Wf Compose

Compose a Weft Workflow

Read the conversation context, scan available skills, identify gaps, and propose a v2 workflow template with loops and skill blocks.

Arguments

$ARGUMENTS

Modes

Step 1: Gather Context

Understand what the user is trying to do:

1. Review the recent conversation for intent (what task, what repo, what outcome).
2. Check git state:

   git branch --show-current 2>/dev/null
   git diff --stat 2>/dev/null | tail -5
   

3. Check if a weft workflow is already active:

   python3 "${CLAUDE_PLUGIN_ROOT}/core/cli.py" status 2>/dev/null
   

4. If --from <template> was provided, load it as the starting point:

   python3 "${CLAUDE_PLUGIN_ROOT}/core/cli.py" preview <template>
   

Step 2: Scan Skill Registry

Build a map of what skills are available:

1. Read the local skills registry, if any (path varies by setup):

   cat "${CLAUDE_SKILLS_REGISTRY:-$HOME/.claude/skills-registry.json}" 2>/dev/null
   

2. List weft templates:

   python3 "${CLAUDE_PLUGIN_ROOT}/core/cli.py" start
   

3. Categorize skills by function (examples — substitute what you have available):
   • Review: staff-review, arch-review, code-review, differential-review
   • Fix/Polish: fix-polish, refactor, simplify
   • Test: infra-test, webapp-testing
   • Plan: aot-plan, spec-first
   • Research: perplexity, context7, research-loop
   • Deploy: deploy-service, pr-ready

Step 3: Gap Analysis

Compare what the user described against available skills:

1. Extract skill references from the user's description (explicit names like "/staff-review" or implicit like "review code", "test it", "deploy").
2. For each referenced skill, check if it exists in the registry.
3. For missing skills, present options:

   Missing skill: /devils-advocate
   Options:
   1. Create a stub skill (I'll generate a skeleton)
   2. Use /staff-review instead (similar purpose)
   3. Skip this step
   

4. Wait for user choice on each gap before proceeding.

Step 4: Generate Template

Build a v2 template JSON:

1. Map each step in the user's described workflow to a template step.

2. For each step, set:

   • name: kebab-case identifier
   • skill: the matching skill name (e.g., "/staff-review"), or null if manual
   • on_fail: "retry" for review/test steps, "block" for critical gates, "continue" for optional steps
   • guards: add logical guards (e.g., no git push before review)
   • description: one-line summary of what the step does

3. For iterative segments (user said "until", "repeat", "loop", "iterate"):

   • Identify the loop boundary (which steps repeat)
   • Set loop_back_to on the last step of the loop, pointing to the first
   • Set max_iterations (default 3, or what the user specified)
   • Set exit_condition from the user's description (natural language)

4. Add schema_version: 2 to the template root.

Step 5: Present to User

Show the proposed workflow in two formats:

ASCII Diagram

Draw the workflow as a flow diagram showing loops:

  ┌────────────┐     ┌───────────────┐     ┌─────────────┐
  │   review    │────>│  fix-issues   │────>│  run-tests   │
  │ /staff-rev  │     │ /fix-polish   │     │              │
  └────────────┘     └───────────────┘     └──────┬──────┘
        ^                                         │
        │        ↻ until clean (max 3)            │
        └─────────────────────────────────────────┘
                          │ done
                          v
                   ┌─────────────┐
                   │    ship     │
                   │  /pr-ready  │
                   └─────────────┘

For linear segments, use a simple arrow chain:

  setup ──> plan ──> implement ──> verify

JSON Preview

Show the full template JSON, formatted for readability.

Prompt

Ask the user:

Approve this workflow? (approve / edit / cancel)
- approve: Save template and optionally start it
- edit: Describe what to change
- cancel: Discard

Step 6: Save and Start

On approve:

1. Save the template:

   echo '<json>' | python3 "${CLAUDE_PLUGIN_ROOT}/core/cli.py" save-template
   

2. Ask: "Start this workflow now? (y/n)"
3. If yes: invoke /wf-start <template-name>

On edit:

1. Ask what to change
2. Modify the template
3. Go back to Step 5 (re-present)

On cancel:

1. Discard and confirm

Design Rules

• Every step with a matching skill gets a skill field. This is metadata — Claude reads it from context.md and knows which skill to invoke.
• Loops are defined by loop_back_to on the last step of the repeating segment. The state machine handles the rest.
• exit_condition is evaluated by Claude (natural language), not by scripts. Keep conditions specific and observable: "no MEDIUM+ issues" not "code is good enough".
• max_iterations defaults to 3. If the user says "until done" without a cap, set it to 5 and note the cap.
• Guards should prevent premature actions: no git push before review, no deploy before tests.
• Template names are kebab-case. If the user doesn't name it, derive from the description.