npx claudepluginhub qmu/workaholic --plugin trippinThis skill is limited to using the following tools:
sh/cleanup-worktree.shsh/ensure-worktree.shsh/init-trip.shsh/list-trip-worktrees.shsh/log-event.shsh/read-plan.shsh/trip-commit.shsh/validate-dev-env.shThree-agent collaboration workflow: Planner (progressive/business), Architect (neutral/structural), Constructor (conservative/technical). Dual objectives: maximize benefit (optimization) and prevent harm (constraint satisfaction).
The leader agent is the sole workflow coordinator. No sub-agent may autonomously advance. After each task, agents STOP and report completion. The leader waits for ALL concurrent tasks before issuing the next round.
Reviews create dialectical tension across perspectives. Requirements:
Planning: Concurrent artifacts, one-turn review, accept/revise/escalate, moderate, plan fixed. Coding: Concurrent launch, review and testing, iteration, done (or rollback to planning).
All scripts use absolute paths from home directory.
| Script | Usage |
|---|---|
ensure-worktree.sh <trip-name> | Create isolated worktree and branch |
list-trip-worktrees.sh | List existing trip worktrees (JSON) |
init-trip.sh <trip-name> [instruction] | Create artifact directories and plan.md |
validate-dev-env.sh <worktree_path> | Check env files, dependencies, ports |
read-plan.sh <trip-path> | Read plan state as JSON |
trip-commit.sh <agent> <phase> <step> <description> | Commit with [Agent] description format |
log-event.sh <trip-path> <agent> <event-type> <target> <impact> | Append to event-log.md |
Script base path: ${CLAUDE_PLUGIN_ROOT}/skills/trip-protocol/sh/
.workaholic/.trips/<trip-name>/
directions/ # Planner: direction-v1.md, direction-v2.md, ...
models/ # Architect: model-v1.md, model-v2.md, ...
designs/ # Constructor: design-v1.md, design-v2.md, ...
reviews/ # round-1-<agent>.md, response-<author>-to-<reviewer>.md
rollbacks/ # rollback-v<N>.md
reviews/ # rollback-v<N>-<agent>.md
event-log.md # Append-only event log
plan.md # Trip state and progress
Each revision is a new file (e.g., direction-v2.md), preserving history. Only the artifact's author may modify the original file; others express feedback through review files.
plan.md tracks trip lifecycle state with YAML frontmatter (instruction, phase, step, iteration, updated_at) and three sections: Initial Idea, Plan Amendments (leader decision log), and Progress (checklist with agent attribution). Update frontmatter at phase transitions; agents append progress entries bundled with artifact commits.
Step identifiers: planning/not-started, planning/artifact-generation, planning/one-turn-review, planning/respond-to-feedback, planning/moderation, coding/concurrent-launch, coding/review-and-testing, coding/iteration-N, complete/done, complete/followup.
Append-only event-log.md with columns: Timestamp, Agent, Event, Target, Impact. Call log-event.sh before trip-commit.sh. Leader-authored events require accurate impact; agent-authored events are best-effort. The commit script emits a warning if event-log.md exists but was not staged. Event types correspond to workflow actions (artifact lifecycle, reviews, gates, testing, rollbacks, phase transitions).
All three agents write simultaneously: Planner writes directions/direction-v1.md (business vision, no codebase exploration), Architect writes models/model-v1.md (structural bridge), Constructor writes designs/design-v1.md (technical plan). GATE: wait for all three.
Each agent writes ONE consolidated review of the other two artifacts to reviews/round-1-<agent>.md. Include: reviewer name, artifacts reviewed, decision per artifact, domain-specific feedback with concerns and proposals, and cross-artifact coherence assessment. GATE: wait for all three.
For each "Request revision", the author writes reviews/response-<author>-to-<reviewer>.md choosing: (a) accept and revise (new artifact version), or (b) escalate to the third agent. GATE: wait for all responses.
The uninvolved third agent reads both positions, writes a synthesized resolution as the next artifact version. After moderation, the plan is fixed -- no further review rounds.
Planning completes when all reviews approved, all revisions accepted, or all escalations moderated.
Maximum 3 review rounds. If consensus is not reached after round 3, the leader invokes forced moderation on all unresolved disagreements and appends a "Forced Convergence" entry to plan.md's Plan Amendments section with the unresolved items and rationale. A forced-convergence event is logged. After moderation, the plan is fixed and the team proceeds to Coding Phase.
Constructor implements + internal tests; Planner builds dev env + plans E2E scenarios; Architect discovers codebase. GATE: wait for all three.
Architect performs analytical review of Constructor's changes. GATE. Planner validates via E2E testing. GATE.
If issues found: Constructor fixes → Architect re-reviews → Planner re-tests. Repeat until approved.
Any agent may propose returning to Planning Phase. Requires 2/3 majority. Proposer writes rollbacks/rollback-v<N>.md; others vote in rollbacks/reviews/rollback-v<N>-<agent>.md. On approval, return to Planning with incremented artifact versions.
After the trip reaches complete/done, the team lead may receive follow-up requests from the user. The team composition rule applies strictly:
Never create new agent team members. The only agents permitted in the worktree are the three designated teammates: Planner, Architect, and Constructor. This applies regardless of the nature of the follow-up request.
Lead handles directly when the task is: answering a question, reading files, making a single-file edit, running a command, or any task that does not require the specialized perspective of a designated agent.
Lead re-invokes designated agents when the task involves: multi-file changes, implementation work (Constructor), structural review (Architect), E2E validation (Planner), or any work that falls within a designated agent's domain. Re-invoked agents retain their original role boundaries, QA domains, and behavioral constraints from the trip session.
Update plan.md frontmatter: set step to complete/followup when follow-up work begins, return to complete/done when it finishes. Log follow-up events to event-log.md.
Step identifier: complete/followup
Applies to projects with user-facing interfaces (web, CLI, API). Planner detects existing E2E frameworks; Playwright is the recommended default. Tests must be CLI-runnable. Projects that are purely library/configuration may skip E2E.
Before implementation, Constructor runs: bash ${CLAUDE_PLUGIN_ROOT}/../drivin/skills/system-safety/sh/detect.sh. If system_changes_authorized is false, use project-local alternatives only.
All trip output must be written in English. This applies to: artifact content (directions, models, designs), review files, event log entries, plan.md content, commit messages, PR titles and bodies, code and code comments, and any agent-generated documentation. The only exception is content written inside the .workaholic/ directory, which follows the consumer project's CLAUDE.md language setting (English or Japanese). When the consumer's CLAUDE.md specifies Japanese for .workaholic/, trip artifacts stored in .workaholic/.trips/ may use Japanese.
Format: [Agent] Descriptive summary of what was accomplished. Description must be a clear English sentence (not file names or terse labels). Every discrete workflow step produces a commit.
Each artifact file: # <Type> v<N>, then metadata fields (Author, Status: draft/under-review/approved, Reviewed-by), then Content section, then Review Notes section.