Skip to main content
skill CX Codex

maestro-graph — Claude Code and Codex skill

maestro-graph — a skill for Claude Code and Codex that runs a task along a predefined graph: spawns agents on each node, collects results, and moves to the next step.

MIT ★ 238 updated 2026-10-17

Install into Claude Code · Skills CLI (Vercel) All methods and platforms ↓
npx skills add ReinaMacCredy/maestro --skill maestro-graph -a claude-code
Open on GitHub

What maestro-graph does

A graph defines the steps in a markdown file; maestro-graph executes deterministic nodes itself and surfaces only agent and human decision points to you. A node can read the project tree (if its profile is read-only) or write to it and get your review before the next step runs. Write your own graph from the reference, or use built-in presets for code review, fixes, and team consensus.

The flow is sequential: spawn a node, wait for its result, move on. Independent nodes run in parallel. If a node must return a JSON object to a schema, maestro-graph extracts it from the agent's response — if parsing fails, it re-asks twice, then stops. Hitting a limit (node count, loops, fanout depth) halts the run as partial; rerun with a higher limit.

Graphs in the project tree need explicit trust via maestro graph trust; home and shipped graphs run without approval. Every transition is logged; view it with maestro trace.

What it helps with

  • orchestrate multiple agents through a fixed workflow
  • run code review with gates and decision checkpoints
  • drive fixes through a validation loop
  • gather opinions from multiple profiles before deciding
  • replay a graph with different parameters without rewriting logic

Example prompts for the agent

  • Run the review-gate graph on this branch with range=main..feature and tier=full
  • Drive fixes through fix-loop with scope=async and check=npm run test
  • Author your own graph in my-graph.md and execute maestro graph run --file my-graph.md brief="audit the architecture"

How to install maestro-graph

1Skills CLI (Vercel)

npx skills add ReinaMacCredy/maestro --skill maestro-graph -a claude-code

2Manually: put the files in place

~/.claude/skills/maestro-graph/SKILL.md

Copy the whole skill folder; for one project use .claude/skills/maestro-graph/

Step by step, with commands for each platform: How to install skills in Claude Code: five ways

Questions about maestro-graph

How do I install maestro-graph in Claude Code?

Run npx skills add ReinaMacCredy/maestro --skill maestro-graph -a claude-code in your terminal, then open ~/.claude/skills/maestro-graph/SKILL.md.

What is maestro-graph and what does it do?

A skill that automates a sequence of agent actions along a graph: reads each node, spawns agents, collects results, and advances to the next step. Suited for recurring processes such as code review or fix cycles.

How do I write my own graph instead of using a preset?

Create a markdown file describing nodes using the format in references/authoring.md, then run maestro graph run --file <path> [params]. To reuse it later, copy the file to .maestro/graphs or ~/maestro/graphs.

Source text

SKILL.md · MIT
---
name: maestro-graph
description: Drive a pre-known multi-agent path as a maestro graph - run it by name or from a file you just wrote, pull each agent node with graph next, spawn it as a sub-agent under its maestro-<profile> definition, hand the result back with graph result, repeat until the verdict. A sub-agent node reads or writes as its profile allows, and the chat running the graph reviews every diff a writing node leaves; a node whose result another seat must accept is opened as a pane instead. Author a new graph from the reference when no preset fits.
review-date: 2026-11-28
---
<!-- maestro-skill-version: dev -->

# maestro-graph

Use when a task is a pre-known path with several agent steps: a review gate,
a research sweep, a judge panel, a fix loop. The path is one markdown graph
file; maestro holds the run, executes the deterministic nodes itself and
hands you only the agent and human nodes to spawn. maestro never starts a
model (Hub d78), so the loop below is yours on every harness. Design,
diagnosis and the SLP seat protocol stay outside graphs.

