Skip to content

feat(antigravity): adopt Carbon Medium default and add /start-small workflow - #125

Merged
futuregerald merged 1 commit into
mainfrom
feat/carbon-medium-default-and-start-small
Nov 15, 2026
Merged

futuregerald merged 1 commit into
mainfrom
feat/carbon-medium-default-and-start-small

Conversation

@futuregerald

Copy link
Copy Markdown
Owner

Summary

  • Updates the packaged Antigravity rules, templates, and model-routing references to make Gemini 4 Carbon (Medium) (2×) the default Coordinator and /start peer-pipeline-swarm implementer (gemini-4-carbon-medium), while preserving Carbon High (gemini-4-carbon-high) and Claude Opus 5.5 High (claude-opus-5-5-high) on adversarial reviews.
  • Adds the Antigravity-exclusive /start-small workflow skill (internal/catalog/antigravity/skills/start-small/SKILL.md), registered in catalog.go (IsAntigravityOnlySkill) and tracks.yaml (antigravity_only: true), so smaller bounded changes can run an Argon Medium (1×) swarm implementer with Carbon High peer reviewers without switching the chat UI model or spending weekly Anthropic Opus quota.
  • Tightens scripts/sync-antigravity-from-local.sh so shared skill references/ directories are not duplicated into internal/catalog/antigravity/skills/.

Background

Context

skill-installer embeds harness-specific templates and skills for Claude Code, Google Antigravity, Zed, and OpenCode. On Antigravity, the opinionated track installs GEMINI.md, GEMINI-BASE.md, peer-pipeline-swarm, baton-pass, and antigravity-orchestration while strictly isolating Antigravity-exclusive skills from non-Antigravity targets.

The problem

Controlled end-to-end A/B benchmarking across a 24-assertion zero-knowledge browser-extension suite showed that gemini-4-carbon-medium outperformed the previous gemini-4-argon-medium default 24/24 (100%) vs 20/24 (83.3%) (2–0 blind review), finished implementation 2.8m faster in 13.5% fewer turns, and reduced downstream reviewer rework by -24.1% (-40.9% less weekly Opus quota). Meanwhile, users still need a 1-command lightweight path (/start-small) for routine bug fixes that stays on 1× Argon Medium without changing the UI dropdown, and Model: "inherit" on a Carbon Medium session would otherwise downgrade invoke_subagent reviewers from Carbon High to Medium.

The fix

  • Carbon Medium default + Carbon High review preservation: Updated GEMINI.md, GEMINI-BASE.md, antigravity-orchestration, peer-pipeline-swarm, start, lifecycle, plan-review, comprehensive-code-review, editor-pass, subagent-driven-development, and future-model-router (SKILL.md + references/model-map.md) so /start defaults to gemini-4-carbon-medium, non-Carbon High UI sessions preserve gemini-4-carbon-high via one-shot swarm workers on non-trivial reviews, and >1-file / >15-LOC post-review or Phase 12 editor-pass fixes delegate to implementer + fix-check.
  • /start-small Antigravity-exclusive skill: Synced start-small/SKILL.md, added "start-small" to IsAntigravityOnlySkill and tracks.yaml (opinionated.antigravity_skills + skills: metadata), and verified C4 cross-reference closure (start/SKILL.md -> ../start-small/SKILL.md).

Pre-existing fixes

  • Guarded the references/ copy block in scripts/sync-antigravity-from-local.sh so skills whose references/ directory belongs to internal/catalog/shared/skills/<skill>/references/ are not copied back into internal/catalog/antigravity/skills/ after a local skill-installer -t antigravity -g run.

Test plan

  • bash scripts/sync-antigravity-from-local.sh — clean sync and PII/org regex sweep (exit 0).
  • internal/catalog/catalog_test.go — updated TestLoadFromDirFS (55 skills), TestSupportedHarnessesAndBatonPassExclusivity, TestEveryCatalogSkillDirectoryHasCompleteMetadata, TestTrackCrossReferenceClosure_C4, TestHarnessTemplatesDistinctAndBatonFreeOutsideAntigravity, and added TestAntigravityCarbonMediumDefaultAndStartSmallCatalogInvariants.
  • main_test.go & e2e/e2e_test.go (-tags=e2e) — verified start-small installs on Antigravity, appears in skill-installer list -t antigravity --long, and is excluded across Claude Code, Zed, OpenCode, and default list --long in isolated t.TempDir() sandboxes.
  • PYTHONDONTWRITEBYTECODE=1 pytest -p no:cacheprovider -q ... — 742 passed (including test_every_antigravity_skill_passes_validation[antigravity/start-small]).

Rollback plan

Revert this PR (git revert <merge-sha>). No state file or persistent schema changes were introduced.

@futuregerald
futuregerald merged commit 282dffc into main Nov 15, 2026
1 check passed
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant