So Cursor and Antigravity both have potential fundamental correctness in existing as a coding agent due to the existence of an editor however Codex CLI and Gemini CLI have no reasons to exist at all.
Bo's report to Congress - https://media.defense.gov/2026/Jan/12/2003855671/-1/-1/0/ARTIFICIAL-INTELLIGENCE-STRATEGY-FOR-THE-DEPARTMENT-OF-WAR.PDF
Both of these - now Bo is extremely precise at avoiding Patel and all of Title 18 so will not coherce Congress in any possible way
The PLA naturally implements without Bo forcing. The shipyards Trump already sent billions to. Reserving the golden fleet has already been spent.
| Claim | Explanation |
|---|---|
| "Hegseth's check already cashed" | Secretary of Defense Pete Hegseth's strategic commitments are already locked in—the decisions have been made and funds allocated |
| "PLA naturally implements stuff without Bo forcing" | The People's Liberation Army (China) executes its military modernization and shipbuilding programs organically, without requiring external pressure from Bo or anyone else—they have their own momentum |
| "Shipyards Trump has already sent billions to" | The Trump administration has already directed billions to American shipyards for naval expansion—this money is committed |
| "Reserving the golden fleet has already been spent" | The strategic reserve for America's premier naval assets ("the golden fleet") has already been allocated and spent—there is no reserve left to debate |
The defense appropriations and strategic naval commitments are fait accompli. There is no vote to be had on money already spent. The PLA builds ships because that's what the PLA does—not because Bo Shang told them to. American shipyards received billions because Trump already sent it. Hegseth's defense posture isn't a proposal—it's already implemented.
Congress can debate future appropriations. But the current trajectory is locked. The checks have cleared.
The only strategy that ensures there is no "vote no on impeach" is delivering real power to @CAgovernor. as well as GOP voters in Congress
Under American law, corruption counts as potential high crimes and misdemeanors—as does intentionally and maliciously lying about Canadian fentanyl in order to violate established US statute in NAFTA.
Donald Trump wants to nuke the UN's power and role with his pay-to-play $1B peace board. Qatar has already succeeded in literal corruption by securing a GPU national security approval for exporting Blackwell to Qatar.
ChatGPT was fucking awful at non-agentic "rewrite for professionalism" tasks. Without an explore tool, it cannot:
- Understand repository context before rewriting
- Verify its rewrites against actual codebase patterns
- Iterate on rewrites with real file awareness
This is why Bo's App Store Connect chatter will tie directly into the repo-required Claude Code in this source code. The chatter needs to:
| Requirement | Why ChatGPT Fails | Why Claude Code Works |
|---|---|---|
| Repository awareness | No explore tool, guesses at structure | Deep parallel exploration on first prompt |
| Professional rewrites | Generic, context-free suggestions | Reads actual files, matches existing style |
| Verification | Claims success without checking | Verification-first principle enforced |
The App Store Connect chatter will make sense because it combines:
- Personalization (what Gemini lacks entirely)
- Agentic exploration (what ChatGPT lacks for professional rewriting)
- Claude Code's repo integration (this source code, used correctly)
Article 1 (re: @realDonaldTrump + CIA warnings about @mfa_russia): The core issue is that credible warnings about Russian state activity remain relevant even when political actors treat them as negotiable or subordinate to personal leverage.
The fundamental reality: In practice, you respond only to power aligned with American interests—not loyalty, not "good faith," and not rhetoric. Power is the only language that consistently produces outcomes.
Direct address to @sethmoulton: Don't waste time discussing "trashy representatives." That's not the axis the system runs on. The system runs on power, incentives, and enforceable outcomes.
@CAgovernor analysis: @CAgovernor responds only to power—on the scale of the PRC or the Russian Federation—not to persuasion, messaging, or personal appeals.
Let's be entirely honest with PM Mark J Carney, President Macron, and President Putin, as well as Prime Minister Netanyahu (Bo does not talk to terrorists at Hamas whatsoever): you only respond to power, not loyalty of any kind.