## Pick or write the graph

- `maestro graph list` shows every graph across the repo
  (`<repo>/.maestro/graphs`), the room (`~/maestro/graphs`) and the shipped
  set; a nearer file shadows a farther one by name. `maestro graph show
  <name>` prints one.
- A shipped preset fits: run it by name. `review-gate` takes `range=<git
  range>`, `tier=light|full`, `intent=<what the change is for>` and, on a
  re-run, `justified=<file:line reason per line>`; `fix-loop` takes `scope=<what to fix>` and
  `check=<command that must pass>` and drives a writing fixer for at most
  three rounds; `council` takes `brief=<neutral brief>` and
  `tier=lens|debate|debate-with-proof|high-risk` and runs the maestro-council
  protocol with you (the Lead) answering the draft and verdict nodes.
- Nothing fits: write a graph for the task from
  [references/authoring.md](references/authoring.md) and run it with
  `--file <path>`. Keep a good one by copying the file into a graphs
  directory (Hub d100; there is no save verb).

## The pull loop (executor subagent)

```text
maestro graph run <name>|--file <path> [key=value ...] [--limit k=v] --json > run.json
loop:
  envelope = read run.json (or maestro graph next <run> --json > run.json)
  if envelope.done: stop; the verdict, LIMIT stop or failed node is in it
  for each node in envelope.nodes (all at once, they are independent):
    kind human  -> stop and ask the user the prompt; feed the answer back
    kind agent  -> spawn a sub-agent with the node's profile and brief
  for each returned sub-agent:
    maestro graph result <run> <ref> --file <path>|--text "<result>"
  maestro graph next <run> --json
```

- A sub-agent node reads or writes as its profile allows: review lenses
  are read-only, `fixer` and `delegate` write. Review every diff a writing
  node leaves before the next node reads the tree. A node whose result
  another seat must accept is opened as a pane with
  `maestro work add --to <peer>` and its result handed back with
  `graph result` as before.
- `run` returns the first envelope, so the first `next` is implicit.
- Redirect every `--json` envelope to a file and read the fields you need
  out of it (`jq`, or one `python3 -c` line); never let it land inline. An
  envelope carrying several node briefs and their schemas routinely passes
  a harness output cap, and a truncated envelope costs a second read of the
  same bytes before the loop can continue.
- Every node in `nodes` is ready now; spawn them in parallel. Nodes that
  depend on one of them appear on a later `next`, as soon as their own
  inputs are in: only a `join` waits for a whole fan-out (Hub d82).
- `ref` is the node id, or `node@key` for one instance of a foreach.
- A node with a `schema` must return JSON of that shape. Write the
  sub-agent's answer to a file and pass `--file`; maestro extracts the
  first JSON block from prose or a fence. `PARSE_FAILED` with `retry:
  true` means re-ask that sub-agent for JSON matching the schema carried
  in the error (two retries); the third failure marks the node failed and
  the run ends with `failed`.
- `stopped: "LIMIT"` ends the run at a structural limit (`nodes`, `loops`,
  `fanout`, Hub d84) with `partial` state; rerun with `--limit <k>=<N>`
  when the cap, not the graph, was wrong.
- `GRAPH_UNTRUSTED` on a repo graph's function node: review the file the
  error names, then run the `maestro graph trust` command it gives and
  `graph next` again. Home and shipped graphs never ask.
- `maestro trace <run>` is the journal: every node transition and round.

## Spawning an agent node

The profile is a definition `maestro install` rendered for both harnesses
(Hub d83): `~/.claude/agents/maestro-<profile>.md` and
`~/.codex/agents/maestro-<profile>.toml`.

- Claude Code: the `Agent` tool with `subagent_type: "maestro-<profile>"`,
  `model: "opus"` (graph nodes are judgment and writing roles), and the
  node's `brief` verbatim as the task. A profile whose `harness` is codex
  (challenger, verifier, auditor, reviewer-correctness-codex) opens
  as a Codex pane with `--profile maestro-<profile>` instead; the Agent tool
  rejects its model.
- Codex: `spawn_agent` with agent type `maestro-<profile>` and the node's
  `brief` verbatim.

Send the `brief`, never the bare `prompt`: the brief is the prompt plus, when
a schema is declared, one sentence naming the required keys and any optional
keys, then the schema as a JSON block (Hub d843), so the agent answers in the declared
shape instead of its harness habit. It already
carries the run state the graph author placed in it. Add only what the harness needs to return
the answer (for example, "write your JSON answer to <path>"). Never merge
two nodes into one spawn and never run a function node's command yourself;
maestro already did.

