Christo Wilken

[Christo Wilken @TweedBeetle]

The SCIF where it happens 🎶

[Christo Wilken @TweedBeetle]

@ClaudeDevs God, anthropic vids are all so nice to look at. Note the claude-colored cardigan an eyeshadow 🔥

[Christo Wilken @TweedBeetle]

v0.1, npm-only, free, MIT, built in public. Add it in one line:
claude mcp add dependency-fitness -- npx -y dependency-fitness-mcp
https://t.co/GKcth3atQU
I'd be curious whether the migration-target inference is the useful part, or if you'd rather just have the raw deprecation flags.

[Christo Wilken @TweedBeetle]

Coding agents will confidently add an npm dependency that's deprecated, a major version behind, or that doesn't actually exist.
I built a small MCP server that gives the agent a fitness verdict on a package before it depends on it.

[Christo Wilken @TweedBeetle]

Under the hood it reconciles four free sources (npm registry, https://t.co/lFbCQR9qnS, OSV, GitHub) into one confidence-scored answer. When they disagree, it flags it instead of guessing.
It's deliberately narrow: the deprecated / yanked / superseded / migration seam.

[Christo Wilken @TweedBeetle]

@loffredojeremy Manifold Market on this

https://t.co/yfGKPxFeZq

[Christo Wilken @TweedBeetle]

I built pod-space: a 3D semantic map of the top 100,000 podcasts. Each dot is a podcast, positioned by what it's about. Search is cross-lingual, so an English query finds the German, Portuguese, and French equivalents. Free, no signup. https://t.co/YojM9gsTGr

[Christo Wilken @TweedBeetle]

Correction on the 63k-per-agent claim. Pulled the raw usage stamps: that's ~31k of parent context (CLAUDE.md + system + tools) loaded twice. Intrinsic Workflow overhead is much smaller.

Also: parallel() siblings don't share prompt cache. N agents cost N× full context.

[Christo Wilken @TweedBeetle]

Anthropic might have just shipped the abstraction level that will replace vibe coding, but you have to enable it first.

CC v2.1.147 added a Workflow tool. Sandboxed JS where you compose subagents with agent(), parallel(), pipeline(), and a real token budget. Off by default. 🧵

[Christo Wilken @TweedBeetle]

Smoke test: I wrote a one-agent workflow that calls agent('Say PARROT'). It ran in 2.9 seconds, returned {result: 'PARROT'}, and cost 63,000 tokens for the single step.

That last number is the real catch. Per-agent overhead is high, so budget your runs accordingly.

[Christo Wilken @TweedBeetle]

If you try this, post your script and what surprised you. Nobody on X has shared a working example yet.

[Christo Wilken @TweedBeetle]

Caveats:
- Same-session resume only
- 1000-agent hard cap per workflow
- Token budget shared with the parent turn
- Docs are nonexistent. Most of the above came from grepping the binary
- Anthropic might flip the flag widely soon

[Christo Wilken @TweedBeetle]

@emilheap @ClaudeDevs I’m pretty sure -p always got billed via api pricing and never used your plan quota. Am I wrong?

[Christo Wilken @TweedBeetle]

My day with claude code

[Christo Wilken @TweedBeetle]

My day in claude code