conduct

Purpose

Walks a feature through the complete SDLC flow by checking artifact state and directing the user to the correct next skill. Run the conduct workflow at any point to see where you are and what to do next (Claude /conduct; Codex $conduct).

Does NOT run other skills internally — it assesses state and directs. The user invokes each skill.

Session Model

Fresh LLM session per executed step. Every step — design through finish — starts without prior-step conversational context and reads its inputs from the persisted artifacts. The session ID/marker under .pipeline/ supports the current step only; it is reset at the next step boundary.

• Design phase (bootstrap → plan): Each step reads the approved artifact produced by its predecessor.
• Build phase (pipeline): The conductor drives the task loop. The selected host agent orchestrates each task by dispatching subagents. Subagent context is isolated and discarded — only a ~2-3 line summary returns to the orchestrator per task. No context compaction needed.
• Ship phase (manual-test → finish): Lightweight steps, context stays bounded.

Retries within the same step resume that step's session so partial work and failure context are retained. A later step never resumes it. With per-step provider routing, session identity is additionally provider-local: another provider starts fresh, and only a same-step retry on the same provider resumes.

The Flow

Step 1:  /bootstrap             → UNDERSTAND
Step 2:  /memory                → UNDERSTAND
Step 2.5: /assess               → UNDERSTAND (existing projects only — skipped for new)
Step 3:  /explore               → DECIDE (context, approaches, + decide product/technical TRACK)
Step 4:  Complexity Assessment   → DECIDE (classify S/M/L, determines which steps run)
Step 5:  Worktree setup          → DECIDE (create feature branch + worktree — all subsequent commits are isolated)
Step 6:  /prd                   → DECIDE (product-only PRD; PRODUCT track only — skipped on technical)
Step 7:  /architecture-diagram  → DECIDE (generate/update current-state diagrams; skipped for Small)
Step 7b: /architecture-review    → DECIDE (skipped for Small, lightweight for Medium — produces ADRs; precedes stories)
Step 8:  /stories               → DECIDE (from PRD FRs on product; technical stories on technical)
Step 9:  /conflict-check        → DECIDE (skipped for Small; root-routes kickback → prd|architecture|stories)
Step 10: /plan                  → DECIDE (technical implementation plan, grounded in the architecture + stories)
Step 10b: /coherence-check       → DECIDE (Medium/Large traceability gate after plan)
Step 10c: /plan-review           → DECIDE (Medium/Large source-grounded gate after coherence-check)
Step 11: /writing-system-tests  → BUILD (skipped for Small)
Step 12: /pipeline or /tdd      → BUILD (task completion and affected-test results; engine build_review is post-build review)
Step 13: Engine-native configured-verifier gate → BUILD (repository-configured aggregate verification)
       ── CHECKPOINT ──         → User reviews build output, can go back or continue
Step 14: /manual-test           → SHIP (validate stories, bug loop via /tdd — auto-skip for non-endpoint features)
       ── CHECKPOINT ──         → User reviews test results, can go back or continue
Step 15: /prd-audit             → SHIP (PRODUCT track only — audit shipped impl vs PRD FRs; GATE; skipped on technical)
Step 16: /architecture-review --as-built → SHIP (shipped code vs APPROVED ADRs; GATE — BLOCKED on an ADR violation)
Step 17: /finish                → SHIP (verify, review changes, present options, and commit the durable shipped record before completion)