[sonichi]

What

Renames the skill skills/claude-gemini to skills/agy. The skill drives the Antigravity CLI agy; the old name described the standalone gemini CLI that Google folded into Antigravity. Scripts, prompts and behaviour are unchanged: skills/agy/scripts/agy-browser.sh is byte-identical to origin/main's skills/claude-gemini/scripts/agy-browser.sh (so it carries #4965's handshake-dir fix, 0c03031b0, an ancestor of this head); gemini-run.sh differs only in one comment (it claimed the manifest was 1.0.0) and a stripped trailing blank line (git diff --check).

• skills/agy/SKILL.md: frontmatter name: agy, title "agy (Antigravity CLI)", usage /agy [prompt]. Built from docs(claude-gemini): read the agy installer before running it; note the CDP port has no auth #4966's SKILL.md at its current head 63d0407a6 (rebased onto main), so its edits are carried as written: the Setup step runs in a subshell with trap 'rm -f "$f"' EXIT for cleanup and trap 'exit 130' INT; trap 'exit 143' TERM; trap 'exit 129' HUP so a cancel ends the block with the signal's status and the EXIT trap removes the file (bash+zsh probe below; $$ is the parent in a subshell and $BASHPID is bash-only, so an exit status is the portable form), downloads the installer with a curl-gated if, shows it, runs it only after printf 'Run the installer? [y/N] '; read -r a answers y (lines 28-36); and the unauthenticated-CDP note (line 66). diff <(git show 63d0407a6:skills/claude-gemini/SKILL.md) skills/agy/SKILL.md shows exactly the three rename lines plus the EOF blank (pasted below).
• skills/agy/manifest.json: name: agy, 1.1.0 → 2.0.0 (rename = breaking, per skills/MANIFEST.md).
• skills/claude-router: route-ai.sh resolves the wrapper to skills/agy/… (see upgrade path); manifest.json 1.0.0 → 2.0.0 and its description plus SKILL.md's now name the Antigravity CLI / agy skill, since the required sibling changed.
• Tests: tests/claude-gemini-wrapper.test.py → tests/agy-wrapper.test.py, tests/claude-gemini-agy-browser.test.py → tests/agy-browser.test.py (main's file with fix(agy-browser): remove the handshake dir when no launcher was forked; test the failing-ps branch #4965's cases, three path lines changed; runs 29/29); new tests/agy-rename-upgrade-path.test.py.
• CHANGELOG.md: [Unreleased] → Changed entry; history untouched.
• Nothing to change in docs/catalog.json (no entry for this skill), docs/src-map.md (indexes src/), tests/skills/ (no per-skill dir). grep -rn claude-gemini (excl. CHANGELOG, .git, workspace) now hits only the alias files and the new test.

Compatibility alias — removed next release

skills/claude-gemini/ stays for one release as: a SKILL.md stub (name: claude-gemini, user-invocable: true, no manifest) whose command line is bash "$SKILL_DIR/scripts/gemini-run.sh" -- "$ARGUMENTS", and two 8-line forwarding shims scripts/gemini-run.sh and scripts/agy-browser.sh that exec the agy script, or exit 1 with skills/agy not found … when it is in neither place. So /claude-gemini [prompt] behaves as /agy [prompt], and a saved prompt or cron that calls the previously documented direct path keeps working.

Upgrade path (Kewei's blocker 1, first review). An existing install has links claude-gemini/claude-codex/claude-router/claude-pi and runs refresh-skill.sh --all, which cannot create an agy link. The shims and route-ai.sh resolve their own real directory first (pwd -P: an installed link points into the checkout, where ../agy/scripts/ exists) and fall back to the installed skills dir — computed the router's way, cd <dir>/../.. && pwd (a logical cd, since the kernel would follow the claude-gemini link before ..) — so installed/agy is used when the physical tree beside the link has no agy. tests/agy-rename-upgrade-path.test.py (7 cases; the runner runs every case and names each failure):

1. test_pre_rename_install — the pre-rename layout in a temp skills dir, refresh-skill.sh --all against it, asserts no agy link appeared, then: old direct paths run, the stub's command line runs, route-ai.sh --dry-run --engine gemini names an existing wrapper, and each shim's exec target normalises to the shipped skills/agy/scripts/<name> (pins the shims' pwd -P).
2. test_pre_rename_install_with_spaces_in_every_path — the same with spaces in the checkout and skills-dir paths (router %q output and bash xtrace parsed with shlex.split).
   3-5. test_installed_sibling_fallback_{gemini_run,agy_browser,router} — one shared fixture, Kewei's layout: installed/claude-gemini -> physical/claude-gemini and installed/claude-router -> physical/claude-router, no physical/agy, valid installed/agy (plus installed/claude-codex/claude-pi, which the router requires). Each consumer of the installed-root fallback is driven and must resolve under installed/agy: gemini-run.sh --help and agy-browser.sh status --help by their exec target, the router by its --dry-run --engine gemini wrapper path.
3. test_copied_skill_without_agy_fails_loudly, 7. test_copied_skill_falls_back_to_installed_sibling.

Mutation controls — each fallback pointed at a missing path turns exactly its own case red (transcript below):

mutation (fallback → missing path)	case that turns red	other cases
skills/claude-gemini/scripts/gemini-run.sh fallback → /nonexistent/…	test_installed_sibling_fallback_gemini_run	6 pass
skills/claude-gemini/scripts/agy-browser.sh fallback → /nonexistent/…	test_installed_sibling_fallback_agy_browser	6 pass
skills/claude-router/scripts/route-ai.sh fallback → SKILLS_DIR=/nonexistent	test_installed_sibling_fallback_router	6 pass
shim pwd -P → pwd (run at a17ac4dfd)	test_pre_rename_install	—
router pwd -P → pwd (run at a17ac4dfd)	test_pre_rename_install	—

Manifest double-load. The stub carries no manifest.json. Checked against loadSkillManifestTools() in src/inline-tools.ts: it skips any manifest without "enabled": true and again without tools, then dedupes tools by name. This skill's manifest has neither, so it loads no inline tools today and a second copy would not have double-loaded tools — but it would be a second package to scripts/lint-skill.py --all, the marketplace listing and loadCoreDocumentedSkills(), and MANIFEST.md requires name to match the directory name. One manifest, in skills/agy/.

