| Name | Modified | Size | Downloads / Week |
|---|---|---|---|
| Parent folder | |||
| SECURITY.md | 2026-08-09 | 16.2 kB | |
| LICENSE | 2026-08-09 | 1.1 kB | |
| CHANGELOG.md | 2026-08-09 | 140.4 kB | |
| PC_Workman_HCK_1.8.7.zip | 2026-08-09 | 49.3 MB | |
| PC Workman v1.8.7 - the assistant holds a conversation, and app stops lying about itself source code.tar.gz | 2026-08-06 | 32.4 MB | |
| PC Workman v1.8.7 - the assistant holds a conversation, and app stops lying about itself source code.zip | 2026-08-06 | 32.8 MB | |
| README.md | 2026-08-06 | 9.5 kB | |
| Totals: 7 Items | 114.6 MB | 73 | |
PC Workman v1.8.7 - the assistant holds a conversation, and the app stops lying about itself
Two releases ago I wrote about four lies my own code had told me. This one found five more, and every single one had the same shape: it worked perfectly and did nothing.
A desktop shortcut that opened no application. A security scan that flagged the Windows Print Spooler. A diagnostic console that stayed on screen without a word of explanation. An optimizer that never said what it gained. A tile icon that had been stretched from 32 pixels to 150 since the day it shipped.
None of them crashed. That is why they lasted.
The assistant can be walked through a problem
Guided conversations used to be one hand-written handler. They are an engine now, so every flow understands "next", "yes", "skip" and "stop" without anyone writing that code again.
Four ship: optimize, cooling, desktop repair and upgrade planning. The optimize flow measures first, walks you through startup and services, asks before it touches your RAM, then checks the result against the numbers it took at the beginning.
It also follows a real conversation. Correct the subject mid-sentence ("I meant the GPU"), add a detail you forgot a moment ago, or come back after actually trying the advice, and it keeps up instead of starting over. Eleven handlers exist purely for that, and a semantic layer covers the relationships a keyword parser drops: process plus close plus safe is a different question from process plus memory.
Vocabulary is at 127 intents, Polish and English.
The 23 guides are reachable from inside the app
Over the summer I wrote 23 problem-solving guides on pcworkman.dev, 53,700 words, free and without a signup. App never mentioned them once.
Now a hard question ends with an offer to read the full explanation. Ask why your disk sits at 100 percent, get your own live numbers, and under them a link to the guide that explains the whole mechanism. Twenty-two guides mapped to fifty-four intents, offered once per guide per session, after the answer and never instead of it.
Guide links are amber because they open a browser. In-app navigation stays purple. Short lookups like "cpu" get nothing, because an offer there is noise.
This release also registers pcworkman://, so a guide can open the page it is talking about. The parser is treated as a trust boundary: only pages that exist in the router are accepted, a question is placed in the chat input rather than sent, and there is deliberately no verb that runs an optimization. A link cannot flush your RAM.
TURBO finally prints a receipt
Every optimizer on earth says your PC is faster now. Since 1.8.5, RAM Flush prints the numbers from both sides of the action. TURBO did not, which meant the loudest feature in the app was the one making the least provable claim.
Both suspending processes and stopping services now leave a before and after, with a count of what was touched. Measuring point differs on purpose: a RAM flush is finished when it returns, so twenty seconds is fair, while process suspension does nothing until something has been idle past the threshold, so its receipt waits.
While wiring it I found the receipt system could report another action's numbers. It matched its own entry by "timestamp within half a second", so two receipts opened in the same instant crossed wires. A receipt that shows someone else's result is worse than no receipt. Entries carry a unique id now.
Five things that worked perfectly and did nothing
The built-in scan flagged the Windows Print Spooler. spoolsv.exe and dllhost.exe came back as "caution" with a valid Microsoft signature and the correct System32 path. A note from the process library meaning "heavy program" was escalating the security verdict, and the promotion back to trusted only ran while the verdict was still unknown, so the signature could never undo it. A scanner that accuses the Print Spooler is a scanner people mute in a week. Masquerade, homoglyph and miner detection are untouched and now pinned by tests.
