A plan comes in. It leaves as finished work, each piece routed to the cheapest model that can do it.
Hand a cheap model a task and it re-reads everything, or confabulates the one ID that actually mattered.
So I pack it instead. A Context Pack is one file: exact needles (ids, paths, values) kept verbatim, the bulk gisted, sized to half the model's window. Workers read what matters and re-fetch the rest.
context-pack.md
Decode headerhow to read this
Taskspec + done
Needleskept exact
7f3a9c2e1b8d/etc/app/config.yamlNEVER push to remote
Gist bulkrecognition only
File indexre-fetch exact bytes
/lib/context_pack.mjs/workflows/swarm.mjs
One file. The needles stay exact. Everything else is gist you can re-fetch.
needle extractor
what the pack keeps
Deploy commit 7f3a9c2e1b8d to the box.
Config lives at /etc/app/config.yaml, port 8080.
NEVER push to the remote before the gate passes.
The rest is background: we tried a few approaches
last week and settled on the simple one, which reads
fine and does not need to be exact to be useful.
4 needles kept verbatim. The rest greys to gist you re-fetch.
Type anything. The exact needles stay sharp, the bulk fades to gist.
A manager routes each task to the cheapest Claude tier that can do it, on your subscription, never metered. Small jobs go direct. Big ones fan out under a cheap large-context model that reassembles the result.
router
Shaikulow
Msonnetmedium
Lfablehigh
XLfablehigh · fan-out
on retry, escalatehaikusonnetopus
Complexity picks the model. No point burning the big one when the small one does the job.
I shipped this repo by pointing the tool at itself. It said "12 ok". I believed it only after diffing every file, because a workflow once told me "ok" for work it never ran.