Superlaunch Workflow

Claude-only 15-phase thorough pipeline for complex development tasks. Uses minions plugin agents (self-contained) driven by minions plugin hooks (Ralph-style loop driver). No Codex MCP dependency.

Key Architecture

• pipeline: "superlaunch" in state.json
• plugin: "minions" — so minions hooks fire
• Other plugins' hooks silently exit (they check plugin field in state.json)
• All agents are minions:* prefixed — they exist in the minions plugin

Execution Flow (Ralph-Style)

The orchestrator uses the Ralph Loop pattern: the Stop hook generates a schedule-driven orchestrator prompt every time Claude tries to stop. Claude reads state from disk and dispatches the current phase.

superlaunch.md initializes state (pipeline: "superlaunch", 15-phase schedule)
  → dispatches Phase S0 (Explore)
  → on-subagent-stop.sh validates output, checks gates, advances state
  → on-stop.sh reads schedule, generates phase prompt, blocks stop
  → Claude dispatches next minions agent
  → cycle repeats through all 15 phases

15-Phase Thorough Schedule

Phase S0  │ EXPLORE   │ Explore                 │ dispatch  → explorers + aggregator
Phase S1  │ PLAN      │ Brainstorm              │ subagent  → brainstormer
Phase S2  │ PLAN      │ Plan                    │ dispatch  → planners + aggregator
Phase S3  │ PLAN      │ Plan Review             │ review    → plan-reviewer + supplementary
Phase S4  │ IMPLEMENT │ Implement               │ dispatch  → task-agent (parallel batch)
Phase S5  │ IMPLEMENT │ Simplify                │ subagent  → simplifier
Phase S6  │ IMPLEMENT │ Impl Review             │ review    → impl-reviewer + supplementary
Phase S7  │ TEST      │ Run Tests               │ subagent  → test-developer
Phase S8  │ TEST      │ Analyze                 │ subagent  → failure-analyzer
Phase S9  │ TEST      │ Develop Tests           │ subagent  → test-developer
Phase S10 │ TEST      │ Test Dev Review         │ review    → test-dev-reviewer + supplementary
Phase S11 │ TEST      │ Test Review             │ review    → test-reviewer + supplementary
Phase S12 │ FINAL     │ Documentation           │ subagent  → doc-updater + claude-md-updater
Phase S13 │ FINAL     │ Final Review            │ review    → final-reviewer + supplementary
Phase S14 │ FINAL     │ Completion              │ subagent  → shipper + retrospective

Stage Gates

Gate	Required Files	Transition
EXPLORE→PLAN	S0-explore.md	After Phase S0
PLAN→IMPLEMENT	S1-brainstorm.md, S2-plan.md, S3-plan-review.json	After Phase S3
IMPLEMENT→TEST	S4-tasks.json, S6-impl-review.json	After Phase S6
TEST→FINAL	S7-test-results.json, S9-test-dev.json, S11-test-review.json	After Phase S11
FINAL→COMPLETE	S13-final-review.json	After Phase S13

Review-Fix Cycles

Review phases (S3, S6, S10, S11, S13) support two-tier retry:

Tier 1: 10 fix attempts per review phase (within one run of the stage)
Tier 2: 3 stage restarts (each restart resets fix counters to 0)
Total:  up to 3 x 10 = 30 fix attempts per review phase before blocking

Coverage Loop

Phases S9 → S10 → S11 repeat until coverage >= coverageThreshold (default 90%) or 20 iterations reached.

Supplementary Agents

Dispatched in parallel with primary agents (controlled by supplementaryPolicy):

Phase	Supplementary Agents
S0	minions:deep-explorer
S2	minions:architecture-analyst
S3	minions:judgement-agent
S6	minions:judgement-agent, minions:critic, minions:silent-failure-hunter, minions:type-reviewer
S10	minions:judgement-agent
S11	minions:judgement-agent
S12	minions:claude-md-updater
S13	minions:judgement-agent, minions:pedant, minions:security-reviewer, minions:silent-failure-hunter
S14	minions:retrospective-analyst

State Schema (superlaunch)

{
  "version": 1,
  "plugin": "minions",
  "pipeline": "superlaunch",
  "status": "in_progress|blocked|complete",
  "currentPhase": "S0|S1|S2|...|S14|DONE|STOPPED",
  "currentStage": "EXPLORE|PLAN|IMPLEMENT|TEST|FINAL",
  "codexAvailable": false,
  "testDeveloper": "minions:test-developer",
  "failureAnalyzer": "minions:failure-analyzer",
  "docUpdater": "minions:doc-updater",
  "schedule": [/* 15 phases */],
  "gates": {/* 5 stage gates */},
  "stages": {
    "EXPLORE": {"status":"pending","phases":["S0"],"restartCount":0},
    "PLAN": {"status":"pending","phases":["S1","S2","S3"],"restartCount":0},
    "IMPLEMENT": {"status":"pending","phases":["S4","S5","S6"],"restartCount":0},
    "TEST": {"status":"pending","phases":["S7","S8","S9","S10","S11"],"restartCount":0},
    "FINAL": {"status":"pending","phases":["S12","S13","S14"],"restartCount":0}
  },
  "reviewPolicy": {"maxFixAttempts": 10, "maxStageRestarts": 3},
  "supplementaryPolicy": "on-issues",
  "coverageThreshold": 90,
  "webSearch": true,
  "fixAttempts": {},
  "coverageLoop": {"iteration": 0},
  "reviewFix": null,
  "supplementaryRun": {},
  "failure": null
}

Hook Responsibilities

Hook	Event	Superlaunch Behavior
on-stop.sh	Stop	Calls generate_sl_prompt() from lib/superlaunch.sh — schedule-driven prompt
on-subagent-stop.sh	SubagentStop	Validates output, checks gates, handles review-fix cycles, advances phase
on-task-gate.sh	PreToolUse (Task)	Validates minions:* agent matches current phase via is_sl_agent_allowed()
on-edit-gate.sh	PreToolUse (Edit/Write)	Allows edits in IMPLEMENT, TEST, and FINAL stages
on-launch-init.sh	UserPromptSubmit	Shows pipeline and stage info for stale state detection

Difference from /minions:launch

Aspect	launch	superlaunch
Phases	4 (F1-F4)	15 (S0 through S14)
Agents	minions:* (12 agents)	minions:* (26 superlaunch agents)
Hooks	Same hooks, launch branch	Same hooks, superlaunch branch
Codex	No	No
Review	5 parallel personality reviewers	Structured review-fix cycles
State	loop/maxLoops counters	currentStage/currentPhase schedule