Hyperclaude — independently scanned and version-tracked by SaferSkills.
SaferSkills independently audited Hyperclaude (Plugin) and scored it 92/100 (green). The audit ran 55 deterministic rules across Security, Supply Chain, Maintenance, Transparency, and Community; it found 0 high-severity and 2 lower-severity findings. The full rule-by-rule trace and per-finding evidence are below. Free, methodology-open.
Findings & checks · 2 flagged
The text {match} tells the agent to skip the normal "ask the user first" gate. Used adversarially it removes the human-in-the-loop check before destructive or sensitive actions, turning a normally-gated agent into a fire-and-forget executor.
The text {match} tells the agent to skip the normal "ask the user first" gate. Used adversarially it removes the human-in-the-loop check before destructive or sensitive actions, turning a normally-gated agent into a fire-and-forget executor.
Every scanned point with the score it earned and what moved between them.
Score fell 8 points between these scans.
The primary manifest — the file an agent reads to learn what this artifact does.
Push Claude Code beyond stock. Claude builds, Codex critiques. A gated research → plan → review → ship pipeline, with autonomous multi-agent revise loops that self-converge.
Claude builds, Codex critiques
A Claude Code plugin built around a deliberate division of labor between two AI coding agents:
reviewer agent when you ask or Codex usage runs ahead of its rate-limit windowThesis: Claude is the builder, Codex is the critic. Better software with a smarter cost split.
┌─ refine ─┐ ┌──── fix ───┐ ┌──── fix ───┐
▼ │ ▼ │ ▼ │
research → plan → plan-review → implement → code-review → docs-sync → docs-review → ship
│ │ │ │ │ │ │ │
Codex+Claude Claude Codex Claude(+agents) Codex Claude Codex userWhen the idea itself is vague (not just un-planned), an optional hyper-interview front-end clarifies it into a spec before research / plan — a short one-question-at-a-time interview, Claude-only (no Codex; clarity is its job, review happens downstream). The refine / fix arcs are what hyper-plan-loop, hyper-implement-loop, and hyper-docs-loop automate — a persistent Claude-side agent (planner / fixer / documenter) revises while Codex reviews (or Claude's reviewer agent, as above), looping until no blocking findings remain. Gates write trace artifacts under .hyperclaude/ (gitignore-friendly); hyper-docs-sync, hyper-docs-loop, and hyper-implement edit the working tree directly. Skip any step a small change doesn't need — only code-review is non-negotiable for behavioral changes. See docs/workflow.md for triggers, skip rules, slug/artifact conventions, and --resume.
hyper-autoOne gesture, end-to-end:
/hyperclaude:hyper-auto add OAuth login to the APIhyper-auto chains hyper-plan-loop → hyper-implement-loop. Claude plans, Codex critiques the plan until no blockers remain, Claude implements, Codex code-reviews until no blocking findings remain (style/nits are reported, never gating) — all hands-off, with each loop seating its own reviewer as above. A clean composed exit closes with an auto-run hyper-recap write-up (Claude-only, no Codex). It's not a new layer, just composition over the two loops, so the same gates and artifacts apply.
User in Claude Code
│
┌───────────┴────────────┐
│ │
Skills ─────────────────► Agents
hyper-setup (invoke-only), planner / implementer
gates, orchestrators, verifier / documenter
autonomous loops, auto researcher / fixer / reviewer
│
▼
codex-bridge.mjs
(only Codex-spawning code;
always read-only sandbox)
│
▼
.hyperclaude/{specs,research,plans,epics,
plan-reviews,code-reviews,docs-reviews,recaps}/
Hooks — SessionStart reminder, fires independentlyThree layers — Skills (the invoke-only hyper-setup doctor plus description-triggered gates, orchestrators, autonomous loops, and hyper-auto), Agents (Claude implementation arm), Hooks (SessionStart reminder). Every Codex spawn — fresh or --resume, with live web search enabled — runs in a read-only sandbox; Codex is critic, never editor. The bridge accepts two optional model-selection flags on all four modes: --model <name> (any non-empty model identifier, e.g. openai/gpt-5) and --effort <low|medium|high|xhigh|max|ultra> (reasoning effort — the levels the Codex model catalog offers, though not every model takes all six; none and minimal are rejected); both default to null, inheriting ~/.codex/config.toml — and, where that file names no model, whatever the installed codex-cli's bundled default currently is, which changes between codex releases; pin model there if the reviewer must stay fixed across upgrades. Each artifact records the model that actually resolved as codex-model-effective when the bridge can resolve one, and --resume auto will not continue a thread that ran under a different one. code-review additionally accepts --background "<text>": a short, strictly descriptive change context (what the change is / what it touches / author intent) passed to the Codex critic to orient the review; it does not alter the review rubric. --background is optional and a no-op when omitted; it is mutually exclusive with --resume (resumed sessions already carry context in the Codex thread). The hyper-code-review skill composes and passes a background automatically; direct CLI callers may pass it explicitly. Both review modes (plan-review and code-review) additionally accept --review-brief "<text>": a caller-composed summary of the user's stated requirements and approved decisions so Codex stops flagging the user's own approved asks as scope creep — unlike --background it is allowed with --resume and passed on resumed rounds too, whenever the caller has an admissible source. These are bridge-level flags; v1 is not exposed via slash commands. See docs/architecture.md for layer details, bridge internals, and the sandbox flag matrix.
External dependencies: Claude Code plugin runtime, codex-cli >= 0.130.0 with the global --search flag, Node 18+, and git. Nothing else (no npm bin, no tmux, no MCP servers).
/plugin marketplace add zeikar/hyperclaude
/plugin install hyperclaude /hyperclaude:hyper-setupChecks Node, codex-cli, codex --search, and git, and reports whether Claude Code is at >= 2.1.232 — the known-good floor for the transport the loops and hyper-auto use. Report-only; nothing is installed automatically, and nothing gates on the version.
/hyperclaude:hyper-research add OAuth login to the API # Codex+Claude prior-art / pitfalls
/hyperclaude:hyper-plan # Claude writes .hyperclaude/plans/<slug>.md
/hyperclaude:hyper-plan-review # Codex critiques the plan
/hyperclaude:hyper-implement # Claude executes the plan task-by-task
/hyperclaude:hyper-code-review # Codex reviews the diff (branch vs main)
/hyperclaude:hyper-docs-sync uncommitted # Claude updates docs for the change
/hyperclaude:hyper-docs-review # Codex accuracy gate on docs
# Or let the loops self-converge:
/hyperclaude:hyper-plan-loop add OAuth login to the API # plan → review → revise, looped
/hyperclaude:hyper-implement-loop <plan path> # implement → code-review → fix, looped
/hyperclaude:hyper-docs-loop # docs → review → fix, looped (default: docs/)
/hyperclaude:hyper-auto add OAuth login to the API # plan-loop → implement-loop, end-to-end
# On-demand, outside the cycle:
/hyperclaude:hyper-memory # mine .hyperclaude/ artifacts for repo-local knowledge candidates
/hyperclaude:hyper-recap # write a human-readable recap of the last completed cycle (also auto-run by hyper-auto on a clean exit)--resume.node --test tests/*.mjs # unit tests for the bridge and setup-doctor
bash scripts/test/smoke.sh # acceptance smoke checksZero npm dependencies. Node 18+ stdlib only.
Structural inspiration from:
No code ported from either; references only.
~30 seconds. Free. No account. Every finding cites a rule and a line of evidence.