Run a first task

This example runs the built-in software-change graph in a local folder, with Git delivery disabled. The template normally creates a PR; this example explicitly uses --delivery none. Git is optional. Add --cwd FOLDER to select another folder; input files stay relative to the invoking folder.

The worker edits the selected folder

Back up work you want to protect, or commit or stash it in a Git checkout.

1. Write the input

The template expects an object with a task string:

input.json

{
  "task": "Add JSON output to the status command and cover it with focused tests."
}

2. Choose the runtime

A uniform runtime applies one agent configuration to every agent node. Replace the model placeholder with an identifier accepted by the selected provider.

runtime.json

{
  "harness": "codex",
  "provider": "openai",
  "model": "YOUR_MODEL_ID",
  "effort": "high"
}

Zeroshot passes model identifiers to the provider unchanged and does not keep its own model catalog. For this local run, Zeroshot reuses the installed Codex CLI's login and provider configuration; no placeholder API key or openai connection is required. A Docker or Platform run instead derives the canonical OPENAI_API_KEY requirement and resolves it from the submission environment or the target's connection store because it cannot inherit host login state.

For another harness, use its file instead. A local run reuses that CLI's own login in the same way.

runtime.json for Claude Code

{
  "harness": "claude",
  "provider": "anthropic",
  "model": "YOUR_MODEL_ID",
  "effort": "high"
}

runtime.json for GitHub Copilot

{
  "harness": "copilot",
  "provider": "github",
  "model": "YOUR_MODEL_ID"
}

runtime.json for Pi

{
  "harness": "pi",
  "provider": "anthropic",
  "model": "YOUR_MODEL_ID",
  "effort": "high"
}

The RuntimePlan reference lists every accepted pair, including the OpenRouter, Bedrock, and gateway lanes that need a connection.

3. Check the request without starting it

--validate-only materializes the template and uniform runtime, then admits the request with its initial input. It does not contact a target or start a run.

zeroshot run \
  --title "Add JSON status output" \
  --template software-change --delivery none \
  --input input.json \
  --uniform-runtime-config runtime.json \
  --validate-only

Successful validation writes {"valid":true}. A rejected request names the input, graph, or runtime problem; the selected target checks connection availability later, when it receives the submission.

4. Start the run

Remove --validate-only:

zeroshot run \
  --title "Add JSON status output" \
  --template software-change --delivery none \
  --input input.json \
  --uniform-runtime-config runtime.json

The foreground command streams newline-delimited JSON while the run advances. Ctrl-C closes that stream without stopping the controller; another terminal can read the durable state:

zeroshot list
zeroshot status RUN_ID
zeroshot logs RUN_ID

For a graph view with node transcripts, open zeroshot ui and select the run in Runs.

Normal completion happens when the graph reaches a terminal node. An explicit zeroshot force-stop RUN_ID also closes the run, while engine or target failures can produce a terminal failure result. zeroshot logs RUN_ID keeps the error records that explain a failure, such as a failed node's error or the runtime cause. A failed foreground run exits with status 1 and its error lists the latest of those records, followed by the zeroshot logs command for the rest.

Try the smaller template

single-worker has one agent node and no independent review loop:

zeroshot template show single-worker

zeroshot template show software-change --delivery none prints the exact graph behind the first example as the same GraphSpec accepted by --graph and the Python SDK.

zeroshot template show auto-research prints the research graph with its default of ten iterations. Supply a positive options.iterations integer in the run input to set the admitted loop bound. It keeps adopted, record-only, and aborted results under .zeroshot/research; add --push to a run when every finalized iteration should be published.