Index of every page on this site: llms.txt
All skills
Run it yourself

Orchestrate several coding tasks or PR reviews in parallel from one Herdr pane or the Herdr app, with any coding agent (Claude Code, Codex, OpenCode, Pi, Cursor or any other CLI agent Herdr supports). Each issue gets its own git worktree, a background Herdr tab and a worker agent; each open PR gets a review worker in a tab of its own. Workers can be different kinds in one run. They stop at gates (spec, implemented, pr), show their gate in the Herdr sidebar and report back to this orchestrator session, which relays the user's decisions. A run is saved to disk, so workers can be resumed with their full conversation later. Use when the user wants to work on multiple issues or tasks at once, e.g. "work on 101 and 102 in parallel", "spin up workers for these issues", "smith 101 102 103", "/smith", "smith 5737 5760" with PR numbers to review them, "use codex workers", "pick N issues and run them in parallel", or "resume yesterday's smith run". Requires Herdr (HERDR_ENV=1) or the Herdr app (hc).

Use this Skill: https://skilld.dev/gh/aslamdoctor/skills/smith

Nothing lands on disk. Nothing to clean up.

Fork this Skill

Edit a local copy. It keeps the author and licence.

README.md

≈1.9k tokens on demand. Your agent reads this file only when SKILL.md points to it.

Smith

An agent skill for working on several coding tasks in parallel from one Herdr pane or from the Herdr app.

You tell your agent which issues to work on. It becomes the orchestrator: each task gets its own git worktree, a background tab and a worker agent. Workers stop at gates (spec, implemented, PR) and report back. The orchestrator brings you their questions with a recommendation and relays your answers, so nothing is shipped or posted without your OK.

/smith 101 102 103

What you get

  • Each task runs on its own branch in ~/<repo>-wt/<branch>, so workers never touch each other's files or your main checkout.
  • Workers write a spec and stop. After you approve it, they implement, commit and stop again. They ship only when you say so.
  • A watcher per worker wakes the orchestrator when a gate is reached, a dialog is waiting, or a worker stops. A Herdr toast and sound tell you when you're needed.
  • When every worker reaches its last gate, the orchestrator asks once whether to close their tabs, their watchers and its own tab. One yes closes them all, the orchestrator tab last, once its reply is done. Worktrees and branches stay for review fixes.
  • Workers are Smith's copies, named after their task type and task: code-<id> for an issue and review-<id> for a PR review. Each worker pane shows its name and gate (code-102 · spec) in the Herdr sidebar. The orchestrator's tab is SMITH-ORC<n>, after its terminal address (e.g. SMITH-ORC7 for t7).
  • Workers can be Claude Code, Codex, Antigravity, Grok, OpenCode, Pi, Copilot, Cursor or any other agent Herdr supports. Other CLI agents work through a custom-command mode, and one run can mix agent kinds.
  • A run is saved to ~/.smith/runs/. Workers can be restarted later in their original conversation, for example to handle review comments the next day.
  • A task that depends on another task's unmerged branch is cut from that branch and shipped as a stacked PR.
  • Base branch, branch naming, commit format, ship flow and test commands come from the repo's instruction files (AGENTS.md, CLAUDE.md and similar). You confirm them once per run.
  • In the Herdr app, workers open as tabs in the orchestrator's space and show up once in the sidebar, nested under the orchestrator. Gate messages are typed straight into the orchestrator's terminal.

Requirements

  • One of these, with the orchestrator running inside it:
    • Herdr 0.9 or newer (the TUI), in a Herdr pane
    • The Herdr app 0.1.16 or newer, with its hc and herdr-app commands on your PATH
  • git, jq and bash (macOS's bash 3.2 is fine)
  • gh if you want the orchestrator to pick issues or open PRs
  • At least one coding agent CLI. Install its Herdr integration (herdr integration install <name>) to get session resume

Install

With the skills CLI:

npx skills add aslamdoctor/skills --skill smith

Or clone the skills repo and symlink this folder into each agent's skills directory:

git clone https://github.com/aslamdoctor/skills ~/skills
ln -s ~/skills/skills/smith ~/.claude/skills/smith   # Claude Code
ln -s ~/skills/skills/smith ~/.agents/skills/smith   # shared location (Codex, OpenCode, Pi and others)
ln -s ~/skills/skills/smith ~/.codex/skills/smith    # Codex

For an agent without skill support, tell it: "Read ~/skills/skills/smith/SKILL.md and follow it."

