Pipeline & Workflows

Chain models together for multi-step workflows

Pipeline mode chains multiple AI models together, where each step builds on the output of the previous step. This enables sophisticated multi-step workflows.

Usage

TUI (Interactive)

/workflow <name> "<prompt>"      # run saved workflow
/workflows                       # manage saved pipelines
/interactive                     # toggle pause between steps

CLI

puzldai run "<prompt>" -P "<agent:action,agent:action,...>" [-i]

Flag	Description
-P	Pipeline steps (e.g., “gemini:analyze,claude:code”)
-T	Use a saved pipeline template
-i	Interactive mode - pause between steps

Examples

# TUI: Run saved workflow
/workflow code-review "Build a rate limiter"

# CLI: Analyze → Code → Review pipeline
puzldai run "Build a rate limiter" -P "gemini:analyze,claude:code,codex:review" -i

Available Actions

Action	Description
analyze	Analyze the task and provide insights
code	Write code for the task
review	Review and suggest improvements
fix	Fix issues identified in previous step
test	Write tests for the code
summarize	Summarize the content concisely
prompt	Generic prompt (default)

How It Works

1. Step 1

   First agent receives the original prompt and performs its action.

2. Step 2

   Second agent receives the original prompt PLUS the output from step 1.

3. Step N

   Each subsequent step receives all previous context, building on prior work.

Saved Workflows

Create and manage workflows through the /workflows interactive interface:

# Open workflow manager (create, edit, delete workflows)
/workflows

# Run a saved workflow
/workflow code-review "Build user authentication"

The /workflows manager lets you:

• Create new workflows with custom pipeline steps
• Edit existing workflow pipelines
• Delete workflows you no longer need
• View workflow details and step configuration

Interactive Mode

Enable interactive mode to pause between steps:

/interactive    # Toggle interactive mode

When enabled, you can review each step’s output before proceeding.

Example Workflow

Pipeline: gemini:analyze → claude:code → codex:review

Step 1: gemini [analyze]
━━━━━━━━━━━━━━━━━━━━━━━
Analysis: The rate limiter should use a token bucket algorithm...

Step 2: claude [code]
━━━━━━━━━━━━━━━━━━━━━━
```typescript
class RateLimiter {
  private tokens: number;
  ...
}

Step 3: codex [review] ━━━━━━━━━━━━━━━━━━━━━━ Review: The implementation looks good. Suggestions:

1. Add thread safety…

## Keyboard Shortcuts

| Key | Action |
|-----|--------|
| `Escape` | Hide view (keeps running) |
| `Ctrl+E` | Return to pipeline view |
| `Ctrl+C` | Cancel pipeline |
| `Space` | Continue to next step (interactive mode) |

## Tips

<Tip>
**Match models to tasks** - Use models that excel at each step (e.g., Claude for code, Gemini for analysis).
</Tip>

<Tip>
**Keep pipelines focused** - 3-4 steps is usually optimal. Too many steps can compound errors.
</Tip>