Ashkan Goshtasb (@Ashkon) header
Ashkan Goshtasb (@Ashkon) profile picture

Self-Designated Adult In-Charge and Chief Concept Exorcist. Ashkan.x

Los Angeles, CA
Joined June 2016
Photo by Ashkan Goshtasb (@Ashkon)Photo by Ashkan Goshtasb (@Ashkon)Photo by Ashkan Goshtasb (@Ashkon)Photo by Ashkan Goshtasb (@Ashkon)Photo by Ashkan Goshtasb (@Ashkon)Photo by Ashkan Goshtasb (@Ashkon)Photo by Ashkan Goshtasb (@Ashkon)Photo by Ashkan Goshtasb (@Ashkon)Photo by Ashkan Goshtasb (@Ashkon)Photo by Ashkan Goshtasb (@Ashkon)
🤖 Made with AI
15
Blind test result, as promised: my council lost. On the PR where CodeRabbit caught a config-folder bug, neither OpenCouncil nor a single Claude caught it. The council took 7.5 min, one Claude took 1 min, and they found the same things. That was quick review, one call per seat. Next: the full spec-debate loop vs one agent, on hard tasks.
1
1
80
Found a bug in my own tool. OpenCouncil's Gemini "plan" seat is meant to be read-only. It wasn't: another CLI flag quietly turned plan mode off, and a test run wrote a file. Fix: run Gemini's plan mode directly, allow reads, refuse writes. github.com/goshtasb/OpenCoun…
72
My AI council signed off on a fix for a stranger's bug. Then CodeRabbit found it hid models from some logged-in users. The council had marked the shaky fact as unverified, and nothing checked it. What broke and what I'm changing: ashkan.us
1
54
a single Claude agent beat the council 10/10 to 7/10 on easy tasks, and a blind test on that same diff that I'll post either way once it's scored.
36
Ran my 3-model council on a stranger's bug in CloudCLI. It wrote the fix, tests proved it, PR is up. Then CodeRabbit reviewed it and caught a config path case the council missed. Now running a blind test: council vs one Claude on the same diff. Posting the result either way.
1
49
Open Council reviewed its own PR this week and found a real bug in itself. The verdict line could say "no objection stands" while the same report listed objections that still stood. One seat raised it, another upheld it, and it got fixed before merge. That's the reason I run reviews through more than one reviewer.
32
Pointed my AI council at an open PR in someone else's repo today. 5 objections. I wrote failing tests before believing any of them. 2 reproduced. The good one: set a .env value that starts with a backtick, and the keys after it quietly disappear. Delete that key later and they're gone for good. One-character fix. npx open-council-cli review
41
Is this true @Android? Why do we always get the short end of the stick?
This is UNACCEPTABLE, Google! Google Maps on iPhone Duo vs Android
1
42
I don't understand, @OpenAI releases a new GPT version on a monthly basis or something? why does @GeminiApp take forever for new releases?
Introducing GPT‑6 Cyber: a major advance in security research that pairs stronger capabilities with better model alignment. The model powered the Linux kernel research we announced yesterday, and security teams at companies like Mozilla, Palo Alto Networks, Calif, and Cisco—alongside our own at OpenAI—are already putting it to work. GPT-6 Cyber is available to select Daybreak Red customers and we are expanding access over time. openai.com/index/introducing…
34
My coding agent's harness ran the build and the tests. Both passed. It opened a PR. GitHub CI failed 14 seconds later: error TS2304: Cannot find name 'config' Nothing lied. It checked the wrong code.
1
2
52
The harness tested the branch. GitHub tests the merge of your branch into main. Only one of those is what actually lands.
1
11
Fix: merge the current base before running any gate. A conflict ends BLOCKED and nothing is pushed. About 20 lines and a regression test. Full write-up: ashkan.us/ Repo: github.com/goshtasb/OpenCoun…
10
How a run of my multi-model CLI goes: Gemini drafts the spec. Claude Code checks it against the real code and pushes back. Grok breaks ties. I approve, then Claude builds it, runs the gates and opens a PR. It takes 10–40 min and it's slow on purpose. Most of the bugs it finds come out while the three are still arguing. github.com/goshtasb/OpenCoun…
1
1
4
91
every AI agent demo is a todo app. show me one that survives a real customer's CSV man
3
457
The demo is easy. The feedback loop is the product.
1
426
Spun up a repo where Gemini, Claude, and Grok CLIs fight over a PRD until they all sign it. You approve. Then they build, verify, and open the PR. No APIs, just the subs you already pay for. Open source. Break it, improve it, send a PR. github.com/goshtasb/OpenCoun…
1
395