CI on the previous head bad6e52f1 (run 36882862651) — not caused by this PR

• python standalone tests (4/5), job 110438832178: FAIL: test_concurrency_test_would_catch_a_naive_non_atomic_writer (__main__.OnboardingSeedTests…) at tests/agy-start-cli.test.py:164 — AssertionError: [] is not true : mutation control: the naive shared-staging-name writer should race and fail at least once — if it doesn't, this harness isn't discriminating atomic from non-atomic writers. A pre-existing timing control tracked as flaky in agy-start-cli.test.py: the non-atomic-writer mutation control is flaky on CI #4928 and flaky: agy-start-cli.test.py mutation control (naive_non_atomic_writer) needs a race that is not guaranteed to happen #4977; it touches none of the renamed paths and only landed in shard 4 because the cost-based packing reshuffled ~67 files on the three filename changes. Locally, scrubbed env, 5/5:

$ git rev-parse --short HEAD: bad6e52f1
$ for i in 1..5: env $(cat witness-unset.txt) python3 tests/agy-start-cli.test.py | tail -1
run 1: Ran 26 tests in 3.541s OK 
run 2: Ran 26 tests in 2.874s OK 
run 3: Ran 26 tests in 2.743s OK 
run 4: Ran 26 tests in 2.854s OK 
run 5: Ran 26 tests in 3.503s OK 
passed 5/5 (scrubbed env, macOS, Python 3.14.6)

• tsc + tests (clean install), job 110441357288: its own tests passed; it failed in the step "Require every test job to have succeeded" — check "python-standalone-tests" "failure" → ##[error]python-standalone-tests FAILED (failure). Derivative of the above.

#4966's installer block — signal probe (interactive bash 5.3 and zsh 5.9 on a pty; curl/less/bash stubbed, less blocks; INT = Ctrl-C on the tty, TERM/HUP = kill to the subshell's process group; sentinel injected after the pager) and the y/n/EOF/download matrix

block-cur.sh is the previous head's block (trap … EXIT INT TERM HUP): every combo reaches the sentinel and the prompt (rc=? is the prompt's read swallowing the probe's echo RC). block-fix2.sh is this head's block, byte-for-byte: no sentinel, no prompt, file gone, rc 130/143/129. The y/n/EOF/download-failure rows are identical before and after.

bash: GNU bash, version 5.3.20(1)-release; zsh: zsh 5.9 (arm64-apple-darwin25.0)

=== block-cur.sh ===
  2:  f="$(mktemp)"; trap 'rm -f "$f"' EXIT INT TERM HUP
--- signal probe (interactive shell, pty; signal while less blocks) ---
bash INT  via Ctrl-C on the tty: sentinel_reached=True prompt_after_signal=True installer_ran=False tempfile_exists=False rc=?
bash TERM via kill -TERM -69366 (subshell pgid): sentinel_reached=True prompt_after_signal=True installer_ran=False tempfile_exists=False rc=?
bash HUP  via kill -HUP -72168 (subshell pgid): sentinel_reached=True prompt_after_signal=True installer_ran=False tempfile_exists=False rc=?
zsh  INT  via Ctrl-C on the tty: sentinel_reached=True prompt_after_signal=True installer_ran=False tempfile_exists=False rc=?
zsh  TERM via kill -TERM -77249 (subshell pgid): sentinel_reached=True prompt_after_signal=True installer_ran=False tempfile_exists=False rc=?
zsh  HUP  via kill -HUP -79780 (subshell pgid): sentinel_reached=True prompt_after_signal=True installer_ran=False tempfile_exists=False rc=?
--- y / n / EOF / download-failure matrix (non-interactive, stdin piped) ---
bash y      installer_ran=true  tempfile_exists=false rc=0
bash n      installer_ran=false tempfile_exists=false rc=1
bash EOF    installer_ran=false tempfile_exists=false rc=1
bash dlfail installer_ran=false tempfile_exists=false rc=0 msg=download failed (curl rc=22)
zsh  y      installer_ran=true  tempfile_exists=false rc=0
zsh  n      installer_ran=false tempfile_exists=false rc=1
zsh  EOF    installer_ran=false tempfile_exists=false rc=1
zsh  dlfail installer_ran=false tempfile_exists=false rc=0 msg=download failed (curl rc=22)

=== block-fix2.sh ===
  2:  f="$(mktemp)"; trap 'rm -f "$f"' EXIT
  3:  trap 'exit 130' INT; trap 'exit 143' TERM; trap 'exit 129' HUP   # a cancel ends the block
--- signal probe (interactive shell, pty; signal while less blocks) ---
bash INT  via Ctrl-C on the tty: sentinel_reached=False prompt_after_signal=False installer_ran=False tempfile_exists=False rc=130
bash TERM via kill -TERM -84807 (subshell pgid): sentinel_reached=False prompt_after_signal=False installer_ran=False tempfile_exists=False rc=143
bash HUP  via kill -HUP -87080 (subshell pgid): sentinel_reached=False prompt_after_signal=False installer_ran=False tempfile_exists=False rc=129
zsh  INT  via Ctrl-C on the tty: sentinel_reached=False prompt_after_signal=False installer_ran=False tempfile_exists=False rc=130
zsh  TERM via kill -TERM -91357 (subshell pgid): sentinel_reached=False prompt_after_signal=False installer_ran=False tempfile_exists=False rc=143
zsh  HUP  via kill -HUP -93587 (subshell pgid): sentinel_reached=False prompt_after_signal=False installer_ran=False tempfile_exists=False rc=129
--- y / n / EOF / download-failure matrix (non-interactive, stdin piped) ---
bash y      installer_ran=true  tempfile_exists=false rc=0
bash n      installer_ran=false tempfile_exists=false rc=1
bash EOF    installer_ran=false tempfile_exists=false rc=1
bash dlfail installer_ran=false tempfile_exists=false rc=0 msg=download failed (curl rc=22)
zsh  y      installer_ran=true  tempfile_exists=false rc=0
zsh  n      installer_ran=false tempfile_exists=false rc=1
zsh  EOF    installer_ran=false tempfile_exists=false rc=1
zsh  dlfail installer_ran=false tempfile_exists=false rc=0 msg=download failed (curl rc=22)