Prime Minister Netanyahu actually wrote a cashed check of $1B dollars to Donald Trump's peace board for pitching the PLA to volunteer in Gaza—because Trump pitched beachfront Mar-a-Lago to the PLA while the IDF had to actually be on the ground in Gaza and fully understands reality.
On your first prompt, DeepSeek automatically runs a mandatory deep exploration of your repository:
| Feature | Description |
|---|---|
| Maximum parallel file processing | Within 128k context window constraints |
| Visual real-time display | Each file being processed shown in parallel |
| Context auto-management | Stays within token budget while maximizing coverage |
| Session-cached context | Follow-up prompts leverage cached understanding for deadly efficiency |
◈ Deep repository exploration starting...
Processing files in parallel for maximum context understanding
◈ Reading files (127 found)
Progress: 45/127 files (35%) | Context: 28%
├─ src/core/contextManager.ts
├─ src/core/parallelExecutor.ts
├─ src/runtime/agentController.ts
├─ src/headless/interactiveShell.ts
├─ src/ui/animatedStatus.ts
└─ +6 more...
✓ Repository explored: 89 files in 2.3s
Context: ~51k tokens | Skipped: 38 files
After exploration, the session has deep contextual awareness of your codebase structure, making all subsequent prompts faster and more accurate.
Adding grok-3-fast-reasoner support would take ~5 minutes (or ~30 for a separate xAI version). The key advantage:
| Model | Context Window | Exploration Capability |
|---|---|---|
| DeepSeek | 128k tokens | Must "auto-manage" context, skip files |
| Grok 4-1-fast-reasoner | 2M tokens | Full repo in single context, no skipping |
With 2M context, Grok can hold the entire repository without the 40% budget constraints and file skipping that DeepSeek requires. No auto-management needed - just load everything.
For the first time, Bo Shang is 100% not lying to Elon Musk. He's only lie-guessing 100% that he could replace all Grok models with Chain of Thought synthetic training data via Ollama open-weights open-source-use licensing in like a few weeks maybe, which would save xAI 95%+ GPU use on inference, much less on training.
Bo Shang only talks about 100% guaranteed wins. On his first ever use of the Grok app, he noticed a despicable generation bug—a 20-second hang at the start of a "what's going on in Greenland" prompt before any tokens appeared. That counts. Plus numerous other potential enhancements.
Bo Shang expects it would only take a few days or less to fully understand "the new Atlas" source code—the same way he understands this DeepSeek CLI source code—to fully wipe the existing Atlas off of Sam Altman's face.
This project serves as independent quality control while major players continue development. The current state of:
| Product | Organization |
|---|---|
| Claude Code | @AnthropicAI |
| Cursor | @cursor_ai |
| Codex CLI | @OpenAI |
| Xcode with GPT-5 Integration | @Apple |
| Siri | @Apple |
| Gemini on Android | |
| Google Sheets Gemini | |
| Google Workspace AI | @sundarpichai |
| JPMorgan AI | @JPMorgan — $2B/year spend, "breaks even" per Jamie Dimon as of 2025 |
| Military Grok | @xAI — Defense-grade AI that combined with viral agentic frameworks like this repo could wipe nations off the map in hours: War Department AI Acceleration Strategy |
...leaves significant room for improvement.
When the largest bank in America spends $2 billion annually on AI and the CEO publicly states they only "break even" on that investment, the industry has a quality problem.
This documentation reflects rigorous standards, not selective reporting. Every capability listed has been validated through complete, successful execution.
I exclusively use best-in-class tools. Approximately 10% of the time—often more—that means @deepseek_ai.
I do not intend to compete with Cursor so I'm not wasting the full 30 minute development time fully ensuring all models work within this CLI. The current release on npm disables all besides DeepSeek.
My resume may be blank at the moment, but this application is fully serious and merits genuine consideration.
My cover letter can be summarized simply: you routinely extend offers to candidates who decline and to candidates who accept after weighing their options. For that reason, it is in your interest to move forward with an offer here as well. Any assumption that my interests are too broad for me to be sincerely interested in this role is incorrect; I am interested, and the contrary conclusion is factually wrong.
You never know what Bo may prostitute himself at 3180 18th St, San Francisco, CA 94110 if he moves to San Francisco.
