fix(office): download-path fixes, drop the vendor name from copy, kill a suite flake

[quickbeard]

Three independent things, one commit each.

1. codevhub skill office download path

opts.size was unreachable (src/lib/download.ts). Number(res.headers.get("content-length")) is Number(null) === 0 when the header is absent, and 0 is finite — so the documented fallback could never run and a chunked response reported total = offset (0 on a fresh download), rendering as Infinity%. The header is now read first and only coerced when present. The progress printer also treats a 0 total like an unknown one.

"Always refetch the setup script" was defeated by a leftover .partial (src/lib/office.ts). Deleting the finished file still left downloadFile resuming from <script>.partial via Range, so an interrupted run could splice stale bytes onto a script that has since been republished — and no expected checksum is passed that could catch it.

README gained a section for the command (documented only in help.ts until now): per-OS bundle sizes, the resumable ~/.codev-hub/office cache, the --download-only / --platform / --dir staging workflow, and the pass-through installer flags.

Tests join the existing runSkillOffice coverage in tests/lib/download.test.ts, against the local HTTP server it already stands up — a new omitContentLength server flag for the two content-length cases, the stale-.partial case, the PowerShell argv shape (unreachable off Windows, so the host is stubbed), and the unsupported-OS / unknown-flag exits. Each was verified to fail with its fix reverted.

2. Vendor name out of user-facing copy

skill office named the bundled skills by their upstream package names in the help text and README; model-limits.ts put the vendor name in the shipped bundle as a plain table key. Help and README now describe the bundle by what it does, and the table key is atob-encoded the way FALLBACK_MODEL already is in const.ts — as its own constant, since coupling it to FALLBACK_MODEL would let a retargeted fallback silently move the 1M window onto another model. grep -i minimax dist/index.js is now empty. Model ids in tests and on the wire are untouched; those are gateway contracts.

3. The SkillPushApp flake

Flagged in the first version of this PR as pre-existing and out of scope — it's fixed here.

It failed ~1 full-suite run in 8 and never in isolation. Instrumenting the test to dump Ink's frame history showed the refusal message does render, and that the app exits ~20ms later — at which point Ink writes an empty final frame. lastFrame() returns "" from then on, so a 20ms poll that misses that 20ms window can never succeed; it burns its budget and reports what looks like a hang.

Fixed with lastNonEmptyFrame, which is stable once the app is gone. Verified by delaying the first poll past the exit: lastFrame() then fails 100%, the settled frame passes. The whole frame history is deliberately not the unit — a prompt renders for one frame before the effect replaces it, so SkillPullApp's not.toContain("❯ ") is an assertion about the settled frame.

The same commit adds renderWithoutRawMode, which mounts an inert tree first so isTTY is already false when the component under test mounts. Four tests mounted the real component and flipped the flag underneath it — a state no terminal reaches, and one ink's handleSetRawMode throws from when useInput's cleanup runs after the flip. That is a fidelity fix rather than the flake fix, and it de-duplicates the same snippet across four files.

14 consecutive full-suite runs clean, against the ~1-in-8 baseline measured on main.

Validation

pnpm fix, pnpm typecheck, pnpm build + node dist/index.js --version clean; pnpm test 1367 passed / 2 skipped across 76 files.

🤖 Generated with Claude Code