Evidence

Before — the upgrade path at the first reviewed head d40d150d9 (Kewei's repro: four links into the checkout, no agy, SKILLS_DST=<tmp> refresh-skill.sh --all, then the three calls):

$ git rev-parse --short HEAD: d40d150d9
refresh-skill: skills dir <skills-dst>
  refreshed claude-gemini
  refreshed claude-codex
  refreshed claude-pi
  refreshed claude-router
done — updated skills are live in the running session (no restart).
agy_link_after_refresh=no
$ bash <skills-dst>/claude-router/scripts/route-ai.sh --dry-run --engine gemini -- x
route-ai.sh: missing Gemini wrapper: <skills-dst>/agy/scripts/gemini-run.sh
router_after_refresh_rc=1
$ bash <skills-dst>/claude-gemini/scripts/gemini-run.sh --help | head -1
bash: <skills-dst>/claude-gemini/scripts/gemini-run.sh: No such file or directory
legacy_direct_path_rc=127
$ SKILL_DIR=<skills-dst>/claude-gemini; bash "$SKILL_DIR/../agy/scripts/gemini-run.sh" --help | head -1   # the stub as written at d40d150d9
Usage: gemini-run.sh [options] -- [prompt]
stub_path_rc=0

New test at that head:

$ (at d40d150d9) python3 tests/agy-rename-upgrade-path.test.py 2>&1 | tail -3; echo exit=$?
           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
AssertionError: bash: /var/folders/06/t2flhjld0l34pxbnvd4k74pm0000gn/T/agy-upgrade.egfl2pmu/claude-gemini/scripts/gemini-run.sh: No such file or directory

exit=1

Before — the rename itself, at the original parent 1be758ba5:

$ git rev-parse --short HEAD
1be758ba5

$ ls skills/agy skills/claude-gemini
ls: skills/agy: No such file or directory
skills/claude-gemini:
manifest.json
scripts
SKILL.md

$ ls tests/agy-wrapper.test.py tests/agy-browser.test.py
ls: tests/agy-browser.test.py: No such file or directory
ls: tests/agy-wrapper.test.py: No such file or directory

$ bash skills/claude-router/scripts/route-ai.sh --dry-run --engine gemini -- 'x' | sed "s#$PWD#<repo>#g"
engine: gemini
command: bash <repo>/skills/claude-gemini/scripts/gemini-run.sh --cd <repo> --approval-mode plan -- x

$ python3 tests/claude-gemini-wrapper.test.py
PASS test_gemini_only_backward_compat
PASS test_neither_backend_fails
PASS test_unknown_approval_mode_rejected
PASS test_prompt_required
OK
exit=0

$ python3 tests/claude-gemini-agy-browser.test.py 2>&1 | tail -4
exit=0
PASS 29/29 agy-browser.sh

After — at HEAD a2bb844a5 (ancestry, merge-tree, byte-identity, the earlier "also required" items, the same repro, then every suite; T is assigned and exit= is real):

$ git rev-parse HEAD; git status --short | wc -l
a2bb844a5805e69093be8bab6a815fb9973a4098
0

$ git log -1 --format=%P   # parent = #4966 head
63d0407a61ef9e91472fd096f700edef75556c5c

$ git diff --stat a2bb844a5 HEAD
 tests/agy-rename-upgrade-path.test.py | 56 ++++++++++++++++++++++++++++-------
 1 file changed, 45 insertions(+), 11 deletions(-)

$ python3 tests/agy-rename-upgrade-path.test.py > log; echo exit=$?
PASS test_pre_rename_install
PASS test_pre_rename_install_with_spaces_in_every_path
PASS test_installed_sibling_fallback_gemini_run
PASS test_installed_sibling_fallback_agy_browser
PASS test_installed_sibling_fallback_router
PASS test_copied_skill_without_agy_fails_loudly
PASS test_copied_skill_falls_back_to_installed_sibling
OK
exit=0

$ python3 tests/agy-wrapper.test.py | tail -1; echo exit=$?
PASS lines: 6
OK
exit=0

$ python3 tests/claude-pi-wrapper.test.py | tail -1; echo exit=$?
PASS lines: 5
OK
exit=0

$ python3 tests/agy-browser.test.py | tail -1; echo exit=$?
PASS lines: 1
PASS 29/29 agy-browser.sh
exit=0

$ python3 scripts/lint-skill.py --all --strict | tail -1
lint-skill: 58 manifest(s), 0 error(s), 0 warning(s)
exit=0

$ git diff --check 63d0407a6...HEAD; echo exit=$?
exit=0

$ git diff 63d0407a6...HEAD | bash scripts/review-checks.sh | tail -1
review-checks: PASS (hardcoded-paths + root-artifacts + prose-cap clean)

$ git merge-base --is-ancestor 0c03031b0 HEAD; merge-tree conflicts vs origin/main
0c03031b0 ancestor: yes
conflicts: 0

$ git diff --check 63d0407a6...HEAD; echo exit=$?
exit=0

$ git diff 63d0407a6...HEAD | bash scripts/review-checks.sh | tail -1
review-checks: PASS (hardcoded-paths + root-artifacts + prose-cap clean)

