AI Agent Workflow Collection

Use when tearing down a live agent-team — a graceful (soft) or forced (--hard) abort that records a durable abort-report postmortem, writes the 'aborted' audit event (the resume signal), sets the team status, and applies the worktree policy.

Internal — first-run Atelier bootstrap into Memex. Seeds Atelier's roles + agent profiles into ~/.memex/agents.db, provisions the atelier store via memex:core:create-store, and writes the bootstrap marker. Idempotent.

Use when starting any new project, feature, or initiative — PM agent grills requirements until all design decisions are resolved.

Use when encountering a bug or unexpected behavior at any phase — diagnoses root cause, writes a regression test, and resumes the interrupted phase.

Live orchestrator dispatch procedure — at plan:approved, drive the deterministic-host pipeline via dispatchhostpipeline (a single awaited coroutine that spawns ephemeral per-attempt claude -p workers and drives the kept WaveDispatcher engine internally), and surface the agent-team meeting / side-query / roster-extension / persona-gap-escalation behaviors. Read by the PM/orchestrator; not a user-facing slash command.

Use when QA is approved and the work is ready to hand off — commits the work to Atelier's feature branch (never pushes or opens a PR — that is the human's step), or abandons it, with pre-flight checks and phase completion.

Use at the end of any session — records current state to the DB so the next session can resume.

Use when review feedback has been posted — evaluates each item, pushes back on factual errors with evidence, implements accepted changes, and requests re-review.

Use when implementation is complete — examines code for correctness, maintainability, and spec compliance.

Use when review is approved and before QA — checks for vulnerabilities, exposed secrets, and insecure patterns.

Use when executing an approved implementation plan — dispatches fresh subagents per task with two-stage review (spec compliance then code quality).

TDD development with Red-Green-Refactor cycle. Use to implement a feature by writing tests BEFORE the code. Trigger automatically when the user asks for TDD, wants to write tests first, mentions "test first", or asks to implement, add, create, fix, correct code, a new feature, a bugfix, or a functionality.

Use when creating a new Atelier skill or rewriting an existing one — structures authoring, runs a self-review gate, and verifies with a subagent before registering.

Internal — routes Atelier operational-state CRUD through Memex Core (insert / update / query / register-role / register-agent / get-agent). Bypasses the Librarian; pure SQL. Not user-visible.

Internal — Tier 2 (structured-row) Atelier writes through memex:index:write with a caller-built librarianoutput. No Librarian subagent dispatch, no LLM. Not user-visible.

Convene a meeting of AI personas (3 to 10 participants) who debate a subject and reach a synthesis. Teams adapt to the theme (dev, design, product, business, life). The user can mix teams, add custom personas on the fly, and size the meeting up or down. Full debate written to a markdown file in the current directory, only the Boss's final synthesis is shown in the console.

Internal — one-shot per-project migration from Local-mode atelier.db to machine-global Memex. Called only when modedetector returns memex AND shouldprompt returns True.

Plan-phase wave 0 (sub-agent team mode) — parallel specialist reads that feed the wave-1 planner synthesis. Read by the PM; not a user-facing slash command.

Plan-phase wave 1 (sub-agent team mode) — planner synthesis that consolidates wave-0 field-analysis docs into the task list. Read by the PM; not a user-facing slash command.

PM wave-dispatch engine (atelier60) — operator-facing procedure for the mode-agnostic wave-barrier scheduler that partitions tasks into ordered waves, enforces the attempt budget + per-attempt wall-clock, validates reply envelopes, and gates each wave on terminal-only closure. Read by the PM/orchestrator; not a user-facing slash command.

Internal — consent-gated, once-per-version offer to apply one of atelier's recommended settings PROFILES (cost-effective | code-quality) to ~/.claude/settings.json on a plugin version upgrade. Triggered when startupcheck() returns a settingsrecoffer.

Internal — Local-mode CRUD for operational state (sessions, tasks, phase bypasses). Direct SQLite, no Librarian.

Team-mode hard rules — atelier multi-party channel contract. Read by scripts/dispatch.py and prepended verbatim to every worker briefing.

Internal — Local-mode raw-body archive helper. Content-addressed disk store under /.ai/raw/.