Multi-agent orchestration for Codex App/CLI, Claude Code, OpenCode & Antigravity.
Thinking models plan and review every step. Worker models execute in parallel. Learnings persist in local, per-project memory.
amplify → spec → scope → dispatch → audit → deploy
Start anywhere. Auto-advance forward. Memory persists across sessions.
Landing site · Installation · Providers · Model Routing · Orchestration · Changelog
Not just another orchestrator — three things set Hyperflow apart:
- Every step is reviewed. Worker → Reviewer is an iron rule at every granularity, sub-phases included. No worker output ships unreviewed.
- Memory that's yours. Learnings, decisions, and pitfalls persist in
.hyperflow/memory/— plain markdown, committed with your repo, never uploaded, never mixed across projects. Hot/warm/cold tiering keeps injection cheap. - Depth that adapts. Triage classifies every task and picks a flow profile (fast → scientific), so a 5-line fix never triggers a 300k-token deep run.
- Compaction with a gauge. Automatic context compaction checks estimated transcript usage first, skips early compacts, and snapshots task state only when compaction is actually useful.
Underneath: a structural thinking/worker model split (expensive models plan & review, fast models execute), 15 persona-stitched experts, intent auto-routing, and four auto-detected providers — all local, no daemon.
Start with a rough idea — the pipeline carries it to shipped. Start at any entry point; the orchestrator picks up and runs forward.
| # | Skill | What it does |
|---|---|---|
| 1 | amplify |
Front door — rewrite a rough prompt into the strongest version (persona standards + 8-dim rubric), then hand off into the chain |
| 2 | spec |
Design-first — multi-dimensional analysis + alternatives; refuses to code before you approve |
| 3 | scope |
Decompose the approved design into a parallel task graph |
| 4 | dispatch |
Fan out persona-stitched workers under per-batch + final-integration review |
| 5 | workflow |
Big-task lane — native Claude Code workflows, custom Codex/OpenCode adapter |
| 6 | audit |
L1–L5 review on the result |
| 7 | deploy |
Pre-push gates (lint · typecheck · build · tests · security) → commit → release → push |
amplify hands off to spec, then spec → scope → dispatch auto-chains; audit and deploy are gates that fire at the end. Enter at spec for design-first work, scope when the approach is clear, dispatch when a task file already exists. scaffold is a one-time project setup — run it once per repo to build the .hyperflow/ cache.
workflow is the big-task lane. Hyperflow routes deep/scientific/system-wide work, large migrations, repo-wide audits, and high-confidence verification prompts to /hyperflow:workflow. In Claude Code v2.1.154+, it asks the native dynamic workflow runtime to create a background workflow with research, parallel execution, adversarial verification, quality gates, and final synthesis. In Codex and OpenCode, it runs the same phases through a portable workflow adapter using provider subagents/tasks when available and inline worker/reviewer phases otherwise.
In Codex App/CLI, /hyperflow:* entries are treated as plugin skill aliases, not native host slash commands. If the host does not expose Hyperflow's AskUserQuestion popup UI, required gates still fire as concise Hyperflow Question chat blocks with numbered choices, then Hyperflow waits for your answer. When Codex subagents are available, Hyperflow maps worker/searcher/writer dispatches to them; otherwise those phases run inline and the chain continues in the same thread.
claude plugin marketplace add Mohammed-Abdelhady/hyperflow
claude plugin install hyperflow@hyperflow-marketplaceCodex App/CLI:
codex plugin marketplace add Mohammed-Abdelhady/hyperflow
codex plugin add hyperflow@hyperflow-marketplaceFirst initialize the project (once), then invoke any skill:
/hyperflow:scaffold # first: set up the project (once per repo)
/hyperflow:amplify "make a login page" # turn a rough idea into a strong prompt
/hyperflow:spec "add user auth with login + middleware" # design → scope → dispatch
/hyperflow:workflow "large migration across the repo" # big-task workflow lane
/hyperflow:trace "tests fail after the auth refactor" # root-cause a bug
/hyperflow:deploy # pre-push gates + ship
Codex-safe equivalent:
hyperflow scaffold
hyperflow amplify "make a login page"
hyperflow workflow "large migration across the repo"
hyperflow trace "tests fail after the auth refactor"
Auto-routing is on by default — say "audit the diff" or "debug this test" and the right skill runs without the /hyperflow:* prefix.
Setup, model routing, and per-provider notes → Installation · Providers.
Invoke a skill. Chain-starters auto-advance through the rest — no always-on orchestrator, no background process, everything in your terminal.
The split is structural, not a setting — each tier does only what it's best at.
| Tier | Models | Role |
|---|---|---|
| Thinking | GPT-5.5 · Opus 4.8 · Gemini 3 Pro | Orchestrate, triage, brainstorm, review every output, run the final integration pass |
| Worker | GPT-5.4 fast mode · Sonnet 4.6 · Gemini 3.5 Flash | Execute in parallel — implement, search, write |
Worker → Reviewer is an iron rule. Independent sub-tasks fan out in parallel; each worker feeds its own thinking-tier reviewer before the batch advances. Every non-trivial phase decomposes into named sub-phases (2a, 2b, 2c…), each with its own reviewer. Per-batch reviewers do L1–L2 spot-checks; a final integration reviewer runs once over the cumulative diff. Each reviewer returns a verdict — APPROVE or NEEDS_REVISION (retry once with findings injected).