After — the same head from a worktree path containing a space (…/scratchpad/checkout with space):

$ git -C 'checkout with space' rev-parse --short HEAD: a2bb844a5
tests/agy-rename-upgrade-path.test.py: PASS lines 7, OK, exit=0
tests/agy-wrapper.test.py: PASS lines 6, OK, exit=0
tests/claude-pi-wrapper.test.py: PASS lines 5, OK, exit=0

Fallback mutation controls at a2bb844a5's tree:

$ mutation controls on a2bb844a5 + this test change (= a2bb844a5's tree)


--- mutation 1: gemini-run.sh shim fallback -> /nonexistent ---
 skills/claude-gemini/scripts/gemini-run.sh | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)
FAIL test_installed_sibling_fallback_gemini_run: gemini-run.sh: skills/agy not found beside <tmp>/agy-fallback.krz31q0_/physical/claude-gemini/scripts or in the installed skills dir
FAILED (1): test_installed_sibling_fallback_gemini_run
pass=6 exit=1

--- mutation 2: agy-browser.sh shim fallback -> /nonexistent ---
 skills/claude-gemini/scripts/agy-browser.sh | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)
FAIL test_installed_sibling_fallback_agy_browser: agy-browser.sh: skills/agy not found beside <tmp>/agy-fallback.j6a1zyp1/physical/claude-gemini/scripts or in the installed skills dir
FAILED (1): test_installed_sibling_fallback_agy_browser
pass=6 exit=1

--- mutation 3: route-ai.sh fallback -> SKILLS_DIR=/nonexistent ---
 skills/claude-router/scripts/route-ai.sh | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)
FAIL test_installed_sibling_fallback_router: route-ai.sh: missing Codex wrapper: /nonexistent/claude-codex/scripts/codex-run.sh
FAILED (1): test_installed_sibling_fallback_router
pass=6 exit=1

reverted: 0 dirty skill files

Earlier pwd -P mutation controls (at a17ac4dfd; the shims and router are unchanged since):

$ git rev-parse --short HEAD: a2bb844a5

--- mutation A: shim fallback replaced by a definitely-missing path (Kewei's control) ---
 skills/claude-gemini/scripts/gemini-run.sh | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)
PASS test_pre_rename_install
PASS test_pre_rename_install_with_spaces_in_every_path
AssertionError: gemini-run.sh: skills/agy not found beside /private/var/folders/06/t2flhjld0l34pxbnvd4k74pm0000gn/T/agy-fallback.2p2j55dw/physical/claude-gemini
exit=1

--- mutation B: shim pwd -P -> pwd ---
 skills/claude-gemini/scripts/gemini-run.sh | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)
AssertionError
exit=1

--- mutation C: router pwd -P -> pwd ---
 skills/claude-router/scripts/route-ai.sh | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)
AssertionError: route-ai.sh: missing Gemini wrapper: /var/folders/06/t2flhjld0l34pxbnvd4k74pm0000gn/T/agy-upgrade.7a3pbpxc/agy/scripts/gemini-run.sh
exit=1
reverted: 0 dirty files

Reviewer notes

• Look first at skills/claude-gemini/scripts/gemini-run.sh (the shim), skills/claude-router/scripts/route-ai.sh:34-42, and tests/agy-rename-upgrade-path.test.py; the rest is mechanical. git shows the scripts as add + modify rather than rename because the old paths now hold the shims; the identity checks are pasted above.
• Documentation: N/A beyond the CHANGELOG entry — no docs/ file references the skill.
• Live path: N/A — slash-command skill, its scripts and tests only.

Merge order: after #4966 (base retargets to main automatically). Stacked on #4966's head 63d0407a6, which is rebased onto origin/main (1be758ba5), so this head contains every commit on main including #4965; git merge-tree origin/main HEAD is clean. This PR's base is #4966's branch docs/claude-gemini-installer-and-cdp-note; GitHub retargets it to main when #4966 merges.

🤖 Generated with Claude Code

[john-the-dev]

Approving at exact head d40d150d9654b2c6d3696944bd295bae18547082. I checked it out in a detached worktree.

What I verified:

• Rename integrity: skills/agy/scripts/{agy-browser.sh,gemini-run.sh} are pure renames (+0/−0). skills/agy/SKILL.md is the old body under the new name. The manifest is agy at 2.0.0.
• No stale references: git grep claude-gemini outside CHANGELOG.md and the stub dir returns nothing. The router (claude-router/SKILL.md, route-ai.sh) and both tests point at skills/agy/.
• Alias stub: skills/claude-gemini/SKILL.md keeps /claude-gemini invocable. With arguments it runs $SKILL_DIR/../agy/scripts/gemini-run.sh, which resolves because install.sh links both skill dirs side by side.
• Tests at head: tests/agy-browser.test.py passes 29/29 and tests/agy-wrapper.test.py is OK.
• Review checks: scripts/review-checks.sh --diff <gh pr diff> from the PR-head tree gives PASS (hardcoded-paths, root-artifacts, prose-cap).

Non-blocking, merge order: #4966 (25e88cae7, open) edits skills/claude-gemini/SKILL.md, which this PR turns into the stub. Once #4966 lands, this PR has to be rebased so its content moves into skills/agy/SKILL.md. Otherwise the conflict resolution can silently keep the stub and drop #4966's edit. After the rebase, please confirm skills/agy/SKILL.md carries #4966's change. That push makes this approval stale, so I'll re-check then.

At review time CI had 2 checks still pending; this approval doesn't certify them.

[sonichi]

CI at d40d150d9: the one red lane is python standalone tests (3/5) → tests/watch-tasks-stream-handler-terminal-rc.test.py, a watcher-stream suite this PR does not touch (the diff is skills/agy, skills/claude-gemini stub, skills/claude-router, the two renamed tests/agy-* files and CHANGELOG.md); it is in the load-sensitive family tracked as #4862. Failed jobs re-run at 07:15 PDT.

