CEO @creately, ok builder, learning distributor, pro lurker.
Joined February 2008
- Tweets1.4K
- Following404
- Followers1.2K
- Likes422
Pinned Tweet
If your agent harness (Claude Code, Codex, OpenClaw) can read your secrets, it can send your secrets.
That’s just how context-based agents work: they ingest local files and send request payloads upstream to figure out what to do with it.
I built mirage-proxy to fix this at the network layer.
It sits between your agent and the LLM API, replaces secrets/PII with plausible fakes before transit, and rehydrates responses locally.
So the model still sees realistic data, your workflow stays the same, and your real keys don’t leave your machine.
• single Rust binary (~5MB)
• sub-ms overhead
• 129 secret patterns + entropy checks
• session-consistent mappings
• daemon + shell integration (mirage on/off/status/logs)
Install:
brew install chandika/tap/mirage-proxy && mirage-proxy --service-install
Open source (MIT):
github.com/chandika/mirage-p…
Chandika Jayasundara retweeted
it’s in gemini, just create it in ai studio. oh, that’s for your personal google one account. for workspace you need gemini business. no, not gemini advanced, that’s ai pro now. unless you need ai ultra. oh agents? you do that in spark actually. no, not gemini api managed agents, that’s different. for coding use jules. unless you mean the agentic ide, that’s antigravity. no, that’s the old antigravity, download the new one. actually gemini cli is being deprecated, use antigravity cli. no the flash model is smarter than the pro model. unless you need pro. if it’s video, use flow. no, flow uses veo. no, nano banana is images. actually that’s in gemini now. unless you’re in search, then it’s ai mode. no, research is notebooklm. anyway it’s all very simple.
Chandika Jayasundara retweeted
If you thought the Chinese models were good, just wait a couple of months, now that all the distillation poisoning has been removed from Claude Code.
Diffusion LLMs may not beat benchmarks but would be awesome for fluid UX where the task is semi-deterministic.
Mercury 2 is live 🚀🚀
The world’s first reasoning diffusion LLM, delivering 5x faster performance than leading speed-optimized LLMs.
Watching the team turn years of research into a real product never gets old, and I’m incredibly proud of what we’ve built.
We’re just getting started on what diffusion can do for language.
3
79
Can we trust agents to pick their own models?
Built an OpenClaw skill that does this. Reads benchmark tables, fetches model cards when you add new providers, rewires its own routing.
4
1
109
Three modes: "work harder" goes full Opus. "Save money" drops to DeepSeek (107x cheaper). Default runs the smart mix.
clawhub install chandika/sonnet-model-router
github.com/chandika/openclaw…
105
The data gives you directional truth:
Sonnet beats Opus on office tasks (1633 vs 1559 Elo)
Opus wins on hard reasoning (75.2 vs 58.3 percent ARC-AGI-2)
Claude destroys GPT on computer use (72.5 vs 38.2 percent)
Start there. Adjust based on what actually works in your setup.
84
Caveat: benchmarks arent real-world performance. A model that scores 80 percent on SWE-bench might choke on your codebase. So the routing is a starting point, not the final answer.
But its a better starting point than "use Opus for everything because its the most expensive."
51
Add Gemini to your setup. The router detects it, pulls the benchmarks, compares against whats running, tells you where it fits.
No config. No dropdown. It figures it out.
37
Sonnet 4.6 dropped a few hours ago. Published a model router skill for OpenClaw — routes Sonnet for tasks it wins at, Opus for the hard stuff.
Why: Sonnet 4.6 is at near-parity with Opus on most tasks. 5x cheaper. 1M token context. Matches Opus on office tasks and document comprehension. Beats prior Opus on computer use. SWE-bench: 79.6% vs 80.8% — basically a tie.
Opus still owns deep reasoning and novel problems.
The thing nobody talks about with benchmarks — stop reading them as leaderboards. They're routing tables. Each row tells you which model to send which job to.
One config change in OpenClaw — subagents run Sonnet, main session stays Opus. Background work goes to the model that's actually better at it. Rate limits stretch further.
```clawdhub install sonnet-model-router```
The broader point: we're in the multi-model era now and most people are still using AI like it's 2024 — one model for everything.
The winners will be the ones who treat model selection like infrastructure, not religion.
github.com/chandika/openclaw…
1
49
Update: v2.1 is live. The model router now teaches itself.
Add a new model to OpenClaw — Gemini, GPT, DeepSeek, whatever — and it:
1. Detects the new model automatically
2. Fetches its model card + benchmarks
3. Compares against your current routing table
4. Tells you exactly where to use it (or not)
Three modes: "work harder" → Opus. "save money" → Gemini/DeepSeek. "normal" → smart mix.
It also ships with a seed registry of 9 models and cross-provider benchmark data. Computer use? Always Claude (72.5% vs GPT's 38.2%). Deep reasoning? Always Opus. Office work? Sonnet beats everything — including Opus.
The meta point: your agent should pick its own models the same way a senior engineer picks their tools. Not by brand loyalty. By the data.
github.com/chandika/openclaw…
‘Our most capable sonnet model yet’ 😂 yeah no shit, why’d you ship a version otherwise!
This is Claude Sonnet 4.6: our most capable Sonnet model yet.
It’s a full upgrade across coding, computer use, long-context reasoning, agent planning, knowledge work, and design.
It also features a 1M token context window in beta.
73