flowsh-cli

flowsh-cli turns workflow YAML into executable shell scripts.

Use it when you want a small, reproducible way to encode a workflow once and rerun it locally or in CI.

Quick Start

uvx flowsh-cli path/to/workflows.yml

That reads the workflow YAML at the path you provide and writes harness scripts to the current working directory.

In this repository, the example workflow file lives at workflows.yml.

Useful flags:

uvx flowsh-cli path/to/workflows.yml --dry-run
uvx flowsh-cli path/to/workflows.yml --workflow wf_example
uvx flowsh-cli path/to/workflows.yml --force
uvx flowsh-cli path/to/workflows.yml --schema
uvx flowsh-cli --examples
uvx flowsh-cli --example simple

You can also run it locally with uv run flowsh-cli.

Use As A Library

flowsh-cli also works as a plain Python library, with no CLI dependencies (no typer, no terminal UI). This is the right choice if you want to parse or render workflows from your own Python code instead of shelling out to the CLI.

uv add flowsh-cli

from pathlib import Path
from flowsh_cli import parse_workflows, render_harness, harness_path

workflows = parse_workflows(Path("workflows.yml"))
for workflow in workflows:
    harness_path(workflow).write_text(render_harness(workflow))

Public API (re-exported from flowsh_cli):

If you also want the flowsh-cli command-line tool (built on typer), install the cli extra:

uv add "flowsh-cli[cli]"
# or, for one-off use without installing:
uvx flowsh-cli path/to/workflows.yml

Workflow Shape

The input file is a single mapping with a workflows list:

workflows:
  - id: wf_example
    name: Example
    params:
      - name: ISSUE_NUMBER
        description: Issue to inspect
        required: true
    steps:
      - type: vars
        values:
          TODAY: date -u +%F
      - type: bash
        run: printf 'today=%s\n' "$TODAY"
      - type: agent
        agent: general
        model: openai/gpt-5
        command: review
        expandPrompt: true
        prompt: |
          Review issue ${ISSUE_NUMBER} and summarize the repository state.
       - type: parallel
         steps:
           - type: bash
             run: echo "worker A"
           - type: bash
             run: echo "worker B"
       - type: for
         in: ITEMS
         item: ITEM
         steps:
           - type: bash
             run: echo "$ITEM"
       - type: while
         condition: '[ -n "$(ls doc/plan/steps/planned/ 2>/dev/null)" ]'
         steps:
           - type: bash
             run: echo "keep looping until the queue is empty"

Harness paths are derived from workflow ids. wf_example becomes example.sh in the current working directory.

Step Types

Agent Behavior

agent prompts are literal by default. Use expandPrompt: true only when you want $VAR or ${VAR} tokens from earlier vars steps substituted at runtime.

expandPrompt: true does plain text replacement only. It does not evaluate shell expressions like $(...), backticks, or globs.

Set capture: VARIABLE_NAME on an agent step when you want the agent's final plain-text answer stored in a shell variable for later vars or bash steps. The captured value is the answer text (extracted from the underlying --format json event stream via jq), not the raw JSON stream — requires jq to be installed.

Set dangerouslySkipPermissions: true only when you want the generated harness to pass --dangerously-skip-permissions to the agent runtime. The YAML alias dangerously-skip-permissions is also accepted.

Inside a parallel step, each child branch runs in its own backgrounded subshell, so capture set in one branch is not visible to sibling branches and does not survive past the parallel block — there is no gather/join mechanism. To share results across branches (e.g. fan out N agents and merge their verdicts afterward), have each branch write its result to its own file (via a bash instruction in the agent prompt, or a bash child step), then read all the files back in a plain bash or vars step placed after the parallel block.

Validation And Safety

The parser is strict:

• workflow YAML must be UTF-8, non-empty, and under 1 MiB
• the root must be a mapping with a workflows key
• duplicate mapping keys and YAML aliases are rejected
• workflow ids must match wf_[A-Za-z0-9_-]+
• workflow and step names must be non-empty single-line labels
• unsafe control characters are rejected in executable fields
• agent names may contain only letters, digits, _, and -

Generated harnesses are non-interactive, owner-executable, and refuse to overwrite existing outputs unless --force is set.

Logs go to .flowsh/logs by default. Set FLOWSH_LOG_DIR to use another relative log directory.

Development

uv sync
make install
make qa
make security
make clean

make qa runs linting, tests, a dependency security audit, and a local build. make security audits hash-pinned runtime, CLI-extra, and development dependencies exported from uv.lock against known vulnerability advisories. It requires network access and fails if the audit cannot complete. The isolated build backend and downloaded CI tools are outside this dependency scan.

GitHub Actions are pinned to immutable commits. CI installs OpenCode 1.14.40 from a versioned Linux release archive and verifies its SHA-256 before extraction. Workflow comments and prompts are passed as data, not interpolated into shell scripts; only repository-owner comments can request agent execution. Agent runs still have their job's declared GitHub permissions, so review untrusted issue content before authorizing write-capable work.

Release

make bump-patch

That bumps the patch version, runs QA, and publishes the package.

Limits

This repository intentionally stays focused on the Python CLI blueprint. It does not include the old TypeScript implementation, workflow templates, DSL explorer, plugin registry, or extra workflow node types.