duketopceo/agentflow

git clone https://github.com/duketopceo/agentflow
cd agentflow
bin/agentflow doctor        # which harness CLIs are on PATH

Globals inside a script: `agent()`, `parallel()`, `pipeline()`, `batch()`,
`phase()`, `log()`, `args`. `batch()` fans out through the harness's async
surface — one OR Batch job on `--harness openrouter` (~40–60% cheaper),
one Devin Cloud session per item on `--harness devin` (`swe-2` models are
free), one async sonar request per item on `--harness perplexity`
(grounded, cited answers — the research surface), or first-party batch APIs
on `--harness openai`/`--harness anthropic` (~50% off, 24h window). Results
align to input order; resumable mid-poll.

## Watching a run

`agentflow watch [run-id]` renders a live per-lane view from the run journal —
spinning in-flight lanes, batch poll progress, per-call token counts, orphan
detection — on an alternate screen. Piped output falls back to `tail`-style
event lines; `status <id>` prints the same totals for scripts. Token usage
is journaled per call on harnesses that report it (devin, via `--export`)
and summed in `status`/`watch`.

## Install as an agent plugin

The repo carries a manifest for each harness; all resolve to the shared
`skills/agentflow/` directory.

| Harness | Install |
|---|---|
| Devin | `devin plugins install -y duketopceo/agentflow` (or `--local /path/to/agentflow`) — reads `.devin-plugin/`; auto-installs the compound-engineering plugin via `requiredPlugins` |
| Claude Code | `claude plugin marketplace add duketopceo/agentflow` then `claude plugin install agentflow@agentflow` → `.claude-plugin/` (live-verified) — though inside Claude, prefer the native Workflow tool |
| Codex | `codex plugin marketplace add duketopceo/agentflow` then `codex plugin add agentflow@agentflow` → `.codex-plugin/` (live-verified) |
| OpenCode | v1 `1.18.29+`: `opencode plugin github:duketopceo/agentflow` → `plugins/agentflow/v1.mjs` (live-verified). v2: `opencode plugin github:duketopceo/agentflow#main::path:plugins/agentflow` → `v2.mjs` (no v2 CLI published yet). Both expose `agentflow_run`/`agentflow_doctor`. See `.opencode/INSTALL.md` |
| Factory Droid | `droid plugin marketplace add duketopceo/agentflow` — `.factory-plugin/` native (`.claude-plugin/` also auto-translates); ships the `worker` droid |
| Copilot CLI | `copilot plugin install duketopceo/agentflow` (verified: skill installs); MCP via `copilot mcp add agentflow -- agentflow mcp` |
| Cursor | `.cursor-plugin/` native; submit listing via cursor.directory or cursor.com/marketplace/publish — root `plugin.json` also covers the open standard |
| Gemini CLI | `gemini extensions install https://github.com/duketopceo/agentflow` — verified live: `gemini mcp list` shows `agentflow` connected |
| Any MCP client | `agentflow mcp` — the `.mcp.json` at repo root wires it (Crush, Claude Desktop, Zed, Roo, Kiro…). Tools: `agentflow_run`/`doctor`/`status`/`tail`/`stop` |
| Ori/OpenRouter/Perplexity | no plugin surface — run `agentflow` CLI directly with `--harness ori` / `--harness openrouter` / `--harness perplexity` or `--via ori` |

<sub>Devin, OpenCode v1, Claude Code, Codex, Gemini (extension + MCP server), the universal
skills registry, and Copilot rows are live-verified; Factory/Cursor rows are
manifest-verified (GUI/no-binary installs). agentflow is not published to npm — the plugin manifests above
(or a clone + `npm link`) are the install path.</sub>

## Quick start

No model calls, no cost:

```bash
bin/agentflow run examples/parallel-audit.mjs --harness echo

bin/agentflow run workflows/deep-research.mjs --harness devin -- "your question"
bin/agentflow run workflows/deep-research.mjs --harness openrouter --cache -- "your question"   # needs OPENROUTER_API_KEY
bin/agentflow run my-workflow.mjs --harness claude --via ori -- args   # claude workers billed through OpenRouter

export const meta = { name: "audit", description: "review every file", phases: ["fan-out"] };

phase("fan-out");
const results = await parallel(files.map((f) => () =>
  agent(`Review ${f}`, { label: `review:${f}`, model: "cheap" })
));
const good = results.filter(Boolean);   // a failed call is null, not an exception

flowchart LR
  S[workflow script .mjs] --> R[runtime: pool + journal]
  R -->|agent#40;#41; x N| P{adapter}
  P --> D[devin -p]
  P --> C[claude -p]
  P --> X[codex exec]
  P --> O[opencode run]
  P --> I[ori code]
  P --> H[openrouter HTTP]
  P -.->|--via ori| W[ori wraps the CLI]
  R --> J[(.agentflow/runs/id.jsonl)]
  J -->|--resume| R

bin/agentflow run w.mjs --detach   # background; journal keeps the record
bin/agentflow status <run-id>      # progress; 'orphaned' if the pid died
bin/agentflow stop <run-id>        # SIGTERM the journaled pid (resumable)
bin/agentflow clean [--all]        # reclaim worktrees of non-live runs (force!)
node scripts/smoke-harnesses.mjs   # reproducible live-adapter matrix