Antigravity Handoff Guide & In-Repo Skill Gaps
Antigravity Handoff Guide & In-Repo Skill Gaps
Section titled “Antigravity Handoff Guide & In-Repo Skill Gaps”Status: Contributor reference for the handoff of the two 2026-06-18 plans to Google Antigravity / Gemini 3.5 Flash.
Research basis: ../architecture/gemini-3-5-flash-antigravity-limitations-2026-06-18.md (execution-target profile) + ../architecture/auto-derivation-design-hygiene-2026-06-18.md.
Plans: ../../superpowers/plans/2026-06-18-track-a-naked-objects-auto-gui.md · ../../superpowers/plans/2026-06-18-track-b-zero-annotation-debugging.md.
1. Why this guide exists
Section titled “1. Why this guide exists”The plans will be executed by Gemini 3.5 Flash inside Antigravity, not Claude Code. Antigravity reads GEMINI.md (its overrides) and AGENTS.md (cross-tool, since v1.20.3), and mounts skills from .agents/skills/. It cannot see Claude’s external ~/.claude/ superpowers plugin cache. So any skill a plan references must live in-repo — and as of 2026-06-18 all eleven do (crates/vox-skills/skills/superpowers/, including deep-research and research). This doc maps them and summarizes each, and lists the handoff prerequisites.
2. Antigravity execution model (what the plans target)
Section titled “2. Antigravity execution model (what the plans target)”- Orchestrator + dynamic subagents with isolated context windows, dispatched in parallel.
- Reliability is the hard constraint: ~48% real-world task completion; mid-task termination leaves no checkpoint; quota is a hard cutoff. Therefore the plans are written so that each task is atomic, ends green, and is committed — a kill wastes at most one task.
- Gemini 3.5 Flash: strong agentic coder; weaker deep reasoning and long-context recall than Pro; prone to API hallucination. Plans counter this with verify-before-use steps and two-strike circuit breakers (see research §3.3).
3. Parallel vs sequential — the dispatch rule
Section titled “3. Parallel vs sequential — the dispatch rule”Each plan task is tagged [PARALLEL-SAFE] or [SEQUENTIAL]:
[PARALLEL-SAFE]— touches a disjoint file set from other parallel-safe tasks in the same wave; an isolated-context subagent can own it. Dispatch these together.[SEQUENTIAL]— modifies a file an earlier task also modifies, or depends on an earlier task’s output; run in order on one agent.
Golden rule: never dispatch two subagents that write the same file. Antigravity’s isolated contexts will not see each other’s edits and will clobber. When in doubt, mark [SEQUENTIAL].
4. In-repo skill map
Section titled “4. In-repo skill map”| Plan reference | In-repo path (use this) |
|---|---|
| Writing plans | crates/vox-skills/skills/superpowers/writing-plans.skill.md |
| Subagent-driven execution | crates/vox-skills/skills/superpowers/subagent-driven-development.skill.md |
| TDD (red-green-refactor) | crates/vox-skills/skills/superpowers/test-driven-development.skill.md |
| Systematic debugging | crates/vox-skills/skills/superpowers/systematic-debugging.skill.md |
| Brainstorming | crates/vox-skills/skills/superpowers/brainstorming.skill.md (NOW NATIVE) |
| Dispatching parallel agents | crates/vox-skills/skills/superpowers/dispatching-parallel-agents.skill.md (NOW NATIVE) |
| Verification-before-completion | crates/vox-skills/skills/superpowers/verification-before-completion.skill.md (NOW NATIVE) |
| Code review | crates/vox-skills/skills/superpowers/requesting-code-review.skill.md (NOW NATIVE) |
| Git worktrees | crates/vox-skills/skills/superpowers/using-git-worktrees.skill.md (NOW NATIVE) |
| Deep research | crates/vox-skills/skills/superpowers/deep-research.skill.md (NOW NATIVE) |
| Research (lightweight web) | crates/vox-skills/skills/superpowers/research.skill.md (NOW NATIVE) |
Update 2026-06-18: all eleven skills are now in the Vox native library under
crates/vox-skills/skills/superpowers/. Antigravity mounts repo skills from.agents/skills/; ensure that path resolves to (or symlinks/copies)crates/vox-skills/skills/superpowers/so Gemini can load every skill above. The §5 entries below are quick-reference summaries — the authoritative skill is the file.
