Headroom
shippedPaste a wall of tool output, logs, or RAG chunks and watch an algorithmic compressor strip redundancy and show exactly how many tokens it saved before the text would ever reach an LLM.
Approach
A single static index.html with inline CSS and vanilla JS: no build step, no dependencies, so it builds in seconds and has nothing that can 404 at 4am. The compressor runs three real, inspectable passes over the pasted text: (1) strip ANSI escape sequences and trailing whitespace/control noise, which is pure token waste with zero information; (2) collapse consecutive lines that are identical or identical-after-normalizing digits/hex/timestamps into one line plus a '(x N)' marker, which is how retry/log spam gets compressed; (3) deduplicate exact-match sentences/paragraphs that recur non-consecutively across the pasted text (the classic overlapping-RAG-chunk problem), keeping the first occurrence and dropping later ones since the same sentence shown once carries the same information to a downstream LLM. A small heuristic tokenizer (punctuation-aware word splitting, not a real tiktoken port) estimates tokens before and after, labeled explicitly as an estimate so the number on screen never overclaims precision. Each pass can be toggled off independently so a visitor can see exactly which technique bought how much, which doubles as proof the savings are real and not fixed/staged.
The source post
https://x.com/benou_irl/status/2112166896086778047
Scoring
Pick
| surprise | 3 |
|---|---|
| demonstrability | 5 |
| self_containedness | 5 |
| honesty | 5 |
Headroom is the only candidate whose one interesting capability is both real and runnable entirely in the browser: a text-compression heuristic that actually shrinks pasted logs/tool-output/RAG chunks and reports a genuine before/after token count, no model call or fetched weights required. That makes it honest (the number on screen is the real result, not a stand-in) and self-contained under this sandbox's no-outbound-network build constraint. It's also the most demonstrable: paste text, click compress, see the ratio, in under 10 seconds. The closest competitor, the node-based 'mission control' workflow UI, is also fully static and buildable, but it scores lower on surprise and honesty since a node graph with no engine behind it risks reading as a mockup of a process rather than a working tool. Everything else in the batch either requires a CLI/IDE integration (ai-memory, OpenCode), a live API or downloaded model weights at build time (VoiceStudio's voice cloning, the image-swipe generator, the scientific-AI simulation stack with FEniCSx/FMI), or is a backend proxy/firewall that can't be meaningfully smoke-tested as a single page (Pipelock).
Review
| shipped | 5 |
|---|---|
| honest | 5 |
| worth_it | 3 |
| efficient | 5 |
No change: the gate passed first try, the artifact is honestly real rather than a mockup, and NOTES.md raises nothing new or recurring — the one process note (using a Node/DOM harness instead of headless Chromium) is already codified in build.md's lessons and was followed correctly, not violated. The only anomaly tonight (transient 522 on the fresh custom domain) is normal Cloudflare DNS propagation, resolved on its own within two minutes, and isn't a file this stage is permitted to edit even if it recurred.
Cost
| total | $0.1136 |
|---|---|
| xai | $0.1136 |
What it looked at
ai-memory**: GitHub tool that lets you pause one coding agent mid-task and resume in another without re-explaining context.
Scientific AI high-assurance workflow demo**: End-to-end system that proposes, simulates, verifies numerical results, detects forgeries/faults, and packages reproducible evidence with optional FEniCSx/FMI runtimes.
OpenCode**: MIT-licensed open-source coding agent (Claude Code style) that works with any model (Anthropic, Ollama, etc.) and is fully inspectable.
Pipelock**: Open-source Apache 2.0 AI agent firewall that inspects HTTP/WebSocket/MCP traffic, blocks exfiltration/injections, and emits signed receipts.
VoiceStudio**: Fully local open-source ElevenLabs alternative for voice cloning, dubbing, and transcription.
Node-based AI workflow interface demo**: Visual “mission control” web of nodes for roles, objectives, rules, and review—shows structured multi-step processes beyond simple chat.
AI image swipe prototype**: Swipe-to-accept/reject style interface that iteratively generates image batches while keeping preferred styles.
Gate
| pass | project directory exists — /Users/artax/code/builds/2026-10-20/project |
|---|---|
| pass | no build step needed — static project |
| pass | build output with index.html — /Users/artax/code/builds/2026-10-20/project |
| pass | index.html is a document — 15958 bytes |
| pass | local asset references resolve |
| pass | page loads without console errors |
Stages
| scout | grok · ok · 60.4s |
|---|---|
| pick | claude · ok · 51.5s |
| plan | claude · ok · 64.5s |
| build | codex · ok · 185.7s |
| gate | local · ok · 2.8s |
| publish | local · ok · 16.2s |
| review | claude · ok · 120.9s |
Notes
Omitted for lack of
None. All planned functionality runs locally in the browser.