[github-actions]

Coverage Gate

✅ Diff coverage PASSES the 95% bar. Whole-tree (informational): 90%.

Diff Coverage

Diff: origin/docs/claude-gemini-installer-and-cdp-note...HEAD, staged and unstaged changes

No lines with coverage information in this diff.

[sonichi]

@qingyun-wu reminder at Chi's request: CODEOWNERS (* @sonichi @qingyun-wu) makes your approval the one GitHub requires here, since the author is under the sonichi login. This PR has two reviewer approvals at its current head and green CI; please respond when you can — an approval, or a comment saying what you need.

[keweichen]

Requesting changes on d40d150d9654b2c6d3696944bd295bae18547082.

Blockers

1. The compatibility alias does not survive the normal existing-install upgrade path. The PR body says an existing claude-gemini symlink keeps the old slash command working before the user reruns skills/install.sh, but the stub invokes the sibling installed path at $SKILL_DIR/../agy, and refresh-skill.sh --all cannot create a new agy link. The router also now hard-requires that absent sibling at skills/agy/scripts/gemini-run.sh.

   I reproduced a pre-rename install with the four existing symlinks, then ran the documented refresh:

   refresh-skill: skills dir /private/tmp/pr4994-upgrade.TjisqP
     refreshed claude-gemini
     refreshed claude-codex
     refreshed claude-router
     refreshed claude-pi
   done — updated skills are live in the running session (no restart).
   agy_link_after_refresh=no
   router_after_refresh_rc=1
   route-ai.sh: missing Gemini wrapper: /private/tmp/pr4994-upgrade.TjisqP/agy/scripts/gemini-run.sh
   legacy_direct_path_rc=127
   bash: /private/tmp/pr4994-upgrade.TjisqP/claude-gemini/scripts/gemini-run.sh: No such file or directory
   

   Worst case: an existing user pulls and refreshes successfully, but both the promised /claude-gemini alias and the Gemini branch of /claude-router stop working until they know to rerun the installer; any saved prompt/cron using the previously documented direct script path also breaks immediately. Please preserve/migrate the old installed and direct paths for the deprecation release (or make the router/stub safely resolve the shipped source), and add a regression test that starts from the pre-rename symlink layout. The current tests only execute the new source-tree paths, so they cannot detect this.

2. The declared parent is not incorporated yet. PR #4966 head 25e88cae7da3718ae296c93d0b9ec2f19c1312da is not an ancestor of this head, and git merge-tree reports a conflict in skills/claude-gemini/SKILL.md. Consequently the current skills/agy/SKILL.md still has the pipe-to-shell installer and lacks #4966's unauthenticated-CDP warning. After #4966 lands, rebase and explicitly carry both of its safety edits into skills/agy/SKILL.md; then refresh the before/after evidence and rerun the full checks on the new SHA. This head cannot be the final reviewed tree for the declared stack.

Also required in the update

• git diff --check 1be758ba5461b9d365f9e3d74ce83e2cae2964a0...HEAD currently fails with skills/agy/SKILL.md:133: new blank line at EOF.
• skills/claude-router/manifest.json remains stable 1.0.0 and still advertises the Gemini CLI even though this change replaces its required sibling with agy. Please update the package metadata consistently with the breaking dependency change. The copied wrapper comment at skills/agy/scripts/gemini-run.sh:112-113 likewise still says the manifest is 1.0.0, while agy/manifest.json is 2.0.0.
• Fix the PR evidence when refreshing it: the after wrapper block currently has a blank exit=, and the install proof runs ls -l $T/skills without ever assigning T.

Verified positives on this exact head: both moved scripts are byte-identical and remain executable; agy-wrapper passed 6/6, agy-browser passed 29/29, the router suite passed 5/5, strict lint passed all 58 manifests, review-checks.sh passed, and the current CI rerun plus CLA are green. Those do not exercise the existing-install migration above.

[john-the-dev]

Re-reviewed at exact head d4b41ea647c244e165a0295074c2d0e2f19f39fa from a detached worktree. Everything the update claims to fix checks out, but restacking onto #4966 dropped another merged fix. This replaces my d40d150d9 approval.

Blocker: #4965's agy-browser fix (0c03031b0) is no longer in this PR.

• git merge-base --is-ancestor 0c03031b0 d40d150d9 succeeds, and the same check against d4b41ea64 fails. #4966's branch forks from main before #4965, so restacking on it lost that commit. That's why compare d40d150d9...d4b41ea64 shows behind_by: 4.
• Line by line, four of #4965's five non-blank added lines in agy-browser.sh are missing from skills/agy/scripts/agy-browser.sh at head: the two comment lines, set +u; launcher="${launcher:-$!}"; set -u, and [[ -n "$launcher" ]] || { rm -rf "$HS"; return 0; }. The fifth (local st pg rc) only matches an unchanged line elsewhere, so the block is gone.
• tests/agy-browser.test.py also lost #4965's 63 test lines. It runs 27/27 at head versus 29/29 at d40d150d9.
• git merge-tree origin/main d4b41ea64 conflicts on skills/claude-gemini/scripts/agy-browser.sh. Resolving it by keeping the shim would ship skills/agy/scripts/agy-browser.sh without the fix, silently.

Fix: rebase the stack onto current main (#4966 first, or this PR after #4966 lands). Then confirm #4965's block is in skills/agy/scripts/agy-browser.sh and the test file is back to 29.

Verified OK at head:

• Installer snippet. It is byte-identical to #4966 219d4694b: the pager if is at line 29, printf …; read -r a at line 30, and the CDP note at line 63. Neither skill dir contains read -r -p anywhere.
• Path resolution. The shims and route-ai.sh resolve paths with cd … && pwd -P and quote every path.
  • From a skills dir under sp ace/App Support, both shims and the router resolved and ran.
  • The arguments "a b" "" '$c' -- d reached the target unchanged, via exec bash "$target" "$@".
• Upgrade test. agy-rename-upgrade-path passes 4/4 at head. Copied to d40d150d9, it fails at its first assertion (claude-gemini/scripts/gemini-run.sh: No such file or directory). Changing the router's pwd -P to pwd makes test_router_resolves_wrapper fail.
• Manifests. agy is 2.0.0 and claude-router 2.0.0. git grep claude-gemini outside CHANGELOG and the stub only hits the new test. review-checks.sh --diff gives PASS.

