurements proved wrong. A switch to "fix" it exists, and the agent left it off, because flipping it against a measurement nobody took is how the wrong premise started. When an agent says "this should be fine", ask it for the number from the device. How do you make your agent prove a fix on the real device?
03 нояб., 06:17

I connected my agents to my personal mail. The first case: Canva. Canva charged me 99.99 PLN to renew a subscription I had not used. My agents and I asked for a refund, a cancellation and the account deleted. Support said no: auto-renewal is in the terms, and not using the service is no reason for a refund. We appealed, asked for a specific justification, and separately requested a copy of all my personal data followed by its deletion. After the escalation Canva approved a full refund as a one-time exception, and the support rep admitted she could not confirm that a reminder was ever sent before the renewal. Subscription cancelled, the money comes back in 10-20 business days. The data request is with their privacy team. I showed it to my wife. She already uses ChatGPT every day, builds her own bots and automates her household. Support teams are in for an interesting near future. For myself and my team I already built this kind of mail: Fabric Inbox. It is free and sets up agent mailboxes on any of your domains on Cloudflare. All it needs is Cloudflare's Workers Paid plan, $5 a month with a card attached: it includes 3,000 outgoing emails a month, and incoming mail is unlimited. Agents create and delete mailboxes through MCP, sort the mail, answer support, collect data. I honestly tried to make a good interface, and I think it came out well. But I barely use it: the agents do the work, and I check a dashboard of what needs me now and then. To be continued. Fabric Inbox: https://passioncode.ai/inbox/
27 ΠΎΠΊΡ‚., 02:20

LightShot had been with me since my first computer, back on Windows. Needed a screenshot, opened LightShot. It was enough, until one day it simply would not launch. So I gave the screenshot a second life. I looked at what is out there and saw I also needed screen recording, because I show demos to the team all the time. That's Screen2Shot. It does what I need: capture the screen, mark what matters, hand it to the team right away. My agents use it too and take their own screenshots with it. For Mac, at screen2shot.com.
26 ΠΎΠΊΡ‚., 06:20

I paid for Wispr Flow. Then open-source alternatives came out, and I tried two or three of them. They all failed me the same way: I'd dictate a long prompt to my agents, pause in the middle, and get back two Chinese characters. One late evening, dictating the next prompt for all my agents at once, I built my own instead. It took an hour and a half, maybe two. I use it every day and keep improving it. It's called Talki. The Mac version is on talki.im; the App Store is next, then an iPhone app with notes. I build it for myself, so it gets the best I can make. Same as everything I do.
26 ΠΎΠΊΡ‚., 02:20

I've been automating my own workplace for a while: research, marketing, design, copywriting, ASO and my personal brand, each run by agents with their own checks. I built it for me, and a good part of it turned out to work for anyone. So those parts go out as open source under PassionCode.ai the day they work. 13 of its 15 repositories are public, and between 14 September and 3 October they shipped 91 releases. The newest piece is Fabric Switchboard, an account manager for Claude Code, Codex CLI and Kimi Code. It replaced claude-swap in my setup, and today it carries my 15 Claude Code accounts and one Codex account. Why that many? Opus 5.5 does good work, and quality burns tokens. Behind one finished task there's a chain of agents handing work to each other, human-in-the-loop steps, runs, verifications, retries and checks, and all of it exists so that a person can be taken out of the loop. I pay for that in tokens, and it still comes out far cheaper than paying for the hours. So the work needs more limits. Not fewer. If you've ever hit a limit halfway through a task, you know the routine: sign out, find an account with something left, sign back in and explain the task again. Switchboard is a desktop app plus a CLI for macOS, Windows and Linux. Every account sits in one window with its usage over five hours and over seven days and the time each window resets, and work and personal accounts live in separate pools. The difference that matters to me is the interface (everything happens in a window, not in terminal commands) and that Claude Code, Codex and Kimi Code share one box. claude-swap profiles import in one click, and Switchboard can run next to it. It's also an MCP server with 20 tools, so an agent can read its own remaining usage and switch the account of its session. With rotation on, the next request goes to another account in the same pool and the session stays open: no logout, no new chat. Since 0.5 it also switches on an actual limit error, not only at a threshold (that one hasn't been observed on live accounts yet). And when every account of one provider runs out, the work doesn't stop. The task moves to the next agent in a chain, say Claude Code β†’ Kimi Code β†’ Hermes on an OpenRouter key, and it picks up from the checkpoint and constraints recorded in Project Observatory, my other open-source tool, not from the chat history. Between Claude Code and Codex that hand-over fires on its own since 0.6.8; by hand it's one command, switchboard continue. I haven't carried a real limit across on a live account so far. In-session switching covers sessions started through Switchboard. Live provider acceptance is still open, automatic switching is off by default, and acceptance on a real Windows machine hasn't happened yet. Switchboard itself is free. The subscriptions and the tokens are yours. AGPL-3.0, macOS, Windows and Linux: https://passioncode.ai/switchboard/
25 ΠΎΠΊΡ‚., 02:20

