workflow-orchestrator

Workflow Orchestrator

A skill that coordinates workflows across multiple skills by evaluating hook metadata, recording pending follow-ups, and running only the actions that are safe and supported in the current host runtime.

When This Skill Activates

This skill should be used when:

• A skill completes its main workflow
• A milestone is reached (PRD complete, implementation done, etc.)
• User says "complete workflow" or "finish the process"

How It Works

┌─────────────────────────────────────────────────────────────┐
│                    Workflow Orchestration                   │
├─────────────────────────────────────────────────────────────┤
│                                                             │
│  1. Detect Milestone → 2. Read Hooks → 3. Record/Run Safe Follow-ups │
│                                                             │
│  prd-planner complete                                       │
│       ↓                                                     │
│  workflow-orchestrator                                      │
│       ↓                                                     │
│  ┌─────────────────────────────────────┐                   │
│  │ declared self-improving follow-up   │ (record/run)       │
│  │ declared session logging follow-up  │ (record/run)       │
│  └─────────────────────────────────────┘                   │
│                                                             │
└─────────────────────────────────────────────────────────────┘

Trigger Configuration

Read trigger definitions from skills/auto-trigger/SKILL.md:

hooks:
  after_complete:
    - trigger: self-improving-agent
      mode: background
    - trigger: session-logger
      mode: auto
  on_error:
    - trigger: self-improving-agent
      mode: background

Execution Modes

Milestone Detection

PRD Complete

Detected when:
- docs/{scope}-prd.md exists
- All phases in {scope}-prd-task-plan.md are checked
- Status shows "COMPLETE"

Actions:
1. Record self-improving-agent as a background follow-up
2. Run or record session-logger if the host supports it

Implementation Complete

Detected when:
- All PRD requirements implemented
- Tests pass
- Code committed

Actions:
1. Ask before running code-reviewer
2. Run create-pr only when the user requested submission
3. Run or record session-logger if the host supports it

Self-Improvement Applied

Detected when:
- Candidate validated with auditable evidence
- One named durable owner changed
- Representative behavior rerun
- Candidate recorded as applied

Actions:
1. Ask before running create-pr
2. Run or record session-logger if the host supports it

Learning Candidate (Skill Complete)

Detected when:
- A skill completes its workflow and produces reusable evidence
- User provides feedback
- Error or issue encountered

Actions:
1. Record self-improving-agent as a background follow-up
2. Run or record session-logger if the host supports it

The self-improving-agent:
- Captures a candidate only when reusable evidence exists
- Excludes raw transcripts and private tool payloads
- Keeps uncertain findings under observation
- Validates candidates only with explicit, auditable evidence
- Applies validated changes only to a named durable owner with a change reference
- Proves the representative behavior after application

Error Handling (on_error)

Detected when:

• A command returns non-zero exit code
• Tests fail after following skill guidance
• User reports the guidance produced incorrect results

Actions:

1. Record self-improving-agent (background) for self-correction
2. Run or record session-logger to capture error context

Hook Implementation in Skills

To declare follow-up metadata, add this section to any skill's SKILL.md:

## Auto-Trigger (After Completion)

When this skill completes, record or run supported follow-ups:

```yaml
hooks:
  after_complete:
    - trigger: skill-name
      mode: auto|background|ask_first
      context: "relevant context"
  on_error:
    - trigger: self-improving-agent
      mode: background

Current Hook Contract

Hook metadata is declarative intent, not proof of CLI automation. Read the current skill front matter before acting. An absent hook means no declared follow-up; an existing hook still requires host support and the permission boundary for the target action.

Universal Learning Pattern

┌─────────────────────────────────────────────────────────────┐
│                  Skill Completes With Evidence              │
└──────────────┬──────────────────────────────────────────────┘
               │
               ↓
    ┌──────────────────────┐
    │ workflow-orchestrator │
    └──────────┬───────────┘
               │
    ┌──────────┴─────────┐
    ↓                   ↓
self-improving-agent  session-logger
    ↓                   ↓
Capture candidate     Save bounded context
    ↓                   ↓
Validate evidence     Log session
    ↓
Apply to named owner
    ↓
create-pr (only if submission was requested)

## Workflow Examples

### Example 1: PRD Creation Workflow

User: "Create a PRD for user authentication" ↓ prd-planner executes ↓ Phase 6 complete: PRD delivered ↓ workflow-orchestrator detects milestone ↓ ┌─────────────────────────────────┐ │ Background: self-improving-agent │ → Records learning proposal │ Auto: session-logger │ → Saves session when supported └─────────────────────────────────┘

### Example 2: Full Feature Workflow

User: "Create a PRD and implement it" ↓ prd-planner → workflow-orchestrator ↓ self-improving-agent (candidate capture only) ↓ prd-implementation-precheck ↓ implementation complete → workflow-orchestrator ↓ code-reviewer → optional candidate capture ↓ create-pr (only when requested) → workflow-orchestrator ↓ session-logger

Each milestone can produce a `self-improving-agent` follow-up, but durable
skill edits still require validation or explicit approval.

## Implementation Steps

### Step 1: Detect Milestone

Check for completion indicators:

```bash
# PRD complete?
grep -q "COMPLETE" docs/{scope}-prd-task-plan.md

# All phases checked?
grep -q "^\- \[x\].*Phase 6" docs/{scope}-prd-task-plan.md

# PRD file exists?
ls docs/{scope}-prd.md

Step 2: Read Trigger Config

# Read hooks from auto-trigger skill
cat skills/auto-trigger/SKILL.md

Step 3: Record or Execute Hooks

For each hook in order (before_start, after_complete, on_error):

1. Check if condition is met
2. Record or execute based on mode and host support
3. Pass context to triggered skill
4. Wait/continue based on mode

Step 4: Update Status

Log what was triggered and the result:

## Workflow Execution

- [x] self-improving-agent (background) - Started
- [x] session-logger (auto) - Session saved
- [ ] create-pr (ask_first) - Pending user approval

Skills with Auto-Trigger

Adding Follow-up Metadata to Existing Skills

To add follow-up metadata to an existing skill, add to the end of its SKILL.md:

---

## Auto-Trigger

When this skill completes, record or run supported follow-ups:

```yaml
hooks:
  after_complete:
    - trigger: session-logger
      mode: auto
      context: "Save session context"

For more complex triggers, specify mode and context:

```markdown
## Auto-Trigger

When this skill completes:

```yaml
hooks:
  after_complete:
    - trigger: next-skill
      mode: background
      context: "Description"
    - trigger: session-logger
      mode: auto
      context: "Save session"
    - trigger: create-pr
      mode: ask_first
      context: "Create PR if files modified"
  on_error:
    - trigger: self-improving-agent
      mode: background

## Best Practices

1. **Log only when supported and appropriate** - Session logging is a bounded optional follow-up
2. **Ask before major actions** - PRs, deployments, destructive changes
3. **Background for analysis** - Reflection, evaluation, optimization
4. **Auto for status** - Logging, status updates, bookmarks
5. **Don't create loops** - Ensure chains terminate