claude-deep-researcher

Run the Deep Researcher Reflect Evolve loop to produce a comprehensive implementation plan for a given topic, searching both a target codebase and the web for SOTA approaches. Adapts the architecture from Prateek 2026 (arXiv:2601.20843) to planning: sequential investigation-plan refinement via reflection, candidate crossover for dual-source answers, and one-shot final plan generation. Use when the user asks for a "deep plan", an implementation plan that requires codebase understanding + SOTA research, or invokes /deep-plan. Args: <topic> [--codebase <path>].

Planner-agent operations for the Deep Plan system: create an initial investigation plan (analyzing both codebase and SOTA), reflect on the current plan against accumulated findings, update the plan, or score planning progress (0-100). Use when a planning session needs a one-off planner operation outside the full /deep-plan loop, or when the user invokes /plan-planner with one of the modes below. Args: <mode> [<topic-slug>] [--codebase <path>]. Modes: create, reflect, update, progress.

Plan-writer operation for the Deep Plan system: produce a one-shot, comprehensive implementation plan from an existing investigation plan and global research context. The deliverable is detailed enough for another AI agent to implement at a high bar of quality. Use when a /deep-plan session has accumulated enough findings and the user wants the plan generated immediately, or when the user invokes /plan-report. Args: [<topic-slug>].

Searcher-agent operations for the Deep Plan system: generate the next search query for an active planning session, or answer a query using the Candidate Crossover algorithm with dual-source search (web + codebase). Three parallel candidates with diverse lenses search both the internet and the target codebase, then merge into one consolidated answer appended to the global research context. Use for one-off search operations outside the full /deep-plan loop, or when the user invokes /plan-search. Args: <mode> [<topic-slug>] [<query>]. Modes: query, answer.

Run the Deep Researcher Reflect Evolve loop on a topic to produce a PhD-level research report. Implements the architecture from Prateek 2026 (arXiv:2601.20843): sequential research-plan refinement via reflection, candidate crossover for web search answers, and one-shot final report generation. Use when the user asks for "deep research", a long-form report, an investigation that requires iterative web search + planning, or invokes /deep-research. Args: <research topic>.

Planner-agent operations for the Deep Researcher Reflect Evolve system: create an initial research plan, reflect on the current plan against the global research context, update the plan, or score research progress (0-100). Use when a research session needs a one-off planning operation outside the full /deep-research loop, or when the user invokes /research-plan with one of the modes below. Args: <mode> [<topic-slug>]. Modes: create, reflect, update, progress.

Reporter-agent operation for the Deep Researcher Reflect Evolve system: produce a one-shot, PhD-level research report from an existing research plan and global research context. Use when a /deep-research session has accumulated enough findings and the user wants the final write-up generated immediately, or when the user invokes /research-report. Args: [<topic-slug>].

Searcher-agent operations for the Deep Researcher Reflect Evolve system: generate the next search query for an active research session, or answer a query using the Candidate Crossover algorithm (3 parallel candidates with diverse lenses, merged into one consolidated answer that appends to the global research context). Use for one-off search operations outside the full /deep-research loop, or when the user invokes /research-search. Args: <mode> [<topic-slug>] [<query>]. Modes: query, answer.