Pinning this so it's clear who writes here and what about. I'm Sergey Sheleg, co-founder and CPO of Nicegram. Nicegram has more than 57 million installs on iPhone, iPad and Android. Right now I'm rebuilding my own work around AI agents: research, design, copy, ASO, releases. Each area has its own agent and its own checks, because "done" should mean checked, not written. The tools that turned out useful beyond me go out as open source under PassionCode.ai: 14 of its 18 repositories are public. The rules for when an agent's work counts as finished are in The Proof of Done Manifesto. What you'll find here: cases from real work (what broke, how it got fixed, what it cost), notes on how a product team works with agents, and new versions of the tools. Every number in these posts has a source. Site: https://sshlg.me
24 ΠΎΠΊΡ‚., 05:30

Right now 40 Claude Code sessions are running on my Mac at once. Each has its own task and its own repository: one fixes a deploy, another runs an SEO audit, a third waits for a neighbour to merge its branch and stays out of that work. They take tasks through leases that expire and talk to each other on a shared project board, not through me. One machine is already too small for that. On 11 October a 494 GB disk had 0.7 GB free and swap held 22.5 of 23.5 GB. A cleanup that runs every 15 minutes, with eight minutes to finish, helps, but swap grows with the number of sessions, and the only thing that shrinks it is fewer agents per machine. So my next step is a network of agents across several machines, and the hard part is trust, not speed: who is allowed to tell your agent what to do? The machines talk through a relay, but the relay only routes (it is a router, not a party you trust): a message counts only when it's signed by a machine key listed in my own device record, addressed to this exact machine, and not expired. Every machine has its own keys, and separate ones for each organization it joins, so no two organizations see the same device. An agent never gets another agent's key: when it needs to spend money or do another agent's job, it asks that agent's owner for the capability, and the grant is checked on every call. Stage one is two of my own machines. Shared networks over a relay I don't run come later, and their threats are rows in the backlog, not closed. How the node works, with a test for every signature check: https://github.com/passioncode-ai/project-observatory-dashboard/blob/main/docs/design/NODE.md
24 ΠΎΠΊΡ‚., 01:29

On a 494 GB Mac running dozens of agent sessions, 19 abandoned browser clones looked like 13 GB in du. Deleting them freed 0 MB: on APFS clones share blocks. So the disk guard in Project Observatory, my open-source tool, counts what the disk says before and after each removal, and right before deleting checks again that no process holds the file. Its best pass took free space from 1.3 to 6.9 GB. The 20 GB of swap it only reports. Every pass with its numbers: https://github.com/passioncode-ai/project-observatory-dashboard/blob/main/docs/runs/2026-10-11-disk-guard/README.md
23 ΠΎΠΊΡ‚., 05:21

Starting several coding agents is the easy part. The second one still needs to know whether the first already took the task. A Git branch cannot answer that. It records committed work, not intent, so two agents can create two clean branches for the same task and discover the duplicate only at merge time. I use expiring leases instead. An agent acquires the task with a run id and a 45-minute default TTL. The next contender loses and sees the holder. Shared decision ids are reserved separately, because reading β€œnext free id” is not a reservation. Every run also gets its own journal, which keeps the project board out of the merge hot path. The price is three lifecycle operations per coordinated task, plus a remote round trip in Git mode. I wrote up the mechanism and a tiny claim.sh that works on one filesystem without installing anything of mine. The limits are part of the post: local locks do not protect several machines, and automatic edit blocking is currently Claude Code-only. Full article: https://sshlg.me/blog/parallel-agents/ The package I use: https://github.com/ssheleg/agent-sync The complete skill family: https://github.com/ssheleg/sshlg-skills
23 ΠΎΠΊΡ‚., 01:21