Superpowers plugin adopted; fleet expanded to 5 agents (2026-07-04)

decision Sat Jul 04 2026 00:00:00 GMT+0000 (Coordinated Universal Time) fleetagentssuperpowersantigravitycopilotcodexgeminihaifree-tier

Superpowers on the fleet — decision 2026-07-04

Locked

• Adopt obra/superpowers methodology (brainstorm → design → plan → subagent-driven TDD/YAGNI/DRY).
• Vendor 11 non-overlapping skills into repos/own/infra/agent-skills/ (single canonical source per skills-cli-canonical).
• Merge 3 direct collisions (tdd, plan, skill-creator) — preserve our locked clauses (no-test-cheating, CC-native plan-mode metadata, CC skill-test harness).
• Fleet grows CC + OpenCode + Kilo → CC + OpenCode + Kilo + Antigravity + GH Copilot CLI + Codex CLI (6 agents, corp-only for Codex).
• No plugin-install-per-agent — npx skills update -g -y fans agent-skills to every CLI agent that supports it. Superpowers's own plugin marketplace bypassed for our fleet.

Grill answers (2026-07-04)

1. Agent scope: CC + OpenCode + Antigravity (reintroduced) + GH Copilot CLI + Codex CLI (corp). Kilo unsupported by Superpowers upstream, stays on agent-skills bundle only.
2. Antigravity reintroduction gate: winget install Google.AntigravityIDE + Google.AntigravityCLI shipped — headless installer now exists. Rebuts the 2026-07-02 drop reason ("standalone IDE, manual install, no headless mode").
3. LLM provider per agent — free-tier verified:

All free-tier verified. Zero card added.

4. Skill collision resolution — merge (not replace, not coexist):

5. Skipped Superpowers targets:

• Kilo Code — not supported by Superpowers upstream (only 10 harnesses listed; Kilo not among them). Kilo already gets the vendored skills via npx skills update.
• Cursor — requires card-on-file. Blocked per no-card-on-file.
• Kimi Code — Moonshot credits expire → card required. Blocked.
• Pi — Primeradiant's paid agent (obra's own). Blocked.
• Factory Droid — evaluated, not currently in fleet.

Vendored skills (11) — from obra/superpowers @ d884ae04edebef577e82ff7c4e143debd0bbec99

• brainstorming — activates before creative work; refines rough ideas through questions
• dispatching-parallel-agents — 2+ independent tasks without shared state
• executing-plans — execute written plan with review checkpoints in a separate session
• finishing-a-development-branch — merge / PR / cleanup after implementation complete
• receiving-code-review — verify feedback with rigor, not performative agreement
• requesting-code-review — verify work before merge
• subagent-driven-development — execute plan with subagents in current session
• systematic-debugging — before proposing any bug fix
• using-git-worktrees — isolate workspace before feature work
• using-superpowers — bootstrap on conversation start
• verification-before-completion — evidence before claims

Env / secrets wiring

Per env-to-system-env-daily rule:

1. pwsh scripts/add-env-key.ps1 GEMINI_API_KEY <key> — you paste key once; script encrypts to .env.enc, pushes to user env
2. pwsh scripts/add-env-key.ps1 HAI_API_KEY <key> — from hai proxy start output on corp laptop
3. GH Copilot Free uses gh auth login (no env var needed)
4. Agents read via ${env:VAR} refs in their config files — no hardcoded secrets

Refresh from upstream

To pull new Superpowers commits:

node scripts/refresh-superpowers.mjs

Deferred — not shipped this turn. Manual refresh: re-run the vendor curl loop from the 2026-07-04 commit message, bump SUPERPOWERS-ATTRIBUTION.md SHA.

Rejected alternatives

• Fork obra/superpowers into repos/frk/ — heavier than vendor. Vendor keeps single source at agent-skills; forking would create a maintenance surface without upside for this use case. Revisit if we're contributing upstream regularly.
• Wholesale replace ours with Superpowers's on the 3 collisions — would lose no-test-cheating anti-cheat clause + CC skill-test-harness affordances. Merge preserves both.
• Plugin-install per agent — would create two skill sources per agent (Superpowers plugin + agent-skills npx bundle). skills-cli-canonical rule blocks this.

Follow-ups

• Antigravity's agy binary confirmed at %LOCALAPPDATA%/agy/bin/agy.exe — install-agents.ps1 updated to install Google.AntigravityIDE + Google.AntigravityCLI on new machines
• install-agents.ps1 updated with Antigravity + GH Copilot CLI steps (2026-07-04)
• 3-place update: AGENTS.md fleet table updated, dropped-list updated, this OKF file created
• New rule filed: none needed — decision is expressive enough at its own file; behavior falls under existing skills-cli-canonical + agent-fleet-parity

Live-verify results (2026-07-04)

• ✅ winget install Google.AntigravityIDE + Google.AntigravityCLI — both installed silently. agy --version returns 1.0.14. Google account sign-in required on first launch of the IDE.
• ✅ ~/.codex/config.toml + ~/.codex/auth.json written from existing .env HAI_API_KEY. No secret leaked to logs.
• ⚠ Codex CLI version pin: 0.125.0. Later versions (0.126+ through 0.142.4) send namespace / custom / tool_search tool declarations that hai's Azure GPT-5 and GPT-5.4 both reject with BadRequest: The following tools are not allowed for model 'gpt-5'. Downgrade via npm i -g @openai/[email protected]. Test with codex exec --skip-git-repo-check "hello" — returns response in ~5s on hai chain.
• ✅ sync-env-to-system-env.ps1 pushed 194 env vars to user env, including GEMINI_API_KEY (53 chars) + HAI_API_KEY (36 chars, UUID form).
• ✅ npx skills update -g -y fanned 53 skills to ~/.claude/skills/. Same set will fan to OpenCode/Kilo on their next skills sync.
• ⚠ GH Copilot CLI still requires installing the standalone binary from github.com/github/copilot-cli — the gh copilot stub is not the same tool. Deferred; skip for now unless you want it.

Known incompats

Cross-refs

• agent-skills-monorepo — canonical bundle mechanic
• fleet-cc-opencode-kilo-2026-07-03 — 3-agent baseline this expands
• skills-cli-canonical — npx skills is the sole sync mechanism
• no-card-on-file — hard rule; all 6 providers respect it
• obra/superpowers — upstream