superpowers-v
Reviewed
at commit f8ba7a4
New version, not yet reviewed
Sends data to OpenAI Codex (through your own codex CLI), only when you pick Codex as the reviewer
Compound V for Superpowers: triple parallel pre-flight (code archaeology + domain-expert advisor + library/doc validator — Context7 MCP when the session has…
0 upvotes
Log in to upvote or save- Made by
- Oleg
- Pinned commit
-
f8ba7a4fd23975b04bc383c1fa9dcc6bd3edf559 - What we checked
-
Adds a brainstorm, build and verify workflow: hooks nudge change requests toward its skills, refuse to stop once while a run is unverified, and block direct edits to its run records, while a band above the prompt shows the run's stage, checks, cost and account quota. It keeps run records in docs/compound-v in your project, reads Claude's session transcripts for cost, and sends code to Codex only if you choose that reviewer.
Review details
Static review of the source at the pinned commit; not run. Delta review of the changes since 8f9990b; the rest carries over from the review at that commit. Version 4.0.0 is a near-total rewrite of the previously reviewed version (most v3 commands, agents and scripts removed; new Python package scripts/cvcore). Shell command hooks (Stop, UserPromptSubmit, SessionStart, PreToolUse on Write|Edit|MultiEdit|NotebookEdit, SubagentStop) each run python3 -I -B scripts/compound-v.py hook <name>, wrapped in || true and failing open. The prompt hook classifies the prompt text with regexes (no text stored; triage.log in docs/compound-v/runs logs only its length and the classification) and adds context telling Claude to use compound-v:brainstorm or compound-v:verify, at most 3 times per session, or a one-time setup notice when .claude/compound-v.json is missing. The Stop hook returns decision block once per tree and stage when an open run is not verified, when code changed before the plan was attached, or when code changed after a change request without a run. The PreToolUse guard denies Write/Edit to docs/compound-v/runs/. SessionStart adds a brief run status and runs its own compound-v-onboard.py staleness check. Small state files (tree ids, nudge counters, cost baseline, a digest) go to a compound-v-<uid> folder in the system temp directory; an HMAC receipt key and lock files go in the repository's .git directory. The module run-band.tsx registers session.start (a 5 s $.clock.every tick that only stats docs/compound-v/runs and, when it changed or every 30 s, runs python3 compound-v.py hud via $.process.run) and ui.render AbovePrompt; it uses $.session.id, $.session.cwd, $.session.usage (account rate limits), $.store (quota baseline), $.fs.stat and $.env.get (CV_DISABLED_HOOKS=run-band turns it off). The CLI runs git (diff, ls-files, write-tree with a temporary index, archive into a temp checkout), hashes ignored .env files to bind receipts (contents not stored), reads the tail of the session transcript for Claude Code's cost-state line, and reads subagent transcripts under ~/.claude/projects/<project> to confirm a reviewer agent ran. Commands the skills have Claude run through Bash (so with the normal permission prompt) execute the plan's test_cmd, live_cmd and setup_cmd under a process-group timeout supervisor, merge worktree lane branches with git merge --no-ff and commit (lane finish), and, with the Codex reviewer, run codex exec -s workspace-write (network off) on a temporary checkout, sending the code, plan and request to OpenAI through the user's codex CLI. Skills declare no allowed-tools; builder and reviewer agents are Opus subagents (up to 200 and 25 turns) that use tokens. The init skill writes .claude/compound-v.json and worktree.baseRef in .claude/settings.json only after showing a diff and getting a yes, and offers (opt-in, disclosed) sandbox excludedCommands with matching Bash allow rules for Playwright and the live receipt, disabling Superpowers for the project, and claude mcp add for a pinned @playwright/[email protected]. The onboard skill writes docs, .claude/rules and, only on a confirmed diff, .mcp.json entries that use npx -y with unpinned packages, and only previews npx autoskills --dry-run behind a confirm. design-lint runs the pinned npx --yes @google/[email protected]. The pr-review skill posts comments with gh or glab only on an explicit instruction. External CLIs (codex, gh, glab, npx packages) were not reviewed. What changed in 4.0.1: the run cost now also covers API responses after the CLI's last cost-state entry; scripts/cvcore/usage.py memory-maps the session transcript and reads the main thread and subagents/**/agent-*.jsonl transcripts next to it (local files, read-only) and prices those responses from a new local scripts/cvcore/prices.json (its pricing-page URL is a citation only, never fetched), showing the figure as an estimate (~$). hooks.py binds a session's run on SubagentStop and the hook states are kept in the same temp folder; receipts.py also accepts a background reviewer subagent that ended with a handback result; run.py can reopen a run the Stop hook closed early. run-band.tsx adds a session.measure handler that updates the account quota (rate limit) baselines per run in $.store (quota:<run>, newest 20 kept via $.store.keys and delete) and layout tweaks. Skill text for build and verify now says to label the cost an estimate. No new network calls, hosts, programs, permissions or allowed-tools; the hook list and the skills' behaviour are otherwise unchanged.
- What it uses
-
- Reads your prompts: yes
- Sees tool calls: yes
- Blocks or changes tool calls: yes
- Reads or writes files: yes
- Runs programs on your machine: yes
- Calls the model (uses tokens): no
- Sends data over the network: yes
- Stores data: yes
- Shows something on screen: yes
Compound V for Superpowers: triple parallel pre-flight (code archaeology + domain-expert advisor + library/doc validator — Context7 MCP when the session has it attached, WebSearch/WebFetch otherwise), disjoint file partitioning, manifest-driven multi-backend dispatch (Claude + headless Codex + Antigravity + Cursor workers), git-diff scope enforcement, crash-resumable runs, adaptive tier-based routing, epic mode for multi-feature builds, V-memory local-first semantic+lexical recall over docs/superpowers (opt-in pure-python embeddings + a deterministic recall→action bridge), and batched parallel dispatch (Opus default, Sonnet for narrow junior-task carve-out), plus /v:onboard — a project-onboarding command that builds a citation-verified knowledge base + AGENTS.md/CLAUDE.md bridge behind a human gate. Auto-intercepts three transitions: brainstorming → writing-plans → execution; a gated pre-brainstorm recon pass and batched elicitation are guidance the brainstorm follows, not enforced mechanisms.
Made this mod? Add a badge to your README
Paste this into your README. The badge says reviewed once the current version has been reviewed, and links readers to this page.
[](https://modsforclaude.com/mods/superpowers-v/)