Increase YYA's capacity for executive functioning
Eight programs so far, across five parts of one job: what is in context, what the team knows, what it is doing, how it gets run, and whether any of it worked.
Context
What goes into the window, and what it costs.
tc
shipped
Like
wc, but for tokens.
Counts tokens in files or stdin with the o200k_base encoding. Plain tc always means o200k_base — there is no configuration to get wrong.
tokdump
shipped
Like hexdump, but for tokens.
Dumps token IDs in the three-column layout hexdump(1) has used since 1978, so you can see exactly where a string splits — and why an emoji costs you four tokens.
Knowledge
What the team and its models read.
metadata-enforcer
shipped
Django-style schemas for Markdown frontmatter.
Declare the shape of a note collection in plain Python, then enforce it. No intermediate YAML DSL, no JSON Schema.
cq
early
The codebase smells a staff engineer would notice.
Mega-files, import cycles, pages that reach into each other, naming that drifts halfway through. Not strong opinions — OSHA best-practices. Run it as a suite and watch the trend.
Tasks
What is being done, and by whom.
Execution
The infrastructure to actually run it.
Evaluation
Whether any of it worked.
invariant
beta
Looker and TensorBoard, on your Postgres.
A stats dashboard for the numbers you actually steer by. No timeseries database to run, and no second copy of your data.
devtee
early
Turns how you work into evals.
Opt-in recording of what actually happens on a machine, used to generate evals from your team's real computer usage.
devtee
early
Turns how you work into evals.
Opt-in recording of what actually happens on a machine, used to generate evals from your team's real computer usage.