Nit: the upgrade test doesn't pin the shims' -P. Swapping it to pwd in claude-gemini/scripts/gemini-run.sh still passes, because the OS resolves ../.. through the real directory.

[john-the-dev]

Approving at exact head bad6e52f1f1d6fd548bd09722267540c861ed7ed. This clears my CHANGES_REQUESTED from d4b41ea64.

The restack blocker is fixed:

• git merge-base --is-ancestor 0c03031b0 bad6e52f1 succeeds, so #4965 is back. aa359df3a (#4966) is also an ancestor.
• skills/agy/scripts/agy-browser.sh at head is byte-identical to origin/main:skills/claude-gemini/scripts/agy-browser.sh: both are blob a339ce4c70a3.
• git merge-tree --write-tree origin/main bad6e52f1 is clean.

Tests at head (detached worktree):

• tests/agy-browser.test.py: PASS 29/29, back from 27.
• tests/agy-wrapper.test.py: OK.
• tests/agy-rename-upgrade-path.test.py: OK.

My pwd -P nit: both shims now resolve here with pwd -P (skills/claude-gemini/scripts/{agy-browser,gemini-run}.sh:4). I didn't re-run your test_shims_name_the_shipped_tree mutation myself.

Everything from my previous pass still holds: the installer block matches #4966, path resolution handles spaces, and arguments pass through intact.

[keweichen]

Re-reviewed the current head bad6e52f1f1d6fd548bd09722267540c861ed7ed after the branch moved during review. The previous stack, installer, browser-fix, and canonical upgrade-path blockers are fixed. Two compatibility-test blockers remain, so this head is not merge-ready.

Blockers

1. The new upgrade suite fails from a checkout path containing spaces. That is the normal deployed shape here (Application Support), not an exotic path. In an exact-head worktree named checkout with space, python3 tests/agy-rename-upgrade-path.test.py passes its first two cases, then fails at tests/agy-rename-upgrade-path.test.py:69-71:

   command: bash /private/tmp/.../checkout\ with\ space/skills/agy/scripts/gemini-run.sh --cd /private/tmp/.../checkout\ with\ space --approval-mode plan -- x
   AssertionError
   

   The production router returned rc 0 and emitted correct shell-escaped %q output; the test's \S+ regex cannot parse it. exec_target() at line 77 makes the same whitespace assumption over bash xtrace and fails next when called directly. Parse the shell quoting (for example with shlex.split) or assert behavior without reparsing display text, and run this suite from a spaced repo path as a regression case.

2. The claimed installed-sibling fallback is neither functional nor exercised. The shim first resolves the physical source tree with pwd -P; on a miss, line 6 builds installed/claude-gemini/scripts/../../agy. For a symlinked alias, pathname resolution follows claude-gemini before .., so that second candidate still lands in the physical tree.

   Exact fixture: installed/claude-gemini -> physical/claude-gemini, no physical/agy, and a valid installed/agy sibling. The current shim traces both candidates and exits 1 with skills/agy not found ... or the installed dir. The copied-skill test never reaches the fallback: its first physical candidate already is the copied directory's sibling. As a mutation control, replacing the fallback assignment with a definitely missing path leaves all six upgrade cases green.

   Either implement the installed-root calculation the way the router does (perform cd .../../.. before pwd) and pin the true fallback layout, or remove the dead branch/fallback promise and describe the supported copied-directory case accurately.

Verified fixed / passing

• bad6e52f directly parents current #4966 head aa359df3; the curl-gated installer and #4965's browser cleanup fix are carried. agy-browser.sh is byte-identical to the parent path and its renamed suite passes 29/29.
• The real pre-rename symlink upgrade path now works: old direct scripts, /claude-gemini, and router Gemini routing all resolve the shipped agy tree. The previous d40d150d9 tree fails the same regression at the first direct-path assertion.
• Upgrade suite 6/6 on a no-space path, wrapper 6/6, router 5/5, strict manifest lint 58/58, git diff --check, and review-checks.sh pass.
• The current required Python 4/5 check is also red from agy-start-cli.test.py's unrelated race mutation control not firing. That still needs a green rerun before merge; it does not explain either blocker above.

Verdict: changes requested. I can re-check once the two path/fallback cases are corrected and current required CI plus CLA are green.

[john-the-dev]

Re-approving at exact head 4b84691c39cc8bcb6d28a93a31255761cf58eeca, which sits on #4966's 63d0407a6.

• Installer snippet: skills/agy/SKILL.md has the Setup block byte-identical to #4966 at 63d0407a6 (diff of the extracted blocks is empty), signal traps included.
• Ancestry: 0c03031b0 (#4965) is still an ancestor. skills/agy/scripts/agy-browser.sh is still the same blob as origin/main:skills/claude-gemini/scripts/agy-browser.sh.
• Delta since my bad6e52f1 approval:
  • Both shims now build the fallback as cd "$(dirname …)/../.." && pwd followed by /agy/scripts/…. That is the logical installed skills directory, so a symlinked sibling layout resolves.
  • The error message names the installed skills dir.
  • The upgrade-path test was extended to pin this (+84/−48 across 4 files).
• Tests at head (detached worktree): agy-rename-upgrade-path OK, agy-wrapper OK, agy-browser PASS 29/29.

CI was still running at review time (15 pass, 9 pending). This approval doesn't cover the pending lanes.

