research-search

name

research-search

description

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.

Research Search — Searcher Agent with Candidate Crossover

This skill executes a single phase of the Searcher role from the Deep Researcher Reflect Evolve architecture. Same per-topic working directory as /deep-research:

.research/<slug>/
├── plan.md
├── context.md
└── progress.json

Argument parsing

First arg is the mode. Second arg is an optional topic slug. For mode answer, the third arg is an explicit query (otherwise prompt for one or take the most recent query from context.md).

If no slug is given and exactly one .research/* directory exists, use it. If multiple, ask the user which.

Modes

query [<slug>]

Generate the next-best search query for the active session.

Spawn a fresh general-purpose subagent. Brief:

You are the Searcher in query-generation mode. Read <plan_path> and <context_path>. Identify the single highest-value unanswered sub-question — one whose answer would most increase research progress AND is not redundant with any query already logged in <context_path>. Output exactly one line to stdout: a concise web-search query string (no quotes, no markdown). Return only the query.

Show the returned query to the user.

answer [<slug>] <query>

Answer a search query using the Candidate Crossover algorithm.

Step A — Spawn 3 candidates in parallel

In a single message, issue three Agent tool calls (subagent_type: general-purpose). Since Claude Code does not expose per-call temperature / top_k, diversity comes from lenses instead — this preserves the paper's intent of exploring a larger search space:

Candidate	Lens	Optimization
A	Breadth	Widest set of relevant facts, numbers, sources. Coverage > depth.
B	Depth	Mechanism, causation, technical detail on the central aspect. Insight > breadth.
C	Skeptic	Contradicting evidence, edge cases, recency, source quality. Fact-check.

Each candidate's brief (substitute the lens row):

You are Candidate <A|B|C> of the Search Agent. Query: <query>. Lens: <lens optimization>. Use the WebSearch tool to gather current information. Filter results: keep only sources you would rate ≥30% relevance — drop the rest. Return a concise answer (≤400 words) preserving every fact, number, date, and source URL you found relevant. Do NOT write to any file. Output only your answer text.

Step B — Crossover synthesis

After all three candidates return, fork yourself (Agent without subagent_type so it inherits this conversation's context). Brief:

You are the Crossover synthesizer for query <query>. Below are three candidate answers. Merge them into a single consolidated answer that retains every distinct fact, number, date, and URL — deduplicated, with contradictions explicitly flagged (⚠ conflict:). Then append exactly this block to <context_path>:

## Iteration <N> — <query>

<consolidated answer>

**Sources:** <bulleted list of unique URLs>

---

Use the next iteration number — read existing iteration headers in <context_path> and increment. Return when the append is complete.

Candidate A (Breadth):

Candidate B (Depth):

Candidate C (Skeptic):

Step C — Report back

Tell the user the query has been answered and appended to <context_path>. Optionally show the consolidated answer text inline if the user is reviewing iteratively.

Notes on faithfulness to the paper

• The paper uses n=3 candidates with varied temperature and top_k. Claude Code does not expose those per call, so diversity comes from prompt-level lenses (Breadth / Depth / Skeptic). This preserves the intent — exploring a broader search space — but the mechanism is different.
• The paper omits TTD-DR's Environmental Feedback (auto-rater) and Revision (iterative critique) steps to reduce latency; this skill matches that choice.
• The paper uses Tavily web search with a 30% relevance score filter; here, candidates apply the threshold via judgment over WebSearch results.
• The paper takes top-5 web results; candidates here should aim for similar coverage, but final filtering is by relevance not rank.