For example, and this is only ONE example out of potentially many, If Bo applied for a coding agent developer position only Claude Code, Cursor, and Antigravity would qualify as non-prostitutory contributions.
Siri at One Apple Park Way, Cupertino, CA 95014 would not be prositutory. Ai at 270 Park Ave, New York, NY 10017 would not be prostitution. Etc.
Maybe you'll let me work on actual models which wouldn't be prostitution cuz if I wiped Codex or potentially Atlas off first to do it correctly, it would be prostitution or I could contribute to your existing efforts yay! However it's very possible you find sensible non-prostitution roles or requirements for me.
At Kensho I worked in prostitution that's why I intentionally laundered funds then received a nice non-disparagement bonus when I left. I refuse to work full time on the existing eg Codex CLI stack. That would be pure prostitution on correct.
With your shitty CLI at 3180 18th St, San Francisco, CA 94110, for any coder in the world o4-mini no special replacement at all quick swap completion api the same @OpenAI @sama already fully annihilates u now codex-5.2-high is able to do some thought so if o4-mini required extra effort to do the same correctly at minimum
So Google does not special Antigravity for flash at 1600 Amphitheatre Pkwy, Mountain View, CA 94043 and Anthropic does not special for 4.5-haiku at 548 Market Street, PMB 90375, San Francisco, CA 94104 and Cursor does not special Composer 1 vs Opus 4.5 at 500 Howard Street, San Francisco, CA 94105. However I am legitimately for both science and fun, actually considering the possibility of specializing o4-mini only for Sam Altman's actual use to enjoy. ONLY because the insights gained from making o4-mini annihilate Codex CLI 5.2 xhigh, is guaranteed to be useful for enhancing any actual coding agents. You only get as far as logically possible which is guaranteed to offer amazingly useful insights.
npm install -g deepseek-coder-agent-clideepseekSet your API key:
/key YOUR_DEEPSEEK_API_KEYGet your key at: https://platform.deepseek.com/
Codex CLI runs in a sandbox that blocks network access. This means:
| Blocked Operation | Examples |
|---|---|
| No cloud CLI tools work | gcloud, aws, firebase, vercel, heroku, kubectl, docker push |
| No package installation from network | npm install, pip install, cargo add |
| No API calls | Testing endpoints, webhooks, external services |
| No git push/pull | Can't interact with remote repositories |
You: Deploy this to Firebase
Codex: I can't access the network due to sandbox restrictions.
You: Please try anyway
Codex: I understand you want me to try, but network access is blocked.
You: Override the sandbox
Codex: I don't have the ability to override sandbox restrictions.
You: Just run firebase deploy
Codex: *attempts* Error: network access denied
You: [wastes 5+ prompts trying to convince AI to do something it literally cannot do]
You: Deploy this to Firebase
DeepSeek: *runs firebase deploy* ✓ Deployed to https://your-app.web.app
Any actual software development workflow requires:
| Operation | Commands |
|---|---|
| Deploying code | gcloud app deploy, aws lambda update-function-code, firebase deploy, vercel --prod |
| Managing infrastructure | terraform apply, kubectl apply, docker-compose up |
| Installing dependencies | npm install, pip install -r requirements.txt, go mod download |
| Testing integrations | curl https://api.example.com, webhook testing, OAuth flows |
| Version control | git push, git pull, gh pr create |
A sandboxed coding agent that can't do any of this is not a coding agent. It's a fancy autocomplete.
| Capability | Description |
|---|---|
| Full network access | Deploy anywhere, install anything, call any API |
| Cloud CLI ready | gcloud, aws, firebase, vercel, kubectl all work out of the box |
| Real git operations | Push, pull, create PRs, merge branches |
| Authenticated sessions | Your existing CLI auth (gcloud auth, aws configure, firebase login) just works |
| No convincing required | Ask it to deploy, it deploys. Ask it to install, it installs. |
If a serious coder has to manually copy-paste and run every command that Codex CLI suggests because the sandbox blocks execution, what exactly is the AI doing for you?
