Skill Orchestrator Patterns

Orchestrator Patterns

This skill defines reusable patterns for the Orchestrator's pipeline stages, enabling compressed prompts.

[!IMPORTANT] Reference this skill for Stage Cycle logic. The Orchestrator uses these patterns instead of verbose per-scenario instructions.

Stage Dispatch Table

Pattern: Stage Cycle

Applicability

• Stages with Init → Review → Revision flow
• Applies to: Analysis, Architecture, Planning, Execution

Init Phase

INPUT: {stage_name}, {agent_prompt}, {artifact_path}

FLOW:
1. Read agent prompt
2. Pass required context to agent
3. Wait for result: { artifact_file, blocking_questions }

DECISION:
- IF blocking_questions → STOP, ask user (Scenario 14)
- ELSE → proceed to Review

Review Phase

INPUT: {artifact_file}, {reviewer_prompt}, {iteration}

FLOW:
1. Read reviewer prompt
2. Pass artifact + context to reviewer
3. Wait for result: { review_file, has_critical_issues }

DECISION TABLE:
| Condition | Action |
|-----------|--------|
| No issues | → Next Stage |
| Issues AND iteration < max | → Revision |
| Critical issues AND iteration = max | → STOP, ask user |
| Non-critical issues AND iteration = max | → Next Stage (with warning) |

Revision Phase

INPUT: {review_file}, {original_artifact}, {agent_prompt}

INSTRUCTION TO AGENT:
"Fix comments from {review_file}. Do NOT change unrelated parts. Preserve structure."

FLOW:
1. Pass review + original to agent
2. Wait for updated artifact
3. Return to Review Phase (+1 iteration)

Expected Results Schema

From Doer Agent (Analyst, Architect, Planner, Developer)

artifact_file: "path/to/artifact.md"
blocking_questions: []  # Empty = proceed

From Reviewer Agent

review_file: "path/to/review.md"
has_critical_issues: true/false

Extended Schema (Developer)

modified_files: [...]
new_files: [...]
test_report: "tests/tests-{ID}/test-{ID}-{slug}.md"
documentation_updated: true
open_questions: []  # Treated as blocking_questions

Extended Schema (Planner)

plan_file: "path/to/plan.md"
task_files: ["path/to/task1.md", ...]
blocking_questions: []

Extended Schema (Plan Reviewer)

review_file: "path/to/plan_review.md"
has_critical_issues: true/false
comments_count: number
coverage_issues: [...]
missing_descriptions: [...]

Extended Schema (Code Reviewer)

Structured footer (control-flow + status). The prose review report — three pillars plus the "Verified" block — is passed to the developer as the report body; it is NOT a structured key.

review_status: "APPROVED | REJECTED"
has_critical_issues: true/false   # sole control-flow key (see DECISION TABLE)
e2e_tests_pass: true/false
stubs_replaced: true/false

Exceptions (Not Pattern-Compatible)

Scenario 13: Completion

No reviewer. Unique flow:

1. Archive current docs/TASK.md (via skill-archive-task)
2. Collect statistics
3. Generate final report

Scenario 14: Blocking Questions

Immediate pause. Flow:

1. Formulate message with blocking questions
2. Wait for user answers
3. Resume at current stage

Stage-Specific Notes

Analysis Stage

• Must apply: skill-archive-task protocol before creating new TASK
• Instruction for Analyst: "Create/Overwrite docs/TASK.md completely. Do NOT append."

Planning Stage

• Iteration limits: Max 1 revision cycle for Plan
• Extended result: task_files[], coverage_issues[], missing_descriptions[]

Execution Stage

• Per-task tracking: TASK {n} of {total}
• Iteration limit: Max 1 fix cycle
• Instruction for Developer: "Fix comments. Do NOT refactor. Run tests."