Imported from npc-live/clawfirm (cmd/whip/SKILL.md). Install upstream with npx skills add npc-live/clawfirm --skill whip. Copyright stays with the author.

WhipFlow

WhipFlow 是 clawfirm 内置的 AI workflow 引擎，使用 OpenProse DSL 编排工作流。 通过内置 whipflow_run 工具直接执行，无需安装任何外部依赖。

执行方式

使用内置工具 whipflow_run：

示例：执行文件

{"file": "~/.clawfirm/workflows/media.whip", "user_inputs": {"track": "美食"}}

示例：直接传源码

{"source": "session \"Hello world\""}

Quick Start

agent worker:
  provider: "claude-code"
  model: sonnet
  tools: ["bash"]
  prompt: "You are a helpful assistant."

let result = session: worker
  prompt: "Echo hello world using bash"

Ask user for input

ask project_name: "What is the project name?"
ask description:  "One sentence description?"

let result = session: worker
  prompt: "Create a README for {project_name}: {description}"

Multi-step with context passing

let step1 = session: worker
  prompt: "Analyze the codebase structure"

let step2 = session: worker
  prompt: """
  Based on this analysis:
  {step1}

  Write a summary report.
  """

OpenProse Language Reference

OpenProse is a DSL for orchestrating multi-agent workflows. Programs consist of statements executed sequentially, with each session spawning a subagent.

File Format

Comments

# This is a standalone comment

session "Hello"  # Inline comment

• Comments begin with # and extend to end of line
• # inside string literals is NOT a comment
• Comments are stripped during compilation — agents never see them

String Literals

Single-line

session "Hello world"
session "Line one\nLine two"
session "She said \"hello\""

Escape sequences: \\ \" \n \t

Multi-line

session """
This is a multi-line prompt.
It preserves:
  - Indentation
  - Line breaks
  - All internal whitespace
"""

Opening """ must be followed by a newline. Content continues until closing """.

String Interpolation

let name = session "Get the user's name"
session "Hello {name}, welcome!"

• Use {varname} to embed variables
• Works in both single-line and multi-line strings
• Escape with \{ for a literal {

Import Statements

import "web-search" from "github:anthropic/skills"
import "code-analyzer" from "npm:@company/analyzer"
import "custom-tool" from "./skills/custom-tool"

Source types: github:user/repo, npm:package, ./local/path

Agent Definitions

agent name:
  provider: "claude-code"   # claude-code | claude | opencode | aider | custom:bin
  model: sonnet             # opus | sonnet | haiku
  tools: ["bash", "read", "write", "edit", "browser"]
  skills: ["skill-name"]
  prompt: "System prompt for this agent."
  permissions:
    read: ["*.md"]
    bash: deny

Properties

Model Selection

Permissions

agent secure-agent:
  permissions:
    read: ["*.md", "*.txt"]
    write: ["output/"]
    bash: deny
    network: allow

Permission values: allow, deny, prompt, or array of glob patterns.

Session Statement

The primary executable construct. Spawns a subagent to complete a task.

Variants

# Simple inline prompt
session "Prompt text"

# With agent reference
session: agentName

# Named session
session analysis: agentName

# With property overrides
session: agentName
  prompt: "Override the default prompt"
  model: opus

Properties

Variables & Context

Let (mutable) / Const (immutable)

let draft = session "Write initial draft"
const config = session "Get configuration"

# Reassign let only
draft = session "Improve the draft"
  context: draft

Context Property

let research = session "Research quantum computing"

# Single context
session "Write summary"
  context: research

# Multiple contexts
let analysis = session "Analyze findings"
session "Write final report"
  context: [research, analysis]

# Object shorthand (for parallel results)
parallel:
  a = session "Task A"
  b = session "Task B"

session "Combine results"
  context: { a, b }

Ask (user input at runtime)

ask name: "What is your name?"
ask count: "How many items?"

Composition Blocks

do: (anonymous sequential block)

do:
  session "Step 1"
  session "Step 2"

let result = do:
  session "Gather data"
  session "Process data"

Named Blocks (reusable)

block review-pipeline:
  session "Security review"
  session "Performance review"
  session "Synthesize reviews"

do review-pipeline

Block Parameters

block review(topic):
  session "Research {topic} thoroughly"
  session "Summarize {topic} analysis"

do review("quantum computing")
do review("machine learning")

Inline Sequence (Arrow)

session "Plan" -> session "Execute" -> session "Review"

let workflow = session "Draft" -> session "Edit" -> session "Finalize"

Parallel Blocks

parallel:
  session "Security review"
  session "Performance review"
  session "Style review"

All branches start simultaneously. Program waits for all to complete.

