Orchestration Patterns

Orchestration Patterns Skill

Purpose

Codify evidence-based delivery and iterative agent loop for orchestrating complex, long-running tasks. These patterns ensure verifiable progress and intelligent escalation.

When This Skill Applies

Invoke this skill when:

• Orchestrating multi-step implementation tasks
• Managing work across multiple subagents
• Running long-running sessions that need checkpoints
• Preparing PRs for merge (mandatory QAS gate)
• Coordinating team handoffs

Simon Willison's Agent Loop

Core Philosophy: "Iterate until success or blocked, then escalate."

┌─────────────────────────────────────────────────────────┐
│  THE AGENT LOOP (for every task)                       │
├─────────────────────────────────────────────────────────┤
│                                                         │
│  1. GOAL DEFINITION                                     │
│     └─ Clear acceptance criteria (from BSA/ticket)      │
│                                                         │
│  2. PATTERN DISCOVERY                                   │
│     └─ Search codebase, docs, previous sessions         │
│     └─ Use: pattern-discovery skill (auto-invoked)      │
│     └─ Or: /search-pattern for explicit code search     │
│                                                         │
│  3. ITERATIVE EXECUTION LOOP:                           │
│     ┌─────────────────────────────────────────────┐     │
│     │  Implement approach                         │     │
│     │       ↓                                     │     │
│     │  Run validation (yarn ci:validate)          │     │
│     │       ↓                                     │     │
│     │  If PASS → proceed to evidence              │     │
│     │  If FAIL → analyze error, adjust, repeat    │     │
│     │  If BLOCKED → escalate to TDM with context  │     │
│     └─────────────────────────────────────────────┘     │
│                                                         │
│  4. EVIDENCE ATTACHMENT                                 │
│     └─ Attach proof to Linear (see templates below)     │
│                                                         │
│  5. QAS GATE (MANDATORY before merge)                   │
│     └─ Invoke QAS subagent for independent review       │
│                                                         │
└─────────────────────────────────────────────────────────┘

Evidence-Based Delivery

Core Principle: "All work requires verifiable evidence - no 'trust me, it works'"

Evidence Types

Phase Evidence Requirements

QAS Pre-Merge Gate

MANDATORY: Before merging any PR, invoke QAS for independent review.

Why QAS Gate Matters

1. Separation of Concerns: QAS validates but doesn't write product code
2. Independent Verification: Catches what implementer missed
3. Bias Prevention: Fresh eyes on commit messages, patterns
4. Evidence in Linear: QAS posts final evidence + verdict to Linear (system of record)

QAS Invocation Pattern

# After implementation complete, before merge:
Task tool: QAS subagent
Prompt: "Review PR #XXX for {{TICKET_PREFIX}}-YYY. Validate:
  - Commit message format (ticket in subject line)
  - Code patterns (RLS, naming, structure)
  - CI status (all checks passing)
  - Evidence attachments in Linear
  Generate validation report to docs/agent-outputs/qa-validations/"

QAS Output Location

All QAS reports go to: docs/agent-outputs/qa-validations/{{TICKET_PREFIX}}-{number}-qa-validation.md

Escalation Patterns

When to Escalate

Escalation Template

**Escalation Required**

**Blocked On**: [specific blocker]
**Attempts Made**:

1. [what you tried]
2. [what you tried]

**Context**:

- Ticket: {{TICKET_PREFIX}}-XXX
- Session ID: [if available]
- Time blocked: X hours

**Request**: [specific ask - what do you need?]

Long-Running Task Checkpoints

For tasks spanning multiple tool calls or sessions:

Checkpoint Pattern

Every 10-15 tool calls:
1. Update todo list with current progress
2. If nearing context limit, summarize state
3. If handoff needed, provide continuation context

At session boundaries:
1. Summarize completed work
2. List remaining items
3. Document any blockers
4. Attach evidence to Linear

State Preservation

**Session Checkpoint**

**Completed**:

- [x] Task 1
- [x] Task 2

**In Progress**:

- [ ] Task 3 (at step X)

**Remaining**:

- [ ] Task 4
- [ ] Task 5

**Blockers**: [if any]

**Next Action**: [specific next step]

Orchestration Workflow Example

# Complete workflow for feature implementation:

1. /start-work {{TICKET_PREFIX}}-XXX
   └─ Syncs to dev, creates branch, sets context

2. Pattern discovery (skill auto-invokes or use /search-pattern)
   └─ Finds relevant patterns before implementation

3. [Implementation with agent loop]
   ├─ Implement
   ├─ Validate (yarn ci:validate)
   ├─ Adjust if needed
   └─ Repeat until passing

4. /pre-pr
   └─ Full validation checklist

5. Create PR with evidence

6. [QAS GATE - MANDATORY]
   └─ Invoke QAS subagent for review
   └─ Fix any blocking issues
   └─ Commit QAS report

7. Merge (only after QAS approval)

8. /end-work
   └─ Updates Linear, cleans up

Authoritative References

• {{PROJECT_REPO}}: Core agentic principles
• AGENT_WORKFLOW_SOP.md: Full agent workflow documentation
• CONTRIBUTING.md: Workflow requirements
• linear-sop skill: Evidence templates for Linear

Anti-Patterns to Avoid