Every task is classified — complexity, scope, risk, ambiguity — and assigned a flow profile, so effort matches the work instead of always running deep:
| Profile | Use when | Workers | Budget |
|---|---|---|---|
fast |
trivial single-file, reversible | 1 | ≤30k |
standard |
simple/moderate, 2–5 files | 1–2 | ≤100k |
deep |
complex / cross-cutting / system-wide | 3+ | 300k |
research |
unknown territory, evaluation | 3+ searchers | ≤80k |
creative |
UI/UX exploration | 1–2 | ≤150k |
scientific |
correctness-critical, proof work | 2–3 + TDD | 300k |
15 composable expert personas — architect, api, db, frontend, ui, security, performance, scientific, refactor, bugfix, test, research, creative, devops, docs. Each task is tagged and the matching personas are stitched into the worker prompt in priority order: security frames every decision first, creative adapts last.
Hyperflow routes very large tasks to /hyperflow:workflow instead of forcing everything through turn-by-turn dispatch. Use it for system-wide changes, large migrations, repo-wide audits, and verification-heavy work.
In Claude Code, the skill uses native dynamic workflows. This requires Claude Code v2.1.154+ with workflows enabled; workflows can be disabled by /config, managed settings, ~/.claude/settings.json, or CLAUDE_CODE_DISABLE_WORKFLOWS=1. Hyperflow does not set /effort ultracode or xhigh automatically. Use /effort ultracode yourself if you want Claude Code's session-wide automatic workflow selection.
In Codex and OpenCode, the same command runs a custom Hyperflow workflow adapter: research and planning, provider subagents/tasks or inline worker phases, adversarial verification, quality gates, per-task commits, and final synthesis. This is not native Claude-style saved workflow support; repeatability comes from the skill, .hyperflow/tasks/, memory, and provider-specific subagent/task configuration.
Learnings live at .hyperflow/memory/ — plain markdown, committed with your repo, never uploaded, never mixed across projects.
- Three tiers —
hot(≤7 days, always injected),warm(8–30 days, tag-matched),cold(30+ days, on-demand, compressed). - Lazy injection — only tag-matched entries load for a given task, so injection cost stays bounded.
- Auto-written by the chain —
auditrecords recurring findings toanti-patterns.md(hot);specrecords structural answers toproject-decisions.md, so the same questions aren't asked twice.
Full walkthrough → Orchestration · Landing site.
Fifteen skills. Three chain-starters auto-advance through the chain; the rest are standalone. Auto-routing is on by default — say the verb and the right skill runs without the /hyperflow:* prefix. In Codex, hyperflow <skill> is the safest portable spelling, with /hyperflow:* handled as an alias.
| Skill | Command | Type | Purpose |
|---|---|---|---|
spec |
/hyperflow:spec |
Chain starter | Design-first analysis + alternatives; auto-chains to scope → dispatch |
scope |
/hyperflow:scope |
Chain starter | Decompose into parallel worker subtasks; auto-chains to dispatch |
dispatch |
/hyperflow:dispatch |
Endpoint | Fan out persona-stitched workers under per-batch + final review |
workflow |
/hyperflow:workflow |
Big-task lane | Native Claude Code workflows; custom Codex/OpenCode adapter for migrations, audits, and verification-heavy work |
scaffold |
/hyperflow:scaffold |
Standalone | Project setup — .hyperflow/ cache + multi-tool shims |
amplify |
/hyperflow:amplify |
Front door | Rewrite a rough prompt into the strongest version (persona standards + 8-dim rubric), then hand off into the chain |
trace |
/hyperflow:trace |
Standalone | Systematic root-cause debugging — 5 Whys, never patches symptoms |
audit |
/hyperflow:audit |
Standalone | L1 quick → L5 exhaustive review on changes, files, or PRs |
deploy |
/hyperflow:deploy |
Standalone | Pre-push gates → commit → release → push (push always asks) |
cache |
/hyperflow:cache |
Standalone | Memory CRUD — show, search, add, prune, archive, compact |
status |
/hyperflow:status |
Standalone | Read-only snapshot — version, memory count, live per-task progress |
background |
/hyperflow:background |
Standalone | List, show, cancel, prune task-level background agents |
sticky |
/hyperflow:sticky |
Standalone | on / auto / off — per-project auto-routing mode |
bridge |
/hyperflow:bridge |
Standalone | Embed the portable doctrine into CLAUDE.md for Desktop / web / IDE |
flush |
/hyperflow:flush |
Standalone | Flush a deferred-commit queue from a prior or crashed chain |
| Provider | Thinking | Worker |
|---|---|---|
| Codex App/CLI | GPT-5.5 (adaptive reasoning) | GPT-5.4 (fast mode) |
| Claude Code | Opus 4.8 | Sonnet 4.6 |
| OpenCode | Claude Opus 4.8 | Sonnet 4.6 |
| Antigravity | Gemini 3 Pro | Gemini 3.5 Flash |
Auto-detected at session start. Override in ~/.hyperflow/config.json. See Model Routing.
- Landing site — the full overview
- Installation · Providers · Model Routing · Orchestration
- Changelog · Privacy · contributor guide in
CLAUDE.md
MIT