Adopts ideas/token-conservation.md, parked 2026-07-28, plus the disciplines the human added: read the compressed form first, discover by mechanism, shrink output at the source, never pay twice for the same bytes. docs/context-economy.md owns all of it and leads with a floor, because every cheap failure mode is also a token saving — skipping the owning doc, guessing a signature, reporting from a diff, thinning a test — and each has already cost this repo a defect. Frugality is for presentation and discovery, never for the artifact or the evidence. Mechanical, not remembered: - scripts/surface.sh generates docs/surface.md — every top-level declaration with its line, 261 lines standing for 3757 of source. The pre-commit hook regenerates and stages it, so it cannot be stale, and verify.sh compares independently for a clone that never set core.hooksPath. The hook refuses a commit with unstaged .go changes, since what it generated describes the working tree, not the commit. - verify.sh --quiet: 48 lines of gate output become 1. The hook uses it. - CLAUDE_LOC_MAX=150, the only budget billed per turn rather than per read. Both new gates were watched failing before being kept: a doctored surface.md, and CLAUDE_LOC_MAX temporarily set to 5. state.md's inventory loses its LOC column. It had already drifted on six files (content.go 381→450, render.go 447→454, web.go 206→217, check 216→223, watch 129→137, chrome 105→110) which is what a number written in two places does; the generated file owns sizes now, the table owns purpose. The subagent question is recorded there as the one open decision, with the case for and against written out in the idea file.
5.0 KiB
Context economy
How the agent spends context. The limit on this project is not typing speed, it is how much work fits in a session before the window fills — so bytes that buy nothing are features not built.
This doc has a floor and the floor wins. Every technique below is forbidden from buying:
skipping the doc that owns a rule (CLAUDE.md §1), guessing an API instead of reading it, reporting
success from a diff instead of a run (SKILL.md §4), keeping a test that was never seen to fail, or
thinning generated code, content, tests, comments or commit messages. Those are all cheaper in tokens
and every one of them has cost this project a defect. Frugality applies to presentation and
discovery, never to the artifact or the evidence. When a saving and the floor disagree, pay.
1. Fewer turns beats fewer bytes
The whole context is re-sent on every tool call, so a six-call sequence costs roughly six times the
window. Batch independent calls into one message. Where a sequence is known in advance or will be run
again, write the script instead: one artifact that sets up, runs, probes and prints its conclusion —
the live-reload check that measured the watcher is the shape to copy, and verify.sh is the same
technique applied to the gates.
2. Read the compressed form first
In order of cost, stop at the first that answers the question:
| Question | Cheapest answer |
|---|---|
| where does X live, what is in this package | docs/surface.md (generated; make surface) |
| what does this package offer callers | go doc ./internal/<pkg> |
| what does this behave like | its _test.go — a table states in 20 lines what 200 implement |
| why is it like this | decisions.md, state.md — never the code |
| what changed and when | git log --oneline -- <path> |
| how big is this before I open it | wc -l, or the manifest line in surface.md |
Then read the range, not the file: grep -n to locate, Read with offset/limit. A whole-file read
is for the doc that owns a rule you are about to assert, and for a file you are about to rewrite.
3. Let the machine find it
Discovery by mechanism beats discovery by reading.
- Break it and read the errors. Changing a signature:
go build ./...returns the exact caller list, complete, in a form a grep can miss. scripts/verify.sh --listanswers "does this gate exist" without running or reading anything.grep -cfor "does this exist",-lfor "which files",-nfor "where". Context flags only when the surrounding lines are the answer.go test -run TestNameover a full suite while iterating; the full suite once, at the end.
4. Shrink output at the source
Filtering inside the call is free. Filtering after it lands is impossible.
./scripts/verify.sh --quiet— two lines on green instead of forty-three. The pre-commit hook uses it. Use the loud form when a gate fails or you are auditing the gates themselves.- Quiet and fail-fast flags by default:
-q,--porcelain,--short,--oneline,-failfast,--statbefore-p, and2>/dev/nullfor known noise. - Pipe through
tail/grep -v/awkin the same call. Ten relevant lines, not four hundred. - Long or noisy runs write to the scratchpad and get queried; never
cata lock file, a binary, or generated output. - Sample rather than enumerate: two bundles out of thirty answer whether the shape holds.
5. Never pay for the same bytes twice
- No re-reading after
EditorWrite— they fail loudly if they did not apply. - A background command's output is read once, after it completes. Not while polling, and not quoted back afterwards.
- No restating the plan, then doing it, then summarising it. State once, report deltas.
- Write a fact down where it belongs the first time it is established (
state.md, an ADR, a resume note in the queue file) so the next session reads a line instead of re-deriving it. A handoff note is cheaper than a compaction summary and loses less. - Report
file:line, never the code — the human has the file, and the reference is clickable.
6. Prose is the cheapest thing to cut and the easiest to overcut
Target density, not brevity: the same findings, numbers and caveats in fewer words. Cut preamble, recap, restatement, option surveys, and hedging. Do not cut a finding, a measurement, a stated assumption, or a caveat — a report that omits the caveat is not shorter, it is wrong.
What is enforced, and what is not
Three of these are mechanical. scripts/hooks/pre-commit regenerates docs/surface.md and stages it,
so no commit can carry a stale one, and verify.sh compares independently for the clone that never set
core.hooksPath. CLAUDE_LOC_MAX bounds the one file billed on every turn. --quiet exists to be
used, and the hook uses it.
Everything else here is discipline — no script can see a redundant read, an unnecessary whole-file Read, or a report that padded instead of informing. That asymmetry is the reason this doc is short enough to be re-read, and the reason its floor is stated first.