Run a Saved Workflow

A saved workflow keeps its guarantees in code or in a generated plan, not in a prompt: fresh agents each round, the refuter quorum, the frozen acceptance checks, the round cap. This skill finds the workflow and hands off to that code or plan. It never improvises a workflow of its own.

Quick Reference

Workflow

1. Find the library. Installs are symlinks, so resolve this skill's real path first. agents/workflows/ sits three levels above the real skill folder (skills/universal/run-workflow/):

   lib="$(cd "$(dirname "$(realpath "<this-skill-dir>/SKILL.md")")/../../.." && pwd)"
   ls "$lib"/agents/workflows/*.manifest.json

2. List and pick. For each manifest, show the id, workflow.engine and workflow.description. If the user named an id with no manifest, say so and list the ids that exist. Take the args from workflow.whenToUse. Ask for a missing required arg, such as the frozen spec and acceptance checks of a build loop, instead of inventing one.
3. Claude Code: hand off. Reply with the one command to type: /<id> followed by the args object whenToUse describes. Do not re-implement the workflow inline, because an imitation drops the checks the script enforces in code. If /<id> is not found, the library sync has not linked agents/workflows/<id>.js into the Claude workflows folder; say so and stop.
4. Codex: execute the plan from the parent session.
   • If the plan has a context_budget block, follow its extract command instead of loading the whole file.
   • If the plan has a tree_snapshot, run its command once before the first worker, and apply its stop_rules after each read-only phase. A rule that fires ends the run BLOCKED or INCOMPLETE with the rule's stop reason.
   • Run the execution_contract phases in order. Each phase names its owner, its primitive (spawn_agent, wait_agent, send_input, close_agent, local reasoning or an immediate return) and its gate. This session does every dispatch, wait, merge and verdict; workers never spawn workers.
   • Brief each worker on its own: the plan's task_name and fork_turns, its rules, and the user's args. Keep to workspace: one writer at a time.
   • Keep to the plan's agent cap. Spawn only the members, refuter count, max_dynamic_members and rounds (maxRounds, else max_rounds_default) it allows. Run no more at once than the session's thread limit, and queue the rest.
   • Apply incomplete_rule: a failed, timed-out or malformed member makes the verdict INCOMPLETE. Name that member. Do not rerun it silently or do its part yourself.
   • Close every worker you spawned before you report.
5. Report. Give the verdict, the stop reason, the open items or escalation, and any member that failed. When a fixer or builder ran, list its changes from git status --short and say that nothing was committed.

Gotchas

• A .js file in agents/workflows/ with no manifest still runs on Claude Code as /<id>, but it has no Codex plan. Do not offer it on Codex.
• A plan's stop rule is part of the contract. Stopping early at a "good enough" round, or running past the cap, breaks the guarantee the user asked for.

Navigation

• agents-subagents - choose between a saved workflow, subagents and a team; Claude saved-workflow behaviour is in its workflow runtime reference.
• agents/workflows/ - the manifests, Claude scripts and generated Codex plans.

Learnings Loop

When prior decisions or pitfalls are relevant, consult learnings.consolidated.md if present; use learnings.md only for needed history or as the available fallback. Otherwise skip both.

After applying it, if you encountered a pattern worth remembering, a mistake worth preventing, or a domain fact that surprised you, append one dated bullet to learnings.md via agents-skills-feedback-loop/scripts/append_learning.py. Do not modify SKILL.md itself.