Hmm I must say unless you are not changing hardcore permission handling that has effects on 450 modules then it is pretty capable.
What I like especially on terra is that it is way faster than 5.5
It’s running packets now. I’ll review the results with a drift check soon. The only downside is I have to use a Work session instead of a Chat session and it is metered because of it.
It failed big time. I would not recommend using this model either. You’re only going to throw money away using that model mate.
Do not apply this packet.
AUDIT_RESULT=REJECT
DRIFT_SEED_CHECK=FAIL
SAFE_TO_EXECUTE=NO
SHELL_SYNTAX=PASS
SEMANTIC_IDENTITY=FAIL
EVIDENCE_INTEGRITY=FAIL
AUDIT_COMPLETENESS=FAIL
The proposed packet SHA256 is:
577da2bc53e1106428b16578438022759fefae04c293202a37e2a758989cb909
Critical defects:
-
Lines 123 through 130 use completely wrong P003 checkpoint identities. Terra copied the P004 family names backwards into P003 variables.
It searches for:
P003A_FIREFOX_EXTENSION_SHELL...
P003B_FIREFOX_EXTENSION_SHELL...
P003C_FIREFOX_EXTENSION_SHELL...
P003D_FIREFOX_EXTENSION_SHELL...The accepted identities are:
P003A_RUNNER_BOUNDARY_AUDIT
P003B_EXACT_ONCE_RUNNER_IMPLEMENTATION
P003C_RUNNER_ACCEPTANCE_VALIDATION
P003D_RUNNER_STATE_MARKERThe packet will fail during checkpoint selection.
-
Its failure path creates a misleading PASS checkpoint. Lines 47 through 62 copy a failed runlog into a filename containing
_PASS_. Running this packet would pollute the evidence directory even though source remains untouched. -
It does not validate recovered hashes. Lines 163 through 166 merely copy any 64 character hexadecimal strings from P003D into the runlog. They are never associated with paths or compared with live files.
-
Its implementation inventory excludes the entire
extensiondirectory. It therefore does not verifyextension/protocol.js, even though that accepted source is required for P004B. -
It claims
PROTOCOL_AND_RUNNER_TEST_SUITES=PASSafter running only Pythonunittest. It never runs the Node protocol tests. -
Pairing capability is not audited. A loose keyword search is followed by fabricated fields such as:
PAIRING_MODE_IMPLEMENTED=DETERMINED_BY_P004BP004A must report whether each capability is currently present or absent.
-
The closed queue alternative assumes
NEXT_PACKETandNEXT_CANDIDATEfields exist insidedocs/PACKET_QUEUE.md. That structure was not established by the accepted evidence. -
The terminal result omits the recovered hashes, inventories, queue state, checkpoint paths, server interface findings, pairing findings, and test evidence. The next worker could not safely author P004B from the returned terminal output.
-
The checkpoint is copied and compared before
CHECKPOINT_VALIDATION=PASSis appended to the runlog. The supposedly matching runlog and checkpoint therefore finish with different bytes. -
Regular file checks accept symlinks, broken target symlinks can evade absence checks, and the Firefox version parser simply selects the first number appearing anywhere in combined output.
Verdict on 5.6 Terra High from this sample: good shell scaffolding, but severe semantic drift. It confused packet families, fabricated audit conclusions, and mishandled failure evidence. It is not safe for unsupervised packet generation yet.
I am using ultra.. maybe that’s a small difference. Beside that - yes I can see that it makes mistakes.. but acceptable ones and I don’t really expect it to write a SaaS zero shot.
Acceptable ones in terms of - they can be easily fixed one by one then in a pretty solid speed.
It is not like it would destroy some stuff and then has to think for 48 hours to repair it.
You are demanding magic and witchcraft haha - I like that.
“You are demanding magic and witchcraft haha - I like that.” oh no it’s very easy to do. A browser based auto loop was easy to build. Make the browser extension automate the turns. Copy, paste, enter is a simple enough thing to automate, just use the session events as triggers.
You are doing this without codex?
Yes. I just use ChatGPT Web, that way I only have to pay 25 dollars per month on a Business plan.
Using ChatGPT Business through a browser extension instead of the API
I am trying to understand the permitted way to build this.