Policy the agent must obey throughout: AGENTS.md (root), plus the Vox-specific rules in CLAUDE.md/GEMINI.md — notably VoxScript-first automation (no .ps1/.sh/.py; use vox run scripts/*.vox), never cargo fmt --all (use cargo fmt -p <crate>), and .md under docs/src/ needs YAML frontmatter.
5. Quick-reference summaries (authoritative skill = the native file)
Section titled “5. Quick-reference summaries (authoritative skill = the native file)”These five skills are now native (paths in §4). The summaries below are a fast reference; load the full .skill.md for complete guidance.
5.1 brainstorming
Section titled “5.1 brainstorming”Use when: a task requires a design choice the plan did not fully specify. Do: (1) State the decision in one sentence. (2) List 2–3 concrete options with one-line trade-offs. (3) Pick one and record why in a comment or commit message. Do NOT start coding until the choice is written down. For Gemini 3.5 Flash: never invent a fourth “clever” option that requires APIs you have not verified exist.
5.2 dispatching-parallel-agents
Section titled “5.2 dispatching-parallel-agents”Use when: 2+ tasks are tagged [PARALLEL-SAFE] in the same wave.
Do: (1) Confirm their file sets are disjoint (re-read each task’s Files block). (2) Spawn one subagent per task with ONLY that task’s text as context (isolated window). (3) Wait for all to return green. (4) Integrate sequentially: pull each result, run the full crate test suite once, resolve any surprise overlap. Never parallelize tasks that share a file. Two-strike rule: if a subagent fails twice, stop it and surface its handoff note rather than re-dispatching.
5.3 verification-before-completion
Section titled “5.3 verification-before-completion”Use when: about to mark any task done or commit.
Do, in order, and paste the actual output: (1) cargo test -p <crate> → must show PASS counts. (2) cargo clippy -p <crate> -- -D warnings → must be clean. (3) cargo fmt -p <crate> (never --all). (4) vox stub-check → must report no stubs. (5) confirm the tree compiles (cargo check -p <crate>). Rule: evidence before assertion — do not claim “done” without pasted command output. If any fails, the task is NOT done.
5.4 code-review (self-review pass for a fast model)
Section titled “5.4 code-review (self-review pass for a fast model)”Use when: a task’s implementation step is written, before its commit.
Checklist: (1) Does every symbol/path I referenced actually exist? (re-grep if unsure — anti-hallucination). (2) Did I add a stub/placeholder? (forbidden — see AGENTS.md/feedback_no_stubs). (3) Does the change match the task’s stated Files block exactly? (4) Did I duplicate logic that already exists (DRY)? Fix inline; do not expand scope.
5.5 using-git-worktrees
Section titled “5.5 using-git-worktrees”Use when: the plan says to isolate parallel file-mutating work.
Do: for [PARALLEL-SAFE] waves that each mutate files, give each subagent its own worktree: git worktree add ../wt-<task> <branch>; work there; commit; then integrate. Clean up with git worktree remove. For this repo’s two plans, most tasks touch disjoint files in one crate, so worktrees are optional — use only if two parallel tasks would otherwise race the same target dir. On Windows, prune stale worktrees before deleting branches.
6. Handoff checklist (run once before starting either plan)
Section titled “6. Handoff checklist (run once before starting either plan)”-
AGENTS.mdandGEMINI.mdpresent and loaded (the latter may need creating fromCLAUDE.mdfor Antigravity). -
cargo run -p vox-arch-checkpasses (baseline). - The plan’s Pre-flight
rgcommands have been run and the real signatures confirmed (anti-hallucination). - Decide track order (all independent at the code level; recommended: Track A → Track C → Track B). Plans: Track A · Track B · Track C.
- Confirm parallel waves: list which task IDs are
[PARALLEL-SAFE]together.