ddx

Operates the DDx toolkit for document-driven development. Covers beads (work items), the queue, executions, agents, harnesses, personas, reviews, spec-id. Use when the user says "do work", "drain the queue", "run the next bead", "execute a bead", "review this", "review with fresh eyes", "fold in this guidance", "review again", "break this down into specs and beads", "make this testable", "check against spec", "what's on the queue", "what's ready", "what's blocking the queue", "create a bead", "file this as work", "run an agent", "dispatch", "use a persona", "how am I doing", "ddx doctor", or mentions any ddx CLI command.

Operates the DDx toolkit for document-driven development. Covers beads (work items), the queue, executions, agents, harnesses, personas, reviews, spec-id. Use when the user says "do work", "drain the queue", "run the next bead", "execute a bead", "review this", "review with fresh eyes", "fold in this guidance", "review again", "break this down into specs and beads", "make this testable", "check against spec", "what's on the queue", "what's ready", "what's blocking the queue", "create a bead", "file this as work", "run an agent", "dispatch", "use a persona", "how am I doing", "ddx doctor", or mentions any ddx CLI command.

Assess readiness, score, classify, and refine ddx beads. Used by ddx try/work hooks before claim or dispatch, after failed attempts, and for operator-invoked refinement.

Run an adversarial review of a plan, spec, bead, or PR by dispatching to multiple harnesses with adversarial framing and aggregating findings. Use when you want pressure-testing against a single-model blindspot, high-stakes review, or multi-round critique before committing to a direction.

Break a spec, plan, epic, or free-form task into execution-ready beads by filing `ddx bead create` calls with inline context, acceptance criteria, and dependency wiring. Use when a governing artifact, approved plan, or epic needs to be translated into a queue of actionable beads.

Run a repeatable prompt or task benchmark across a matrix of agent conditions, aggregate comparable outputs, and produce a concise report. Use when evaluating prompt changes, model/power choices, harness passthrough constraints, or regression behavior across a suite. Composes `ddx run` and the compare-prompts workflow; replaces benchmark behavior that should not live in core CLI routing.

Dispatch the same prompt across N agent invocations (different harnesses, models, personas, or prompt variants) and aggregate the results into a single comparison report. Use when you need to benchmark prompt variants, run a quorum-style consensus check, compare model outputs side-by-side, or replay a prompt under altered conditions. Replaces the removed core quorum flag and benchmark CLI.

Estimate effort for a bead, spec, or task by dispatching to one or more harnesses and aggregating structured complexity, time, and risk findings. Use before scheduling work, breaking into beads, or prioritizing a backlog.

Re-run a previously executed bead under altered conditions (different harness, model, profile, persona, or base revision) and diff the new attempt against the prior baseline. Use when investigating why an attempt failed, A/B-testing a model upgrade on a known task, or sanity-checking a regression. Replaces the deprecated replay CLI from the old task-dispatch namespace.

Support DDx prose work by preserving voice, improving execution-useful clarity, applying deterministic prose findings with judgment, and keeping DDx terminology, headings, tables, paths, commands, and legitimate lists intact.

Route HELIX methodology work to the right planning, alignment, design, review, execution, or release workflow. Use when the user asks to use HELIX, work with HELIX artifacts, align documents, frame requirements, design a change, evolve specs, review work, decide what is next, or manage HELIX-governed work without naming a specific helix-* skill.

Generate or edit images via OpenRouter with the Gemini 3 Pro Image model. Use for prompt-only image generation, image edits, and multi-image compositing; supports 1K/2K/4K output.

Aggressively refine a draft plan, design, or proposal through adversarial multi-reviewer critique and automatic folding. Returns a sharper plan, not a pile of review notes. Invoke when the user asks to "review [this/the] plan", "review with codex", "fresh eyes review", "re-review", "refine this plan", "sharpen this plan", "pressure-test this", "get a second opinion", "critique this", "what does codex think", "peer review", or any phrasing that asks for external scrutiny and improvement of a draft. Runs fresh-eyes (general-purpose subagent) and codex (ddx agent run) in parallel, folds findings back into the plan between rounds, loops until convergent or until a reviewer surfaces a question that needs user input.