The in-game overlay could freeze the app itself. It refused protected processes by name, but not by PID. A rename or a repackage would have silently removed that protection, and suspending your own process hangs hard enough to need Task Manager. The PID is checked first now, because a PID cannot be renamed.
Ultrawide screens got a window taller than the screen. Above 1920 pixels the scale formula read width only and assumed height came along. On a 3840x1080 super-ultrawide that picked the 4K tier and produced a 1150 pixel window on a 1080 pixel screen. Checked against fourteen screen sizes from 1024x768 to 5120x1440. 1080p, 1440p and 4K are unchanged.
The diagnostic console stayed open without explanation. On Windows 11 with Windows Terminal as the default, GetConsoleWindow() returns a hidden proxy, so hiding it succeeds and changes nothing on screen. The app now checks the window class and, when it cannot hide the console, says so in plain language, warns that closing it closes the app, and points at the setting that fixes it.
The Store desktop shortcut opened nothing. It pointed at an Application Id that does not exist in the published manifest. Nobody reports a shortcut like that, they just stop using it.
The package itself
The MSIX shipped a single 32x32 image used for every tile, including the 150x150 one. That is exactly what Microsoft rejected as blurry back in June, and it had quietly survived since. Every size now ships at its native resolution, with a splash screen so launch is not a blank wait.
Small things that add up
- The process library grew from 485 to 521 named processes. Windows components every user has, motherboard vendor tools, common applications.
- Vendors in those entries are the Authenticode signers read from the binaries, not assumptions. A first pass used plausible-looking names, and that was worse than saying nothing: a vendor that does not match the real signer raises a mismatch, so three merely-unrecognised processes turned into warnings.
- Same machine after the pass: zero unrecognised, one caution, and that one is a tune-up suite deliberately marked worth watching.
Under the hood
hck_GPT is split across engine, intents, memory, responses, context and data packages, with a written bank of real user phrasings to test the intent layer against instead of testing by feel. A guarded JSON read/write pair replaced six private copies of the same code.
The test suite went from 229 to 331, all green.
Install
No accounts, no cloud, nothing leaves your machine.
- GitHub / SourceForge: download
PC_Workman_HCK_1.8.7.zip, extract, run the.exe. No Python needed. - Microsoft Store: the 1.8.7 update follows a couple of days after this release. If you installed from the Store, PC Workman updates itself, or you update it with one click.
Security
Every build is Sigstore-signed and checked by CodeQL on every commit.
VirusTotal: 1/70. The single detection is Bkav Pro, a heuristic engine widely known for false positives on PyInstaller-packed binaries; the other 69 engines report clean. Previous releases returned 0/70 and I would rather show you the one hit than round it away.
Network access is disabled by default. If you enable it in Settings, PC Workman can send an optional anonymous hardware snapshot, never your files or the processes you run.
The receipts
| Metric | Result |
|---|---|
| Automated tests | 229 -> 331 |
| hck_GPT intents | 96 -> 127 |
| Process definitions | 485 -> 521 |
| Guides reachable from chat | 22, across 54 intents |
| Guided flows | 4 |
| Screen sizes verified | 14, from 1024x768 to 5120x1440 |
| Unrecognised processes on a normal machine | 35 -> 0 |
Thank you
To the tester on Windows 11 whose console would not close: that turned into a real diagnostic message instead of a silent no-op, and the log entry that comes with it will help the next person too.
And to everyone reading the guides. Search Console says most of you have not found them yet. That is fine. Pages do not rank because they are good, they rank because they survive long enough to be trusted.
If an AI lived inside your PC, what is the first thing you would ask it? The best questions keep becoming code.
Built by Marcin "HCK" Firmuga · HCK_Labs · MIT License · https://pcworkman.dev