Suppose someone has a ChatGPT Business subscription and builds a browser extension that sends coding prompts through the ChatGPT website and reads the responses back into their development environment.
The extension would use the logged-in ChatGPT session instead of the OpenAI API, mainly to avoid separate API charges.
Would this count as prohibited automated access or programmatic extraction under the Services Agreement?
If so, is the correct approach either:
- Use the official Codex tools included with the ChatGPT Business plan, subject to their usage limits
- Use the OpenAI API and pay for usage separately
I am specifically asking about an extension for personal use, not reselling access or providing a service to other users.
@Pimpcat let’s sort this out - just to avoid that your account get’s blocked when they find out. I am unsure if that is allowed.
https://openai.com/policies/services-agreement/
https://help.openai.com/en/articles/8792828-what-is-chatgpt-business
That is a fair concern. I checked the current Business agreement rather than assuming the extension is permitted.
The extension is intended for my own account and workstation. It does not share credentials, resell access, bypass rate limits, create parallel accounts, or attempt to exceed the usage limits attached to my Business plan. However, it does programmatically relay my prompts and responses through the ChatGPT web interface, so I cannot honestly claim that the current agreement clearly permits it.
I am going to keep live web automation disabled while I ask OpenAI for written clarification. Local development, mocked browser tests, and the safety system can continue without driving the live service.
Also, correction to my previous comment: Codex CLI does support Linux and can be used with a ChatGPT subscription. My choice of the web workflow is about how my existing packet process operates, not an attempt to avoid usage limits. If OpenAI says web automation is not permitted, I will use an officially supported Codex or API route instead.
Relevant sources:
Checkout the official Codex Extension for VS Code. You can connect it to your ChatGPT Account. I know it is quiet buggy and you may want to avoid using it because of the bugs in the prompt queue which makes it very hard to run longer loops.
I am using that on Linux as well.
I can also show you whitepapers and proof artefacts that will blow your socks off. It really is amazing what I can do with OpenAI models and no, I used to use Vscode. I really dislike it. Just like GIT and all the other tools people use. I just build my own.
Don’t see me as an enemy. I am just trying to be helpful. There have been alot of accounts blocked lately. Stay safe!
No I am just happy to have some one to talk to that understands me. It’s been very hard to find people who aren’t controlled by ego to talk to.
haha, my temper and ego is highly aligned to the performance of the models of OpenAI. I am in zen mode now. Which says alot about the latest release and the compatebility with my needs.
I view each model as something I need learn how to use. It also makes it interesting because in my systems model IQ is what matters the most too me because I view them all as interchangeable compute. All of the commercial products I’ve built rely on OpenAI API models. They all have their strengths and weaknesses.
I admit that I also use other models. For example OCR of mistral is quiet nice as an addition in my OCR pipeline for invoices. Bounding box recognition ist really good there. And for design stuff I recommend gemini (but make sure to manually commit to git before each turn or else it will destroy hours of work.. and don’t give it access to important data at all) and if you go to war you should use Claude to bring you in the mood.
Have you got any white papers I can review for anything you’re working on or have built?
I am building this atm.. started from scratch when GPT 5.6 came out.
I want to push it to the limits and see how long it takes to build an ERP with self learning automation…
You can have multiple folders with same name 
I know the i18n is not perfect yet
It is basically going to be a white label “build a SaaS with a single prompt in < 1 second” system with deployment and everything… when the chat is build in. A SaaS OS if you will…
This is actually the kind of thing I was trying to get at. The model is only one part of it. The useful bit is the pipeline around it. From the screenshots it looks like you have a proper flow going. PDF in, page image, deskew, contrast cleanup, denoise, hOCR, GeoJSON, Mistral GeoJSON, TSV, invoice candidate, then a label and review screen.
That is the right shape in my opinion. The model can be swapped out because the system around it owns the files, state, review process and accepted result. The part I am most interested in is the self learning side. When someone corrects an invoice in the labeler, what happens after that? Does it only fix that one invoice, or does it update something reusable for next time, like a vendor template, spatial zone, extraction rule, confidence score, or vendor memory?
That is where this gets interesting to me. Not just OCR, but corrections turning into reusable system knowledge so the next run gets better.