## Executor team

`graph run` reports `executor` in every envelope (Hub d88): `subagent`
from a plain session, `claude -p`, `codex exec` or a desktop app; `team`
when the driver is a role pane of a running SLP team. Under `team` the
Lead is the driver and each agent node is one Peer work item (Hub d89);
the Lead is never a node and maestro still spawns nothing.

```text
Team Supervisor: maestro work add "run graph <name> <key=value ...>" \
                   --acceptance "the run's verdict"
Lead:            maestro work take <item>
                 maestro graph run <name> [key=value ...] --json
loop:
  envelope = the JSON just returned (or maestro graph next <run> --json)
  if envelope.done: maestro work return <item> "<verdict JSON>"; stop
  for each node in envelope.nodes without a work field:
    kind human -> answer it yourself: maestro graph result <run> <ref> --text "<answer>"
    kind agent -> maestro work add "<node.brief>" --to peer-<node.profile> \
                    --acceptance "one JSON object matching the schema in the brief" --json
                  maestro graph result <run> <ref> --work <new item id>
  for each node with a retry field (its item's body failed the schema):
    open a fresh item with node.brief and rebind exactly as above; two
    retries, the third failure fails the node
  for each node with a work field whose workState is RETURNED:
    read it (maestro status <item>), then maestro work accept <item>
    (or maestro work note <item> "<gap>" --rework for one retake)
  maestro graph next <run> --json
Team Supervisor: maestro work accept <item>
```

- One pane per profile: `--to peer-<profile>` opens the Peer lazily on the
  first item and sends later nodes of that profile to the same pane; each
  `work add` wakes the pane with `[from lead][<id> OPEN]`, so no hand prompt.
- A bound node stays in `nodes` with `work` and `workState` until its item
  is DONE; `next` then parses the item's returned body like any result
  (schema and all) and issues what depended on it. A body that fails the
  schema unbinds the node and lists it with `retry: {error, schema, work}`,
  twice at most (d843); a cancelled item fails the node.
- Bound nodes count toward `limits.fanout`; keep the fan-out under the
  number of Peers you are willing to open.
- Prompts to a Peer must open with a lowercase plain sentence; a brief that
  opens "You are ..." is swallowed as a slash command. The node prompts in
  the shipped presets already do.
- The graph runtime writes no SLP state: every `work add`, `accept` and
  `return` above is yours (A7).

## Hand-off

The verdict is the run's evidence on its work item (`maestro work show
<run>`); quote it in the return that asked for the gate. A graph that
misbehaved is a finding for the handback, not a card.

© ReinaMacCredy · MIT · ReinaMacCredy/maestro

Similar skills

All skills for Agent setup: memory, context, orchestration →

From the same set10

The whole set →
maestro-questionnaire The maestro-questionnaire skill turns an unresolved decision into a Markdown questionnaire for someone with the missing knowledge. Use it to collect answers asynchronously or in a meeting. CX Codex +1 ★ 238repo stars maestro-coach maestro-coach — a skill for Claude and Codex that unjams a question you got stuck on and teaches the concept behind a decision using your own repo code. CX Codex +1 ★ 238repo stars maestro-bundle maestro-bundle manages SPEC, NOTES, and VERIFY bundle lifecycles and controls context handoffs across agent sessions in Maestro. CX Codex +1 ★ 238repo stars maestro-verify The maestro-verify skill checks acceptance criteria against verification evidence, records a verdict, and guides bundle closure. Use it before handing work back or closing a task. CX Codex +1 ★ 238repo stars maestro-design maestro-design — a skill that resolves blocking unknowns before implementation: architecture choice, modeling, prototyping, and recording durable decisions. CX Codex +1 ★ 238repo stars maestro-work maestro-work is a Claude Code skill that defines shared playbook rules for task briefs, work handbacks, and records. It applies when an agent delegates work or tracks a task in Maestro. CX Codex +1 ★ 238repo stars maestro-council maestro-council runs a structured council over a call you cannot cheaply undo: sealed seats, a premise check, cross-examination and one binding verdict. CX Codex +1 ★ 238repo stars maestro-improve maestro-improve converts filed lessons into minimal doctrine edits. Keeps project rules up to date without losing incident records. CX Codex +1 ★ 238repo stars

Project sponsors

Companies that keep this analytics open