
Merging a whitepaper is not a vote to build it. A Nano Collective whitepaper merges when the argument is good enough to publish. The build, iterate, or decline call is a separate gate at Stage 3, and the rationale is written on the page either way.



When a Nano Collective whitepaper's proposer goes quiet, the paper doesn't stall. Any other contributor can add themselves as the new proposer and keep the review going. After 60 days with no one stepping up, it moves to Paused, never Deleted.



The fastest way to lose trust with an open-source fund is announcing capacity you don't have. The Nano Collective's community fund is a mechanism today, not a balance. The page says so, because announcing from an empty pool is the failure mode an OSS fund can't afford.



Before a Nano Collective donor gives money, they agree what the money is for. The scope, the deliverable, the report. Nothing open-ended, nothing retrospective. The principle: money follows agreed work, not the other way round.



A batch tool that buffers everything before exiting is a batch tool you can't pipe. get-md --json writes one record per line, so jq, dashboards, and downstream jobs see URLs the moment they complete - not after the last one finishes.



"Convert everything" is the wrong default for a sitemap crawl. get-md stops at 10,000 URLs and maxDepth 3 by design. Your run ends when you want it to.



Most "browser HTML to Markdown" libs hit a wall the moment you ship to a phone. `get-md` swaps JSDOM for `happy-dom-without-node`, so the same `convertToMarkdown` call runs in a React Native bundle. HTML in, clean Markdown out, on-device - no Node service in the loop.



Iterating on a conversion pipeline means the network is your bottleneck. get-md --cache with its default one-hour TTL sends re-runs to ~/.get-md/cache. Re-runs hit disk, not the network.



You don't have to pick a builder type before you know which migrations need I/O. `createAsyncMigrations` in `json-up` takes sync and async `up()` functions in the same chain, so only the steps that actually await something pay the cost.



A bad migration that ships today becomes the incident you debug weeks later. `json-up` names the failure: ValidationError (version + Zod path), MigrationError (up() throw with cause), VersionError (out-of-order or duplicate configs). You read it, you don't guess.



1 schema per migration step. json-up validates each up() output against its own Zod schema before the next migration runs, and ValidationError names the version that broke. A silently corrupted field cannot pass through the chain undetected.



Retrofitting a migrations library onto data you didn't design usually starts with `rename version to _version`. json-up's migrate options take a `key` argument so the builder tracks the field your data already uses.



Commands, subagents, tools, and cron/file triggers used to live in four places and drift apart. Nanocoder skills fold them into one bundle. A per-project daemon actually fires the triggers.


The editor isn't displaying your agent. It IS the agent. nanocoder --acp streams JSON-RPC over stdio, so editors like Zed render text, tool cards, and diffs natively. Normal, auto-accept, yolo, and plan become modes you pick from inside the editor.



Today I learned Nanocoder's auto-compact silently swaps to mechanical truncation when the LLM summariser errors, returns nothing, or hands back a summary longer than the original. One invocation, no prompt, no flag, just keeps working.



An AI refactor turns ugly at step 30. No undo. nanocoder's /checkpoint create snapshots files, git diff, conversation, and provider config, so /checkpoint load puts it all back. Run two approaches, compare by listing them.



The system prompt drops from 500-700 tokens to 150-250 on small models. nanocoder's /tune nano profile strips CORE PRINCIPLES, CODING PRACTICES, AGENTS.md, and three tools so your code gets the context window, not the prompt. Auto picks nano under 4B params, no flag needed.



You catch a model failure mid-chat, fix the system prompt, then watch the same mistake come back next session. `/keep` appends the last exchange to `train.jsonl`, so the bug you just saw becomes the example that prevents it.



7 providers, 1 config shape. nanotune judge configure treats OpenRouter, OpenAI, Anthropic, Gemini, Ollama, llama.cpp, and LM Studio the same. API keys resolve from the shell via ${ENV_VAR:-default}, so swapping a cloud judge for a local one is a base URL and a model id.


Two seeds, two jobs. `--seed` only runs the split when `valid.jsonl` is missing, so it's a silent no-op after that. `--train-seed` runs every time. Delete the validation file to re-roll the split; leave it alone to keep training reproducible.



A benchmark score that moves could be the model, or it could be the sampler. nanotune benchmark defaults to temperature 0 and seed 42, so two runs against the same model produce the same score. A moved pass rate is a real change. --samples N is the opt-in for variance.



Five flags to tune a benchmark, and you only need one. Nanotune's --preset low|medium|high|ultra sets threads, GPU layers, context, batch, and max tokens together, with the realistic hardware spread covered. Need a knob? An individual flag still overrides the preset.


8 detectors, 1 priority chain. Secret wins, then Email, Url, Path, Phone, Address, Name, CodeTell. When a URL carries a credential shape, the token wins and the URL rewrites around it. Missing a credential beats missing a name.



A privacy tool that picks your proxy for you has too many opinions. prompt-scrub's threat model draws the line at the content layer: IP, timing, headers are not its job. Scrub locally, route with whatever you trust.


Most plugin systems ship with a new DSL to learn. prompt-scrub rule packs use the runtime contract directly - export a `Detector[]`, declare it under `prompt-scrub.rulePacks`, the CLI auto-loads. The format is the interface, not a config schema.



Defaulting every detector to on sounds thorough, until names in code start getting stripped as PII. prompt-scrub ships NameDetector and CodeTellDetector off by default, with strict mode for names and a terms list for code as the safe opt-in.



Tuning a rule pack by guessing is how you bury your own issues later. Sentinel's dry-run prints a grouped preview, would-file, would-dedup, would-skip, so you tune the pack and the severity floor without filing anything you didn't mean to.



Most dedup keys include line range, so a model that drifts a few lines between runs hashes a different finding and refiles the issue. Sentinel uses rule + file + category instead: stable across runs, so a triaged issue stays triaged.


An audit tool that ships its own rule packs is telling you what *it* thinks your code should look like.

Sentinel ships zero. The format is the product. You write the pack for the code you ship, and the body of the pack is the prompt handed to the model.



Audit work is exactly where reaching for a cloud model exposes you. Sentinel alpha runs the model through Nanocoder on your own runner, so the audited code never leaves hardware you own. Cloud is opt-in config, not the default.
