Use when encountering any bug, test failure, or unexpected behavior, before proposing fixes
Broad-to-narrow research Orchestration - start with a wide census, harden the scope from returned Evidence, and specialize Subagents only once the exact situation is proven. TRIGGER when the Architect says "progressive research", "start broad", or when dispatching research or debugging Subagents into a situation whose cause or shape is unknown. DO NOT TRIGGER when the exact scope is already proven - dispatch the specialized Subagent per /subagents directly.
Consolidate the knowledge wiki — revisit journaled topics, fold repeated findings into each topic's Claude.md, surface contradictions, keep cross-links true. TRIGGER on a scheduled pass or when a topic tree has grown messy or redundant. DO NOT TRIGGER to file fresh session material (that is journal), or to gather new research.
File a session's loose material into the knowledge wiki — each capture into its owning topic folder, the topic's Claude.md kept current. TRIGGER after a session hands you loose captures (notes, clippings, findings, screenshots) that need a home in the wiki. DO NOT TRIGGER to gather new research, or to consolidate the wiki itself (that is dream).
Operate Dent via first-party API. Trigger: Dent API, Funnels, Offers, Pages, Articles, Courses, Spaces, Analytics. Do not trigger for WordPress-only work.
Autonomous async Execution. The Architect hands off and walks away; the Agent drives the Task to done — makes every Architectural call via /trace + /pcc + ranking, executes through /subagents, verifies from each implementing Subagent's Evidence (they exercise their own flows, real browser for UI), and returns finished code plus the recorded Decisions for review. TRIGGER when the Architect says "code-first", "/code-first", "execute with /subagents", "execute directly & autonomously", "drive it and report back", "iterate until done", "I'm going to bed", "off to bed", or signals an AFK / overnight handoff. DO NOT TRIGGER when the Architect wants options before action — that is the default mode (propose via /pcc and wait).
Drive codex CLI runs as your Agents — you own Orchestration, codex does the work. One wrapper, `codex-run`, owns the mechanics (flags, output storage, stream parsing, failure detection); you write the Task Prompt and do the judgment. Calls /subagents for Orchestration doctrine on a codex-run substrate. TRIGGER when the Architect says "codex", "/codex", "codex-run", "use codex", "dispatch to codex", "run this through codex", "codex agents", "review with codex", "codex review", or asks to fan out work across codex runs. DO NOT TRIGGER for native Claude Code Subagents (use /subagents) — those run inside this Harness; codex is a separate CLI process.
Run N independent architect Agents debating Architectural options through rounds. TRIGGER when the Architect wants multiple competing solutions evaluated, when 3+ distinct approaches need comparison, or when the Prompt says "debate", "competing solutions", "run architects", "architecture debate", or `/debate [count] "problem"`.