docs: correct claims the code does not support

[ribdsp]

Three prose claims that the code does not support. Docs and one comment only — no behaviour change, and registerDynamicTool is left exactly as it is, marker included.

What was wrong, and what it says now

1. CLAUDE.md — "every stub has been implemented"

There are none left — every stub from the original build has been implemented.

False in two ways. traces/src/lib/webmcp/register-tools.ts:133 still reads:

export function registerDynamicTool(_hypothesisId: string): void {
  throw new Error('registerDynamicTool: not implemented')
}

and the TODO: Day 6 marker is still on line 131, directly above it. A repo-wide grep (excluding node_modules) finds exactly three references to the symbol: the two definition lines, and one comment in tool-status-banner.tsx. Nothing calls it.

Now names the stub, says it throws, says nothing calls it, and states the consequence — promoting a hypothesis does not grow a 17th tool. The rest of the section (why markers must not be deleted) is untouched.

2. CLAUDE.md — the test count

All 285 tests pass.

Measured on this branch: 25 files, 303 tests. Updated to 303.

3. tool-status-banner.tsx — the toolchange comment

toolchange is how a surface that grew a tool mid-investigation shows up here without a reload — the promoted-hypothesis tool from registerDynamicTool is the case worth demoing.

Read as documentation of something the app does. It reads as a conditional now, and says plainly that registerDynamicTool is a stub that throws with no callers, so the only thing that can move the counter is the host changing its own tool list. The rest of the comment — the EventTarget note, ChatGPT Desktop, why the count is changes rather than a total — is accurate and unchanged.

Everything else was checked and is correct

Every other absolute claim in CLAUDE.md and README.md was verified against the code rather than assumed:

Claim	Verified against	Result
16 tools, and all 16 names	registry.ts, register-tools.ts	correct
7 predicate shapes	evaluate-predicate.ts	correct
60 lines / 1200 chars budget	compress-dom.ts:16-17	correct
60-char text, 20-char value truncation	compress-dom.ts:19,100	correct
Compression 2,026→663 chars, 17 lines, 3.06× (+2.68×, 2.57×)	re-ran scripts/measure-compression.mjs	correct
3 recordings at 184 / 210 / 213 KiB	public/recordings/	correct
209–223 events	counted per fixture: 222 / 209 / 223	correct
4 blocking tools, 64 MB ceiling, src/app/fonts.ts	source	correct

docs/tools.md:579 was already fixed and is untouched. The tagline is untouched in all four places.

Notes

One drift from the brief. The test-count line is CLAUDE.md:39, not :37 (:37 is the section heading). Fixed by content.

One understatement left in place, deliberately. CLAUDE.md:43 describes no-eval.test.ts as failing "if eval( or new Function appears anywhere". That is true, but the test enforces five patterns across all of traces/src, not two — it also catches setTimeout/setInterval with a string first argument and .innerHTML assignment from anything named like args|input|params|predicate|payload. Understated rather than false, so it is reported here rather than rewritten; correcting it properly means rewriting the paragraph, which is a bigger edit than this PR is for.

Forward dependency. "does not grow a 17th tool" is correct at 16 tools. If a seventeenth tool lands, that word has to move with the rest of the 16→17 sweep.

Verification

npx tsc --noEmit     exit 0
npx vitest run       25 files, 303 tests passed
npm run lint         exit 0 (8 pre-existing warnings, unchanged from #16)

[vercel]

The latest updates on your projects. Learn more about Vercel for GitHub.

Project	Deployment	Actions	Updated (UTC)
traces	Ready	Preview	Aug 31, 2026 6:05am