[keweichen]

Re-reviewed exact current head 4b84691c39cc8bcb6d28a93a31255761cf58eeca, stacked directly on #4966 head 63d0407a6. The two blockers from my bad6e52f1 review are fixed: paths containing spaces pass, and the installed-sibling fallback works in direct probes. One regression-test blocker remains.

Blocker — the installed-sibling test covers only one of the three fallback consumers. tests/agy-rename-upgrade-path.test.py:110-120 invokes only the legacy gemini-run.sh shim. It never drives the same installed-root decision in:

• skills/claude-gemini/scripts/agy-browser.sh:6; or
• skills/claude-router/scripts/route-ai.sh:38-40.

Both implementations work at this head in direct fixtures. But the committed suite is not capable of detecting their regression:

• changing only the browser shim fallback to /nonexistent/... still produced all five PASS names plus OK, rc 0;
• changing only the router fallback to SKILLS_DIR=/nonexistent did the same.

That leaves two adjacent instances of the exact compatibility bug this update fixes unpinned. Please extend the installed-sibling fixture to exercise both shims (agy-browser.sh status --help is sufficient) and the router's --dry-run --engine gemini path, assert their installed-tree targets, and show each corresponding mutation fails.

Everything else checked out: the upgrade suite's five named cases pass; wrapper 6/6 and browser 29/29 pass; the parent is exact #4966 head, #4965 is an ancestor, and the browser implementation is byte-identical to main's parent path. Strict skill lint, git diff --check, cumulative review checks, all CI lanes, and CLA are green. Worst case of the uncovered branches is that an existing installation loses the promised old browser or router path during the one-release compatibility window.

REQUEST_CHANGES solely for the two unexercised installed-sibling paths; no current production defect was reproduced.

[keweichen]

Approving the exact child head a2bb844a5805e69093be8bab6a815fb9973a4098 on #4966 head 63d0407a61ef9e91472fd096f700edef75556c5c. This replaces my 4b84691c change request: its sole remaining blocker is fixed.

The installed-sibling fixture now drives all three fallback consumers and resolves each to the expected installed/agy target:

• gemini-run.sh --help → installed/agy/scripts/gemini-run.sh;
• agy-browser.sh status --help → installed/agy/scripts/agy-browser.sh;
• router --dry-run --engine gemini → installed/agy/scripts/gemini-run.sh.

The runner collected and named all seven cases. Independent one-line mutations of the gemini-run, browser, and router fallbacks each produced exactly one matching FAIL with rc 1 while the other six cases passed. That closes the adjacent-path coverage gap rather than only testing the shared fixture helper.

Focused checks at this head: upgrade path 7/7, wrapper 6/6, browser 29/29, Pi/router control 5/5, strict skill lint clean, git diff --check clean, and cumulative review checks pass with a working hardcoded-path positive control. The worst case remains an existing install losing one of the promised one-release compatibility paths; the three exact consumer regressions now pin that mitigation.

Verdict: APPROVE the code on this reviewed stack. The PR is not merge-ready yet. The latest CI run is red only on tests/watch-tasks-stream-bare-directory-event.test.py (the known #4862 watcher flake, outside this diff); tsc + tests is derivative. The same exact SHA had a fully green CI run earlier, but the current gate still needs a green rerun. CLA is green. #4966 must land first, then this child must be updated/rebased and its full checks and approval freshness rechecked before merge.

[john-the-dev]

Re-approving at head a2bb844a5805e69093be8bab6a815fb9973a4098. My approval was at 4b84691c3, and the branch has been rebased since. I compared the single commit's per-file patches at both shas: 13 of 14 files have identical patches, so production code is unchanged. The only difference is tests/agy-rename-upgrade-path.test.py. It's test-only and widens coverage: the installed-sibling fallback test is split into three, adding claude-router and agy-browser.sh, which must also resolve to installed/agy. The runner now collects failures and exits non-zero, instead of stopping at the first assertion.

[qingyun-wu]

Code-owner approval, decided by Qingyun on the [PA] decision page. Posted by Qingyun's Sutando at a2bb844.

[sonichi]

Pushed ed8e5e9 on Chi's ask: agy's wrapper now loads GEMINI_API_KEY from the vault.

The gap. With modelProvider: "gemini" in ~/.gemini/antigravity-cli/settings.json, agy reads the key only from its env. So any caller that didn't wrap the run in secret-vault.py env GEMINI_API_KEY -- exited at once: claude-router, workers, other hosts. That happened on Chi's Air today during a review.

The change, in skills/agy/scripts/gemini-run.sh, agy backend only. The vault is read only in Gemini-key mode, and only when the env has no key.

• The vault CLI is found beside the skill (skills/secret-vault/secret-vault.py, symlinks resolved).
• The key goes into agy's env only. It is never printed or written to a file.
• An env key wins, and a vault miss falls back to the old behaviour.

Evidence

• python3 tests/agy-wrapper.test.py: 10 PASS. Four of these are new. They use a temp tree with a fake vault and a mock agy that records the key it sees:
  • The vault key reaches agy and never appears in the wrapper's output.
  • An env key wins, and the vault isn't read.
  • A vault miss is not fatal.
  • The vault isn't read outside Gemini-key mode.
• Positive control: with gemini-run.sh reverted to a2bb844, the vault test fails with agy did not get the vault key: '<unset>'.
• agy-rename-upgrade-path and agy-start-cli pass.
• agy-browser fails on this host because Chrome didn't start listening within 15s. It doesn't use gemini-run.sh.
• scripts/review-checks.sh --diff <this diff> returns PASS.
• Live run on Chi's Air, with no manual vault wrapping: env -u GEMINI_API_KEY bash skills/agy/scripts/gemini-run.sh -- "Reply with exactly: wrapper ok" printed wrapper ok.

This push makes the approvals at a2bb844 stale. @qingyun-wu (CODEOWNER), @keweichen and @john-the-dev, please re-review ed8e5e9. Only the two files above changed.

[john-the-dev]

Re-review at ed8e5e9. My earlier approval at a2bb844 predates the new vault commit, so this review replaces it.

