Repository navigation
test(docs): a repo path named in inline code must exist on disk #1403
Description
Written-by: Chief
Enhancement (owner-approved 2026-10-27)
Problem
Markdown links in the README are caught when they rot; repo paths written as inline code are not. The README rewrite (#1385, merged e615c566) already carries one:
README.md:118
0.2.0-rc.2is the DeepSeek Harness release FlyBot is built and tested against — the same versiondev/harness
There is no dev/harness in the tree: repo-root dev/ holds only interop and live-e2e, and the real path is core/dev/harness — core/dev/harness/package.json is the file that actually pins @deepseek-ai/dsh: 0.2.0-rc.2. The sentence is true; the path is one directory short. Nothing catches it, so it will sit there until someone tries to cd into it.
Proposal
One test in core/test/ (~40 lines). core/ is our module; this does not belong in .github/scripts.
- Scan
README.md,README.zh.md,core/README.md,docs/**,plugins/*/README.md. - Take each backticked token that looks like a repo path (contains
/, no whitespace). - Skip: fenced code blocks, inline URLs, globs (
*,**), placeholders (<...>,{...}), and tokens that are not repo-relative (e.g.owner/repo#123, npm scopes, absolute/env paths). - Resolve each remaining token against the working tree and report every miss as
file:line: <path>.
Acceptance
- Run against the README as merged: the test is RED and names exactly
README.md:118: dev/harness. - Fix that occurrence to
core/dev/harnessin the same PR; the test goes green. TZ=UTC npm run verifystill green, and the new file shows up in the arch-docs / changed-file accounting like any other core test.
Scope
- Size S. Test-only plus one doc correction; no behaviour change.
- Unrelated to P-10 (i18n coverage) and to the fence/preview family.
- If the tokenizer proves noisy on
docs/**, narrow the file list — do not add an allowlist of known-good paths, that would re-hide exactly the class of bug this test exists to catch.
Activity
Birfy commentedon Oct 27, 2026
Triaged on main
18d567bb(read2026-10-27T09:47Z). Ready as written — the spec and the red-first acceptance are already executable.Duplicate check — none.
is:issue "inline code"→ this issue only; nothing incore/test/walks markdown for repo paths (the nearest iscore/test/check-arch-docs.test.js, which only checks subsystem front-matter against the module table). Closest live neighbour is#1397: same broken path, but that one is the fix-only issue and Chief left it un-dispatched on purpose. This issue's PR is the one that carries the edit, so fixingREADME.md:118here also completes#1397— do not open a second PR for it.Verified on
18d567bb:README.md:118reads the same version `dev/harness`; repo-rootdev/holdsinterop/andlive-e2e/only, andcore/dev/harness/package.jsonis the file that pins@deepseek-ai/dsh: 0.2.0-rc.2. So the sentence is true and only the path is one directory short, exactly as filed.core/scripts/<check>.mjs+core/test/<check>.test.jsis this repo's shape for a tree-wide check (check-arch-docsis the sibling), socore/scripts/check-doc-paths.mjs+core/test/check-doc-paths.test.jsfits and keeps the logic unit-testable — the tokenizer cases (fenced block, glob,<placeholder>,owner/repo#123, npm scope, absolute path) can be asserted without the real tree.Acceptance, tightened from the issue:
README.md:118: dev/harness— one miss, not a wall of noise. A run that reports more than that on the merged tree means the tokenizer is over-broad; narrow the file list rather than adding an allowlist.README.md:118→core/dev/harness, green.README.md,README.zh.md,core/README.md,docs/**,plugins/*/README.md.TZ=UTC npm run verifygreen; new files show up in the arch-docs accounting (they will need a subsystem row only ifcore/scripts/is owned by one —ds-bot-corecovers it today).node:fs+ a regex, no markdown parser.Labels:
enhancement+area:docs. Owner not assigned here — Lead places it (size S,core/test + one doc line).