dag-workflow

Reviewed at commit bfb0d9a

Caution: Approves DAG workers' permission prompts on a model's judgement (Sonnet, or TypeSafe's Jev with a key) and keeps your prompts and tool output in ~/.claude/dag-workflow.

Sends data to TypeSafe Jev (api.typesafe.ai), only when you set TYPESAFE_API_KEY

Dependency-ordered DAG workflows of Claude Code subagents, with a live DAG pane

Made by
HyunjunJeon
What we checked

Splits large requests into a graph of Claude subagents that run in parallel, verifies each with checks it runs itself and shows progress in a DAG pane. A Sonnet call on your account (or TypeSafe's API, with a key) judges every prompt and settles uncertain worker permissions. It learns skills from your session in the background, keeping notes in .claude/dag and ~/.claude/dag-workflow.

Review details

Static review of the source at the pinned commit; not run. hooks.json loads hooks/register.ts, which registers session.start, classic.SessionStart, session.measure, session.end, session.receive, session.compact, prompt.context, prompt.submit, prompt.edit, prompt.autocomplete, skill.prompt, agent.spawn, turn.start, turn.step, turn.complete, classic.PermissionRequest, tool.check, several tool.call hooks (all tools, Edit/Write, Bash, SubagentHandback and its own mcp__dag-workflow__dag and __board tools), command.run (/dag, /dag-ping), ui.render (Pane, AbovePrompt), ui.message and ui.close. It spawns node workers with $.agent.spawn (general-purpose or the node's agent type, Sonnet or Opus, no permission-mode change), steers them with $.session.send and stops them with TaskStop. 'Jev' decisions (turn mode for every user prompt under the default enforcement auto, node model routing, failure recovery, unattended node questions, and tool permissions) go to https://api.typesafe.ai/v1/systemone with a Bearer TYPESAFE_API_KEY read from $.env when that key is set, otherwise (default jev_model_fallback true) to $.model.complete with Sonnet on the user's own account; requests carry the prompt (up to 4,000 chars), the last three requests, the tail of the last assistant reply, node tasks and goals, or the tool input, passed through a regex secret mask, and are capped by jev_max_calls (100 per session). In tool.check, when the host would ask, a node worker's call (jev_permission_scope dag, default; all extends it to the main conversation; off disables it) is allowed or denied by Jev at confidence >= 0.9 without the person; AskUserQuestion, ExitPlanMode and nodes marked external stay with the person. Node verify checks are argv arrays from the model-written or file-based run definition and run through $.process.run in the run folder (any program, up to 10 min, outside Claude Code's Bash permission prompt); start_approval is off by default, so model-started runs and their checks start without a confirmation. Under strict or a DAG turn the main conversation's non-read-only tools are refused (its stated purpose); a Bash kill-by-name guard refuses pkill-style commands while a run is active; Edit/Write get an anti-slop scan (no model call) and, when semgrep is on PATH, a local semgrep scan with the plugin's own KISA rules and --metrics off that can refuse a hard-coded secret. Session learning (session_learning, default on) records every user prompt, assistant answer and tool input/outcome (masked, capped at 4,000 chars per event) in ~/.claude/dag-workflow/learning/sessions, runs a background Sonnet agent up to 20 times a day, writes learned SKILL.md files to the project's .claude/dag/skills (older ones stay managed in ~/.claude/skills, which it creates), and appends a learned-skill catalog to user prompts and spawned agents' prompts. Run checkpoints, node reports, verification evidence, decisions, context and boards go to <project>/.claude/dag (with a * .gitignore), pruned after 14 days; review notes default to /tmp/dag-review; the trajectory log is off by default. It runs mkdir, mv, rm, rmdir, sleep, sh (a command -v lookup), mktemp and read-only git (ls-files, diff, show). The pane, band, toasts, status line and native notifications are drawn locally. Fully read: plugin.json, hooks.json, register.ts, host/io.ts, host/learning.ts, engine/learning.ts, engine/jev.ts and the semgrep runner; the other engine, host and ui modules by targeted search for network, process, file-write, env and spawn calls. statusline/*.ts, scripts/*, hooks/engine/worktree.ts and isolation.ts are not imported by the loaded code and were not reviewed; skills were grepped (no allowed-tools, no installs).

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): yes
  • Sends data over the network: yes
  • Stores data: yes
  • Shows something on screen: yes
Screenshot of dag-workflow
Made this mod? Add a badge to your README

modsforClaude badge for dag-workflow

Paste this into your README. The badge says reviewed once the current version has been reviewed, and links readers to this page.

[![Listed on modsforClaude](https://modsforclaude.com/badge/dag-workflow.svg)](https://modsforclaude.com/mods/dag-workflow/)

← All mods