fix(ci): pin Bun to the workspace version and untrack a local node_modules symlink

[dmitrii-f-t27]

What and why

Every Tests / * job on #517 and #518 is cancelled at the 20-minute timeout-minutes while it is still running Install dependencies (bun ci). No test ever runs (#518 server-api log).

Two problems combine to cause this:

1. A local symlink was committed. trios/agent-server/apps/server/node_modules points to /Users/playom/queen-patches/work/browseros-deploy/.... It was last touched in ce13b7e (feat(queen): a lane can be signed with a GitHub login #511). The .gitignore rule node_modules/ matches directories only, so a symlink, which is a file, was not ignored.
2. CI gets the latest Bun instead of the pinned one. oven-sh/setup-bun@v2 runs without a version. The packageManager: bun@1.3.6 pin is in trios/agent-server/package.json, not at the repo root, so CI installs the latest release, 1.4.2, which hangs on that dangling symlink.

Once installs worked, the suites ran for the first time since 2026-09-23 and surfaced latent test problems. Commits 2–6 fix each one, so the whole matrix is green.

Change

1. b3f92ee — CI install
   • Untrack the symlink (git rm --cached). Nobody's checkout loses anything.
   • Change .gitignore from node_modules/ to node_modules, so a symlink can't be committed again.
   • In test.yml, set setup-bun to bun-version-file: trios/agent-server/package.json, which resolves to 1.3.6.
2. 583d8ea — get_page_content
   • The test read the live https://example.com 57 ms after opening it and found no text.
   • It now writes the text into about:blank, the same way the neighbouring get_page_links test does.
3. 26938a6 — killProcessOnPort
   • It used lsof -ti :<port>, which also lists clients connected to the port. One of those clients was the bun test process itself, so it sent SIGTERM to its own run (exit 143, no junit report).
   • It now kills only listeners (-sTCP:LISTEN) and never process.pid.
4. 3d57649 — wait_for finds text on page
   • The test waited 10 s for "Example Domain" on the live example.com and never saw it.
   • It now opens a data: page that adds the text 500 ms after load, so it still exercises the waiting.
5. 8321349 — search_dom class-selector test
   • It searched once, straight after new_page. That is the load race dom.test.ts already documents.
   • It now uses the file's own searchUntil helper, like its two sibling tests.
6. 350d559 — salvage rename test
   • This test runs real git over 205 files. It hit bun's 5 s default once on a loaded runner, while passing on the runs before and after.
   • It now has an explicit 30 s budget.

Verification

bun ci was run on a git archive export of trios/agent-server:

Tree	Bun	Result
base 6bf7efe (symlink tracked)	1.4.2	hangs, killed by timeout 120
base 6bf7efe	1.3.14	installs
b3f92ee	1.3.6 (the pin)	4654 packages, 12.63 s
b3f92ee	1.4.2	2325 packages, 2.23 s

CI: #518 carries all six commits. On its head bf15d74 all 14 suites pass, 2436/2496 with the rest skipped:

• server-tools 244/244
• server-api 1315
• server-pglive 3/3
• server-integration 10/11

Local runs used the BrowserOS AppImage, headless, with --no-sandbox --disable-dev-shm-usage as in CI:

• get_page_content fails before commit 2 and passes 3/3 after it.
• input + navigation-newtab-guard exit 143 before commit 3 and pass 18/18 after it.
• wait_for passes 2/2, repeated.
• search_dom passes 13/13, three runs in a row.
• queen-salvage-guards passes 13/13.

biome check is clean on every changed file.

The workflows audit, code-quality, trios-logic and release-agent-extension also call setup-bun without a version, and this PR leaves them unchanged. Removing the symlink is what unblocks them; pinning them is a separate decision.

Unblocks #518 and any other PR into feat/queen-supervisor. All six commits are also carried in #518, so they no-op there once this merges.

🤖 Generated with Claude Code

https://claude.ai/code/session_01SJ8KjRoGNBoHBoDR92fAo2

[dmitrii-f-t27]

cla is red for a reason unrelated to this PR. The CLA action stops before it checks any signature: Please add a personal access token … for writing signatures in a remote repository, because PERSONAL_ACCESS_TOKEN is empty in this repository (log). It failed the same way on #517, which was merged, and on #518.

Nothing in this PR can fix it: the token secret has to be set, or the CLA workflow disabled for this fork. A re-run would fail the same way.

The change this PR targets is working: on this head the Tests / * suites install and run instead of hanging in bun ci. shared, server-agent, server-lib, server-skills, server-pglive and eval finished green within about a minute.

---

Generated by Claude Code

[github-actions]

✅ Tests passed — 2435/2495

Suite	Passed	Failed	Skipped
✅ agent	87/87	0	0
✅ build	9/9	0	0
✅ cdp-protocol	5/5	0	0
✅ eval	93/93	0	0
✅ server-agent	272/272	0	0
✅ server-api	1305/1364	0	59
✅ server-browser	6/6	0	0
✅ server-integration	10/11	0	1
✅ server-lib	279/279	0	0
✅ server-pglive	12/12	0	0
✅ server-root	68/68	0	0
✅ server-skills	31/31	0	0
✅ server-tools	244/244	0	0
✅ shared	14/14	0	0

View workflow run

[dmitrii-f-t27]

One more base breakage fixed here (aa74689). #522 mounted /queen/contributor-keys but didn't add it to trios/tools/route-guard-audit.mjs. As a result, feat/queen-supervisor now fails four route-guard.test.ts cases: 46 mounts against a pin of 45, 23 /queen mounts against 22, and an unguarded mount with no allowlist entry. I reproduced the failures on the base head (e36fab3).

The route has its own guard. Its bearer must equal QUEEN_CONTRIBUTOR_PROXY_TOKEN (at least 32 bytes, compared with timingSafeEqual), the request needs a verified x-queen-contributor-id, and the route is off while the token is unset. Its only caller is the render proxy, which sends no browser Origin, so the trusted-origin check would refuse it. I allowlisted it with that reason and re-measured the pins. No other count changed.

After merging the base, test:api passes locally: 1305 tests, 0 failed. The same fix is ported into #518 and #521.

---

Generated by Claude Code