Usage

Start your agent inside the repo, in a Herdr pane or a Herdr app terminal, then ask for parallel work. Smith picks the backend from where it runs, and asks once if both are available:

  • /smith 101 102 103
  • "Work on 101 and 102 in parallel"
  • "Pick 3 open issues that don't depend on each other and run them in parallel"
  • "Use codex workers for these issues"
  • "Resume yesterday's smith run"

A typical run:

  1. The orchestrator detects the project setup and shows it as a table to confirm.
  2. It spawns one worker per task, then sends each a kickoff prompt.
  3. Each worker writes a spec and stops. You get the approach and its open questions, with a recommendation for each.
  4. After you approve, the worker implements, runs lint and tests, commits and stops.
  5. When you say ship, the worker pushes and opens the PR using your project's ship flow. Review-bot replies and comments come to you for approval before they're posted.

Agent support

Tier Agents What works
A claude, codex, agy, grok, pi, opencode, copilot, cursor, devin, droid, kimi, qodercli, qwen, letta, hermes, mastracode, kilo, omp, with the Herdr integration installed Everything, including resume with full context
B The same agents without the integration, plus amp, kiro, maki, muse Start, prompt, live state and watchers. Resume falls back to a catch-up prompt
C gemini, cline Like B, but Herdr's state detection is less tested
D Any CLI agent Herdr doesn't recognize (--cmd) Prompts are pasted into the pane. Progress is tracked through gate lines only

The orchestrator can be any of these too. An agent that can run a background command and gets woken when it exits (Claude Code, for example) uses that. Other agents are woken by a Herdr prompt from a detached watcher.

In the Herdr app, workers are hc terminals and fewer kinds are checked so far:

Tier Agents What works
A claude, agy Everything, including resume with full context
B pi Start, gates and watchers. Resume falls back to a catch-up prompt
C Everything else Untested. Add a row to scripts/app/launch.tsv and try it before relying on it

How it works

SKILL.md                 orchestrator instructions
references/herdr.md      Herdr steps: spawn, send, watch, board, resume
references/herdr-app.md  the same steps for the Herdr app
references/prompts.md    kickoff, approval, ship and relay templates for workers
references/gotchas.md    known problems from real runs and their fixes
scripts/spawn.sh         worktree + background tab + worker, recorded in run.json
scripts/send.sh          prompt a worker and confirm it was delivered
scripts/mark.sh          workers call this at each gate (gate log, run.json, sidebar label or app message)
scripts/watch.sh         wait until a worker needs the orchestrator, then report or wake it
scripts/board.sh         status table for the whole run (both backends)
scripts/done.sh          is every worker at its last gate? the orchestrator then offers cleanup once
scripts/close.sh         close worker tabs and stop their watchers (keeps worktrees and run.json); --self closes the orchestrator tab too
scripts/close-self.sh    close the orchestrator tab once its turn ends (started by close.sh --self)
scripts/agents.tsv       per-agent integration, resume arguments and transcript paths
scripts/app/             Herdr app versions: preflight, spawn, send, watch, close, launch.tsv

Each run keeps its state in ~/.smith/runs/<repo>-<date>/ (<repo>-<date>-app/ for the Herdr app): run.json holds one record per task (branch, worktree, tab, pane, agent kind, session ID, gate, PR), and gates/<id>.md is the gate log.

Notes

  • Worktrees share whatever the main checkout shares. Tests that use one database can collide when workers run them at the same time, and a local dev site usually points at the main checkout, not at a worktree. See references/gotchas.md.
  • New worktree paths can trigger first-run dialogs in some agents (folder trust, login, model selection). The orchestrator brings these to you instead of answering them.

License

MIT

Source: SKILL.md on GitHub

skilld matched fixed text patterns in SKILL.md and file names. Patterns miss obfuscated code.

skilld run checks every file with the same patterns. It asks for approval before it loads a Skill with a behavior marked Needs approval.

1 warning1mo3 checks · Risk SAFE
  • Gen Agent Trust Hub1mo

    This skill acts as an orchestrator for parallel coding tasks by managing multiple worker agents. It presents an indirect prompt injection surface as it reads external data from GitHub and worker outputs, but it implements mandatory user approval gates for all code-writing and shipping actions.

  • Socket1mo

    No alerts

  • Snyk1mo

    Risk: MEDIUM · 1 issue

Signed by skilld at c123387. This ties the file your Agent reads to that commit on GitHub. It does not review the instructions.

Last checked against GitHub 2 days ago.

Activeupdated 4 days ago

README badge

README badge for aslamdoctor/skills/smith