Named Parallel Results

parallel:
  security = session "Security review"
  perf     = session "Performance review"
  style    = session "Style review"

session "Unified review report"
  context: { security, perf, style }

Join Strategies

# Race — return on first completion
parallel ("first"):
  session "Try approach A"
  session "Try approach B"

# Any N — wait for N successes
parallel ("any", count: 2):
  session "Attempt 1"
  session "Attempt 2"
  session "Attempt 3"

Failure Policies

# Continue even if some branches fail
parallel (on-fail: "continue"):
  session "Task 1"
  session "Task 2"

# Ignore all failures
parallel (on-fail: "ignore"):
  session "Optional enrichment"

Fixed Loops

Repeat

repeat 3:
  session "Generate a creative idea"

repeat 5 as i:
  session "Iteration {i} of 5"

For-Each

let items = ["a", "b", "c"]
for item in items:
  session "Process: {item}"

for item, i in items:
  session "Item {i}: {item}"

# Inline array
for topic in ["AI", "climate", "space"]:
  session "Research this topic"
    context: topic

Parallel Fan-Out

parallel for item in items:
  session "Process {item} concurrently"

Unbounded Loops

Uses discretion markers (**...**) for AI-evaluated conditions.

# Loop until condition
loop until **the task is complete**:
  session "Continue working"

# Loop while condition
loop while **there are items to process**:
  session "Process next item"

# With safety limit
loop until **all bugs fixed** (max: 10):
  session "Find and fix a bug"

# With index variable
loop until **done** (max: 5) as attempt:
  session "Attempt {attempt}"

# Basic (always needs max to avoid infinite loop)
loop (max: 50):
  session "Process next item"

Multi-line Conditions

loop until ***
  the document is complete
  all sections reviewed
  and formatting is consistent
***:
  session "Continue working"

Pipeline Operations

let items = ["a", "b", "c"]

# Map — transform each element
let results = items | map:
  session "Process this item"
    context: item

# Filter — keep matching elements
let good = items | filter:
  session "Is this valid? Answer yes or no."
    context: item

# Reduce — accumulate into single result
let combined = items | reduce(summary, item):
  session "Add '{item}' to summary: {summary}"

# Parallel map
let results = tasks | pmap:
  session "Process concurrently"
    context: item

# Chaining
let result = topics
  | filter:
      session "Is this trending? Answer yes or no."
        context: item
  | map:
      session "Write a one-line pitch"
        context: item

Inside map, filter, pmap bodies, item is the current element. Inside reduce, variables are named explicitly: reduce(accVar, itemVar).

Error Handling

Try/Catch/Finally

try:
  session "Attempt risky operation"
catch:
  session "Handle the error"
finally:
  session "Always clean up"

# Capture error context
try:
  session "Call external API"
catch as err:
  session "Log and handle"
    context: err

Throw

throw "Precondition not met"

# Re-raise in catch block
try:
  try:
    session "Inner operation"
  catch:
    session "Partial handling"
    throw  # re-raises to outer catch
catch:
  session "Handle re-raised error"

Retry

session "Call flaky API"
  retry: 3

session "Rate-limited API"
  retry: 5
  backoff: "exponential"

Backoff strategies: "none" (default), "linear", "exponential"

Choice Blocks

AI picks the best option based on criteria.

choice **the severity of issues found**:
  option "Critical":
    session "Stop deployment, fix critical issues"
  option "Minor":
    session "Log issues for later and proceed"
  option "None":
    session "Deploy immediately"

Conditional Statements

All conditions use discretion markers (**...**) for AI evaluation.

if **the code has security vulnerabilities**:
  session "Fix security issues"
elif **the code has performance issues**:
  session "Optimize performance"
else:
  session "Proceed with normal review"

Complete Example

# Research + write + review workflow

agent researcher:
  provider: "claude-code"
  model: sonnet
  tools: ["bash", "read"]
  prompt: "You are a thorough research assistant."

agent writer:
  provider: "claude-code"
  model: opus
  tools: ["write", "edit"]
  prompt: "You write clear, concise documentation."

ask topic: "What topic should I research?"

# Parallel research from multiple angles
parallel:
  market  = session: researcher  prompt: "Research market landscape for {topic}"
  tech    = session: researcher  prompt: "Research technical aspects of {topic}"
  trends  = session: researcher  prompt: "Research current trends in {topic}"

# Write a comprehensive report
let report = session: writer
  prompt: """
  Write a comprehensive report on {topic}.

  Use these research inputs:
  - Market: {market}
  - Technical: {tech}
  - Trends: {trends}
  """

# Iterative improvement loop
loop until **the report is polished and publication-ready** (max: 3):
  report = session: writer
    prompt: "Review and improve this report:\n{report}"

# Final quality check
if **the report meets publication standards**:
  session: writer  prompt: "Format the final report for publication"
else:
  session: writer  prompt: "Flag issues and create an improvement plan"

Skill Invocation

Invoke a Claude Code skill (/<name>) as a session step.

# Basic invocation
skill commit

# With named parameters
skill review-pr args="123"

# Capture output
let review = skill simplify

Tips

• One goal per session — focused prompts produce better results
• Use ask for values that change per run
• Name outputs with let so you can reference them later
• loop until is ideal for iterative refinement tasks
• parallel for independent tasks that don't need each other's output
• try/catch around network calls or operations that might fail
• choice when the AI should decide which path to take