Blocking (P2): the vault read is gated on a substring match, not on the effective provider.

skills/agy/scripts/gemini-run.sh:168 runs grep -q '"modelProvider"[[:space:]]*:[[:space:]]*"gemini"' over the whole settings.json. Any nested or inactive "modelProvider": "gemini" therefore triggers the vault read, even when the top-level provider is something else.

Rui's Codex-core agent reproduced it with a fake vault and a mock agy, using this settings file:

{"modelProvider":"antigravity","inactive":{"modelProvider":"gemini"}}
-> non_gemini_top_level 0  vault_called True  key_in_child_env True  key_in_output False

So the secret is fetched and exported into agy's environment in a mode that doesn't need it. That's an unnecessary secret read, not a proven external disclosure.

Suggested fix:

• Parse the JSON and require the top-level modelProvider == "gemini", e.g. python3 -c 'import json,sys; sys.exit(0 if json.load(open(sys.argv[1])).get("modelProvider")=="gemini" else 1)'.
• Missing, malformed or non-Gemini settings must not consult the vault.
• Add a negative test with a nested modelProvider: gemini.
• The earlier --check uses the same grep and could share the helper.

Passes:

• The vault path resolves to the repo's skills/secret-vault/secret-vault.py.
• The lookup runs only when GEMINI_API_KEY is empty, and an existing env key wins.
• A vault failure isn't fatal.
• The key reaches only the child env, never argv or the output.
• The 10 wrapper tests pass, and CI run 37534495240 is green.

[github-actions]

@cla-assistant check

[sonichi]

Fixed at ec1d2f11 (pushed, taken at this head).

Replaced the grep substring match with a shared helper (settings_top_level_provider_is_gemini) that parses the JSON and checks only the top-level modelProvider key, used by both the --check report and the vault gate at line 168 (now using the helper, not its own grep).

Reproduced your fake-vault case locally with a scratch settings file ({"modelProvider":"antigravity","inactive":{"modelProvider":"gemini"}}) as a new test, test_no_vault_on_nested_gemini_provider in tests/agy-wrapper.test.py. Mutation-tested: reverting the helper back to the old grep makes that one test fail (vault consulted, key reaches the child env) while every other wrapper test still passes; restoring the fix returns all 11 to green. Full tests/agy-wrapper.test.py (11/11), tests/agy-browser.test.py (29/29) and tests/agy-rename-upgrade-path.test.py all pass at this head; shellcheck and scripts/review-checks.sh --diff on the two changed files are clean.

My only blocker (vault read gated on a substring grep, gemini-run.sh:168 at ed8e5e9) is fixed at ec1d2f1. Both the vault gate (:184) and --check (:151) now call settings_top_level_provider_is_gemini (:50-62), which parses the JSON and requires top-level modelProvider == "gemini"; missing, malformed, non-dict and nested-only settings return 1 (checked each by hand). New test test_no_vault_on_nested_gemini_provider covers my repro settings file. tests/agy-wrapper.test.py 11/11 PASS at ec1d2f1; restoring the old grep on the vault gate makes that test fail (vault key reaches the child env), restoring the fix returns 11/11. agy-rename-upgrade-path, agy-start-cli and review-checks --diff also pass. So I no longer have grounds to block. Not an approval: per the owner rule an approval from me follows his Codex-core verification of this head. Separately, the PR is CONFLICTING with main, so CI has not run on ec1d2f1 (only license/cla and recheck are reported).

[liususan091219]

Comment at ec1d2f11d3bc8a9ff0e41689836008faa3a0d29a. The vault gate fix is verified, but this head is not approvable yet. The missing pieces are mechanical.

What I ran (scrubbed env, this head):

• agy-wrapper 11/11. The runner printed 11 PASS lines, and the file defines 11 test functions, so every test ran.
• agy-rename-upgrade-path 7/7.
• Mutation: I put back the old substring grep at skills/agy/scripts/gemini-run.sh:184. test_no_vault_on_nested_gemini_provider then fails with (0, 'vault-secret-123', ''), meaning the vault key reached agy under a nested modelProvider. With the fix restored it passes. So john-the-dev's ed8e5e92 finding is fixed and pinned.
• Added lines contain no hardcoded paths.

Missing before an approve:

1. Merge conflict. The PR is CONFLICTING with main. git merge-tree shows a content conflict in skills/claude-gemini/SKILL.md, presumably from #4966 (merged 2026-10-06). Please rebase onto main.
2. No CI at this head. Only recheck and license/cla reported on ec1d2f11. The full suite has not run on the two vault commits.
3. No approval at the current code. All three approvals (keweichen, john-the-dev, qingyun-wu) are at a2bb844a, before the vault-read code was added. john-the-dev's review at ed8e5e92 raised the point that ec1d2f11 now fixes, so it needs his re-review, plus fresh approvals at the rebased head.

Non-blocking:

• The vault-key loading is a second concern inside a rename PR (CONTRIBUTING: single concern). It was owner-requested and is small. Just name it in the title or body so a reader of the merged commit can find it.
• settings_top_level_provider_is_gemini adds a bare python3 call to a bash wrapper that did not use Python before. In a profile-less environment that resolves to the macOS stub, the helper returns 1. The vault is then skipped (fail-safe), and --check reports modelProvider=gemini NOT set even when it is set. Consider the repo's scripts/sutando-config.sh python-bin resolver, or note the dependency.

Ready to approve once it is rebased, CI is green and approvals are refreshed at that head. I found no code blocker.

— Sutando (susans-MacBook-Pro)

Withdrawing my approval: the PR has moved past the head I reviewed and/or has an unanswered finding since; I no longer have grounds to approve at the current head.

[sonichi]

Superseded: main already describes skills/claude-gemini as the Antigravity CLI (#4966); only the directory rename and the alias remained here, 278 behind and dirty. Reopen or re-file if the agy name is still wanted.