feat(framework+cli): fw-4.42.0 / cli-3.43.0 — Antigravity CLI succeeds Gemini CLI

[montfort]

What

Google retired Gemini CLI; its successor is Antigravity CLI (agy). StrayMark's Gemini-era surfaces are retired with it — and the framework gains the mechanism that makes retiring a distributed path possible at all.

The correction that reshaped this

The obvious reading of "discontinue Gemini CLI" is half wrong. Verified against agy 1.1.10 installed locally and against agy-customizations, the skill Google ships inside the binary as the authoritative spec of its customization system:

Surface	What Antigravity actually does
GEMINI.md	Reads it. A first-class Rules file alongside AGENTS.md and .agents/rules/*.md.
~/.gemini/	Its own global config dir. The binary carries a resolve_gemini_dir symbol; ~/.gemini/skills/, settings.json and GEMINI.md are live.
<project>/.gemini/skills/	Not a discovery location. Workspace roots are .agents/, .agent/, _agents/, _agent/.
<project>/.agent/workflows/	Read by nothing. Right root, wrong shape — agy wants skills/<name>/SKILL.md, and the binary contains no agent/workflows string.

The product was retired; the on-disk contract was kept. So GEMINI.md stays (deleting it would remove a rules source from the very CLI we're adding support for), retitled to the Antigravity lineage with identity antigravity-v{version} and a note explaining why the filename persists — without it the next reader deletes it as a leftover.

Meanwhile .agent/workflows/ — the channel the README advertised as the Antigravity surface — was read by nothing at all.

Changes

New channel dist/.agent/skills/, generated from the Claude source rather than hand-mirrored. That choice came from the data: .gemini/skills/ had drifted behind .claude/ in 7 of 15 skills because nothing regenerated or gated it, while the generated .codex/ channel was byte-perfect. A git mv would have imported the stale files. gen_codex_skills → gen_minimal_skills, emitting both minimal-frontmatter channels under the existing CI gate.

New retired: manifest key. Retiring a distributed path was previously impossible to complete: update_files only copies, so a directory dropped from files: survived in every existing installation until someone ran straymark remove. Deletion is provenance-gated — a file goes only when its hash still matches what .checksums.json recorded. Operator-edited files and files StrayMark never installed are kept and reported distinctly, because telling someone they "modified" a file they wrote themselves erodes trust in the report.

Also: --agent gemini → --agent agy (the guidance no longer derives the source dir from the agent name — Antigravity reads .agent/skills/, not .agy/skills/); auditor-CLI prose across skills, governance, CLI output and website; docs in EN/es/zh-CN, including the zh-CN CLI-REFERENCE that was two releases behind.

The line I drew: CLI vs model

The Gemini CLI was retired; the Gemini models were not, and agy serves them. The audit skills already require auditor: to name the backend model rather than the CLI. So prose telling the operator which CLI to open moved to agy, while model ids, Sentinel telemetry, test fixtures and the Charter-template rule justified by an observed Gemini-auditor behavior across 2 cycles were left exactly as written. Rewriting the actor in a piece of evidence would falsify the evidence.

Verification

• cargo test: 953 passed, 0 failed.
• New retired_paths_test.rs + four unit tests covering the three provenance outcomes, full-subtree cleanup, the no-op case, and the pre-4.42.0 manifest parse (#[serde(default)] is load-bearing — every older .straymark/dist-manifest.yml lacks the key).
• Manual e2e on a simulated fw-4.41.0 install: two pristine files pruned, one operator-edited file kept and flagged as modified, one operator-authored file kept and flagged as not installed by StrayMark, parents cleaned only when empty.

The e2e caught two defects before they shipped: repair returned early on a healthy installation, so the sweep would never have run where it mattered; and the first report conflated "you edited this" with "this was never mine". A unit test caught that only the retired root was cleaned, leaving emptied skill subdirectories behind.

Root confirmed empirically. .agent/ vs .agents/ was the one assumption the code couldn't settle — Google's doc lists .agent/ as an alias, the binary's path templates only build .agents/. Three headless agy -p probes failed to distinguish them (print mode doesn't exercise workspace customization discovery at all, even in a trusted dir with a git root). The operator ran it interactively: agy listed both probe skills. .agent/ is a real alias; shipped channel unchanged.

Release

fw-4.41.0 → fw-4.42.0, cli-3.42.0 → cli-3.43.0. Minor, not major: two directories leave the adopter's tree but neither was consumed by a live runtime, the content reappears in .agent/skills/, and migration is automatic via retired:.

Decision in AIDEC-2026-08-05-002, work in AILOG-2026-08-05-004.

🤖 Generated with Claude Code