AI agent native DevOps bash script orchestrator.

ShellFlow is a minimal shell script orchestrator for mixed local and remote execution. You write one shell script, mark execution boundaries with comments, and ShellFlow runs each block in order while resolving remote targets from your SSH configuration.

• Split a shell script into @LOCAL and @REMOTE execution blocks.
• Run each block fail-fast, in order.
• Reuse the shared prelude before the first marker for every block.
• Pass the previous block output forward as SHELLFLOW_LAST_OUTPUT.
• Export named scalar values from a block into later block environments.
• Emit either a final JSON report or streaming JSON Lines events for agents.
• Support bounded @TIMEOUT and @RETRY directives without embedding workflow logic.
• Provide non-interactive, dry-run, and audit-log modes for automated execution.
• Resolve remote targets from ~/.ssh/config or a custom SSH config path.

uv tool install shellflow

shellflow run playbooks/hello.sh

uv tool install shellflow
shellflow --version

npx skills add longcipher/shellflow

This installs the agent skill from this repository for writing and reviewing Shellflow playbooks.

To upgrade to the latest version:

uv tool upgrade shellflow

git clone https://github.com/longcipher/shellflow.git
cd shellflow
uv sync --all-groups # uv sync --refresh --reinstall --no-cache

uv tool install --force .
shellflow --version

uv pip install -e .
shellflow --version

Shellflow recognizes two markers:

• # @LOCAL
• # @REMOTE <ssh-host>

Shellflow also recognizes bounded block directives at the top of a block body:

• # @TIMEOUT <seconds>
• # @RETRY <count>
• # @EXPORT NAME=stdout|stderr|output|exit_code

<ssh-host> must match a Host entry in your SSH config. Shellflow then connects using that SSH host definition, which means the actual machine can be resolved through the configured HostName, User, Port, and IdentityFile values.

Example:

#!/bin/bash
set -euo pipefail

# @LOCAL
# @EXPORT VERSION=stdout
echo "runs locally"

# @REMOTE sui
uname -a

# @LOCAL
echo "remote output: $SHELLFLOW_LAST_OUTPUT"
echo "version = $VERSION"

Example ~/.ssh/config entry:

Host sui
    HostName 192.168.1.100
    User deploy
    Port 22
    IdentityFile ~/.ssh/id_ed25519

With that config, this block is valid:

# @REMOTE sui
hostname

This is intentional:

• Shellflow accepts configured SSH host names, not arbitrary free-form targets.
• Unknown remote targets fail early with a clear error before spawning ssh.
• You can override the default config path with --ssh-config.

Each block runs in a fresh shell.

• Shell options from the prelude are copied into every block.
• Shell state like cd, shell variables, aliases, and export commands does not persist across blocks.
• Explicit context values are passed forward through environment variables.

Example:

# @LOCAL
echo "build-123"

# @LOCAL
echo "last output = $SHELLFLOW_LAST_OUTPUT"

Named exports are additive to SHELLFLOW_LAST_OUTPUT:

# @LOCAL
# @EXPORT VERSION=stdout
echo "2026.03.15"

# @REMOTE sui
echo "deploying $VERSION"
echo "last output = $SHELLFLOW_LAST_OUTPUT"

Lines before the first marker are treated as a shared prelude and prepended to every executable block:

#!/bin/bash
set -euo pipefail

# @LOCAL
echo "prelude is active"

# @REMOTE sui
echo "prelude is also active here"

Shellflow is designed to be the execution substrate for an outer agent, not an embedded planner.

• Use --json when you want one final machine-readable run report.
• Use --jsonl when you want ordered event records while the script runs.
• Use --no-input for CI or agent runs where interactive prompts must fail deterministically.
• Use --dry-run to preview planned execution without running commands.
• Use --audit-log <path> to mirror the structured event stream into a redacted JSONL file.

Recommended agent flow:

1. Generate or select a plain shell script with @LOCAL and @REMOTE markers.
2. Add bounded directives only where needed: @TIMEOUT, @RETRY, and @EXPORT.
3. Run with --json or --jsonl.
4. Let the outer agent decide whether to retry, branch, or stop based on Shellflow's structured result.

Shellflow intentionally does not provide:

• Conditional directives such as @IF stdout_contains=...
• A workflow DSL or embedded ReAct loop
• Heuristic destructive-command detection

Those decisions belong in the outer agent or automation layer.

shellflow run <script>
shellflow run <script> --verbose
shellflow run <script> --json
shellflow run <script> --jsonl
shellflow run <script> --no-input
shellflow run <script> --dry-run
shellflow run <script> --audit-log ./audit.jsonl --jsonl
shellflow run <script> --ssh-config ./ssh_config
shellflow --version

Examples:

shellflow run playbooks/hello.sh
shellflow run playbooks/hello.sh -v
shellflow run playbooks/hello.sh --json
shellflow run playbooks/hello.sh --jsonl --no-input
shellflow run playbooks/hello.sh --dry-run --jsonl
shellflow run playbooks/hello.sh --audit-log ./audit.jsonl --jsonl
shellflow run playbooks/hello.sh --ssh-config ~/.ssh/config.work

Useful commands:

just sync
just test
just bdd
just test-all
just typecheck
just build
just publish

Direct verification commands:

uv run pytest -q
uv run behave features
uv run ruff check .
uv run ty check src tests
uv build

Shellflow supports both local publishing and GitHub Actions release publishing.

just publish

uv publish uses standard uv authentication mechanisms such as UV_PUBLISH_TOKEN, or PyPI trusted publishing when supported by the environment.

The repository includes:

• .github/workflows/ci.yml for lint, type-check, test, and build verification.
• .github/workflows/release.yml for publishing to PyPI when a tag like v0.1.0 is pushed.

Recommended release flow:

git tag v0.1.0
git push origin v0.1.0

To use trusted publishing with PyPI:

1. Create a pypi environment in GitHub repository settings.
2. Add this repository as a trusted publisher in the PyPI project settings.
3. Push a v* tag.

The release workflow then runs verification, builds distributions with uv build, and uploads them with uv publish.

shellflow/
├── src/shellflow.py
├── tests/
├── features/
├── playbooks/
├── pyproject.toml
├── Justfile
└── README.md

Apache-2.0