Added#
trazum-vscode, the editor extension. Live token cost while writing a prompt: the count, the budget that covers the file and the pattern it came from, and what the deterministic rules would recover, in the status bar and its hover. The plan called this unblocked since 0.10.0 and unscheduled ever since, because an extension is a distribution commitment rather than a feature. This arc takes the commitment.It sends the buffer nowhere, in any form, ever.
@trazum/coreruns in the editor's own process against text already on the machine, and the promise is a test rather than a paragraph: every source file in the package is scanned for a way out, the permitted set is empty, and importing either core module that exists to make calls fails the build.Every judgement lives in
reading.ts, which has never heard of an editor. A VS Code extension is normally tested by downloading VS Code and driving it — a network dependency, a version to track, and a suite that cannot run without a display. Here the extension is a wire: it hands a string and a config to a pure function and renders what comes back, and a guard asserts it performs no arithmetic and formats no figure of its own.It depends on
@types/vscodefor nothing. The editor supplies thevscodemodule at runtime and never installs it, so the contract this extension relies on is written out insrc/vscode.d.ts, exactly as wide as what it touches. 1.85.0 already paid for the alternative: an install-only dependency that turned an optional package into one the repository could not compile without.nullis not zero, in the one place a reader would be misled. The rules run when the typing stops, not on the keystroke, so a fresh reading has not measured them. It says so rather than showing0, which would tell somebody their prompt is already tight when nobody has checked.The first version of this workspace could not be packaged at all, and nothing caught it because nothing had tried to ship it. It was named
@trazum/vscode— a fine npm workspace name and an impossible extension name, since a marketplace identifier ispublisher.nameand a slash cannot appear in one. It istrazum-vscodenow, and the rules a packager enforces are checked by a test rather than by installing the packager:@vscode/vsceis fetched bynpxfor the one run that writes a.vsix.The icon is generated by
npm run draw:iconand asserted byte-identical to a fresh run, the same claim the architecture picture carries. It draws the proportion bar this product prints on every surface, in the palette the rest of it uses, and the PNG is written by hand — a zlib stream and four chunks — rather than by taking an image dependency.npm run package:vscodewas run rather than described. It producestrazum-vscode-1.86.0.vsix: 8 files, 8.39 kB — the two built modules, the icon, the licence, the README and the two manifestsvscewrites. No sources, no tests, no source maps, which is.vscodeignoredoing its job and is now the thing that observation confirmed rather than the thing a config file claims.That run also found what it leaves behind. The guard against runtime state in this repository checks the tracked tree, deliberately, because an untracked file is a local mess rather than everybody's — but
git add -Abefore a commit sweeps up whatever the last command produced, which is exactly how sixty waiver records reachedmainand sat there for two releases.*.vsixis ignored, and a test asserts it, so running the documented command cannot cost somebody that.
Fixed#
The release document told the owner to configure three package pages out of four. The trusted-publisher instruction listed
@trazum/core,@trazum/cliand@trazum/mcp— a list typed when there were three packages and left alone when@trazum/tokenizer-openaimade it four. Somebody following it exactly would have configured three, and the release publishes the new package last, so the page that fails is the one the instruction never named.A derived guard holds it now: every package the workspace globs say this repository publishes must be named in the paragraph that hands over the settings table. The first version of that guard read the whole section and a plant walked past it, because the paragraph above the instruction names the newest package for a different reason.
Two claims in that document that the 1.85.0 release disproved. It carried the heading "still not working" and the standing instruction "assume tags will not publish and release by hand". The 1.85.0 merge published all four packages through the workflow, with provenance, and created the tag and the release with no human step. What the run does not settle is which credential authenticated the upload, because provenance is signed with the job's OIDC identity either way — so the document now says that too, and names the step that answers it rather than guessing.
A token scope that goes stale the moment a package is added. The fallback token is granular and scoped to named packages, so a package added later is not in it, and the failure is the same
E404as a missing trusted publisher. Written down beside the instruction that creates the token.The extension's wire shipped with no behavioural test at all. The guard on it asserted textually that it computes nothing and writes no status text of its own — an assertion about what it does not do. Nothing checked that it does the right thing, and a wire has plenty of room to be wrong: the workspace-relative path it builds for budget matching, the four separate branches that hide the item, the debounce that keeps
optimizeoff the keystroke path, and the listener that must ignore a document which is not on screen.shim.test.jsrunsactivate()against a fake editor, resolved through a loader the wayapps/webresolvesnext/server. Everything under the fake is real: the core, the reading module, and a config parsed off a real file on disk. Eight violations were planted and every one fired on the test that claims it. The workspace-relative path is proved through the budget rather than by watching the call — the config scopesprompts/*.txt, which can match the relative path and cannot match the absolute one, so a shim that passed the wrong one shows a bare count and the budget silently stops applying in the editor whiletrazum checkstill enforces it.That guard's own first version was bound to its neighbour: it matched the first binding called
path, which is the config path insideprojectConfig, and reported the shim as broken. Rebound to the binding that is handed toread.The fake editor was the only editor this repository will ever run against, and nothing held it to the contract.
src/vscode.d.tsis hand-written, the fake is plain JavaScript reached through a loader, andtscnever sees the two together — so a fake that lost a member, took fewer arguments than the editor passes, or grew one VS Code does not have would leave the whole shim suite green about an editor nobody ships.contract.test.jsholds it, and holds the declaration to being no wider than what the shim touches, which is the install-only dependency growing back one member at a time.Its arity check was too loose at first and a plant proved it: a fake that dropped the optional priority argument passed, and only failed by accident. It requires the declared count exactly now, because a fake cannot observe an argument it does not take.
Four comments named a guard that does not exist. Two source files pointed at
shim.test.jsand one atcontract.test.js— files promised in a shipped release and never written.memory.tspointed at a postgres suite under a name it does not have, a rename that took the file and left the sentence. Anddraw-icon.mjspointed at an icon suite that has never existed under any name.A reference to a guard that is not there is worse than no reference: it reports a check where there is none, and the reader stops looking.
named-guards.test.jsscans every source file and document in the repository and fails on a name that is not a file.The icon's palette said it came from the product and one third of it did not.
draw-icon.mjsclaims the colours are "the product's own, taken fromdocs/assets/demo.svg", and the guard that was supposed to hold that claim held a hardcoded copy of one colour instead — two copies of the same number, agreeing with each other and with nothing. Bound to the demo, the ground and the accent were there and the unspent-cell colour was in no other surface of this product. It is#363329now, the rule colour the demo already draws on the same ground, and the icon is regenerated.