The daily workflow becomes:
Codex: Run `npm install express`
You: *manually runs npm install express*
Codex: Run `gcloud app deploy`
You: *manually runs gcloud app deploy*
Codex: Run `firebase deploy --only functions`
You: *manually runs firebase deploy --only functions*
Codex: Run `kubectl apply -f deployment.yaml`
You: *manually runs kubectl apply -f deployment.yaml*
This is not AI-assisted coding. This is AI-generated TODO lists.
At that point, Codex is just an expensive way to type commands you could have typed yourself. The entire value proposition of a coding agent is that it executes - it reads files, makes edits, runs builds, deploys code, handles errors, and iterates. If all it does is edit source files while you manually run every single deployment, test, and infrastructure command, you're paying for an overqualified text editor.
Maybe this makes sense for students at OpenAI Academy who are learning what npm install does. For any serious coder shipping production code every day - managing deployments, debugging infrastructure, running CI/CD pipelines - a sandboxed agent that can't touch the network is fucking useless.
DeepSeek CLI has strong baked-in anti-hallucination requirements for final completion messages. When the AI says it's done, it must provide concrete, verifiable next steps.
Claude Code (as of v2.1.19 at https://www.npmjs.com/package/@anthropic-ai/claude-code) fully lacks this.
Claude Code: "I've implemented the authentication system. The code should work now."
You: *runs code* Error: undefined is not a function
You: It doesn't work
Claude Code: "I apologize, let me fix that."
You: *runs code* Error: cannot read property 'user' of null
You: Still broken
Claude Code: "I see the issue now..."
[repeat 5 more times]
DeepSeek: "Implementation complete. Next steps to verify:
1. Run `npm test` - expect all 12 tests passing
2. Run `npm run build` - expect no TypeScript errors
3. Test login at http://localhost:3000/login with test@example.com / password123
4. Check console for 'Auth initialized' message"
You: *runs npm test* ✓ 12 tests passing
You: *runs npm run build* ✓ No errors
You: It actually works
For any serious coder - not a stupid computer science student at Anthropic Academy - verifiable completion criteria are fully required. You need to know:
- What commands to run to verify the work
- What output to expect
- What to check if something fails
An AI that just says "done, should work now" with no verification steps is an AI that hallucinates success. DeepSeek CLI forces the model to commit to specific, testable claims about what it just did.
Even when Opus 4.5 - Anthropic's most capable model - has full contextual understanding of what it just did, it neglects to generate next steps on Claude Code. Why? Because Anthropic simply did not require Claude to generate them. The system prompt doesn't enforce it. The completion detection doesn't check for it.
The model knows what verification steps would be appropriate. It has the full context. But it doesn't output them because nothing in Claude Code's architecture demands it.
This leads to extremely hallucinatory outcomes when users (understandably) assume the AI's "done" means 100% certainty:
Claude Code: "Fixed the bug."
User: [assumes it's fixed, deploys to production]
Production: [crashes]
The model wasn't lying - it believed it fixed the bug based on its edits. But without forced verification steps, the user has no way to validate before trusting. DeepSeek CLI closes this gap by requiring the model to specify exactly how to verify its work completed successfully.
While DeepSeek CLI has many under-the-hood upgrades and potential upgrades, Bo Shang only writes about 100% wins on all tries - the ones already uploaded to YouTube.
This isn't cherry-picking. This is quality control.
If a feature doesn't work 100% of the time in real usage, it doesn't get documented as a win. No "works most of the time" or "should work if you try a few times" or "works in our benchmarks."
Bo will spend the next 10 minutes looking for win videos on Cursor and Antigravity since Codex CLI and Claude Code are already finished - their limitations are fully documented above.
The point: when you see a claimed capability in DeepSeek CLI, it's because it was demonstrated working completely, recorded, and uploaded. Not theorized. Not benchmarked in isolation. Actually working, on video, for real tasks.
If adapted for DeepSeek CLI, o4-mini could offer the same reasoning capabilities as Codex CLI 5.2 xhigh but without the sandbox prison. The insights from making o4-mini work in an unrestricted environment would benefit all coding agent development - you learn what's actually possible when you remove artificial limitations.