lunasearch — Cheap Parallel Discovery Lane (depends on cxc-search)

cxc-lunasearch is a dependent tool of cxc-search, not a standalone search skill. It fans out cheap Luna-model subagents for wide discovery, then hands every candidate back to the main agent, which runs the cxc-search proof ladder (Tier 1 discover, Tier 2 open-the-source) to settle claims. Luna discovers; cxc-search proves; the main model synthesizes. Never use lunasearch without this handoff — Luna snippets are leads, not evidence.

Hardcoded Spawn Path (no catalog probe)

The session surface is pinned on its first turn. V1 is the default unless the model catalog selects V2 (sol/terra; luna stays V1) or features.multi_agent_v2 selects it for a fallback model. If spawn_agent is not visible on V1, tool_search for it first (structure/60_native_capabilities.md §1). Fan the lanes out as N spawns before waiting; V1 wait_agent returns final status plus content, while V2 wait_agent is a no-content mailbox. Reuse a lane with V1 send_input(agent_id) or V2 followup_task(task_name). V1 also has close_agent/resume_agent; V2 has only interrupt_agent. The concurrency limits are V1 agents.max_threads (default 6) and V2 max_concurrent_threads_per_session (default 4, root included).

The user of this skill is already running on Luna, so do not call catalog_list or any model-picker probe before spawning. Hardcode the model directly on every spawn call:

agent_type: "explorer"
model: "gpt-5.6-luna"
reasoning_effort: "low"

If the spawn call returns a model-not-found / invalid-model error, do not retry with a probe. Fall through to serial dispatch immediately: re-issue the same spawn without the model field so the subagent inherits the main session model. State plainly which path each agent took. No silent fallback to 5.5 and no catalog round-trip — the error itself is the signal, and the serial retry is the recovery.

Default reasoning_effort is "low" — Luna lanes are cheap discovery, not deep reasoning. Keep final judgment in the main session regardless.

Subagent Skill Attachment (attach cxc-search, not prose)

Do not hand-write a tool directive in the spawn message. Attach cxc-search through the preferred [$cxc-search](skill://<abs SKILL.md path>) form, or the plugin-native $codexclaw:cxc-search fallback when the path is not link-safe, so each Luna subagent can load the proof ladder where the surface delivers the skill (Tier 1 web_search + Tier 2 open-the-source) at launch. The skill body is the single source of truth for the tool list; this skill only adds the lane assignment and the Luna model.

The shared payload form is a link-form mention in the spawn message. V1 parses it on the child's first turn. On plaintext V2 provider/proxy paths, the codexclaw spawn hook inlines the recognized skill's full SKILL.md body; native ChatGPT-backend V2 gives the hook ciphertext, so normalization and inlining are no-ops there. When no body can be inlined, the hook instead appends a plaintext [CXC-SKILL-AFFORDANCE] block telling the child to self-load any $cxc-<folder> / $codexclaw:cxc-<folder> mention from <skillsDir>/<folder>/SKILL.md; fork inheritance remains a secondary channel:

message: "[$cxc-search](skill://<cxc-search SKILL.md absolute path>)
TASK: one lane in a Luna search swarm. LANE: <source class / query family>. Run 5-10 distinct queries; open the source for every result that matters. Return 3-5 findings with URLs, dates, source type, primary-or-lead flag. No edits, no questions."

On the v1 surface the structured items channel is equivalent (exact selection) when routing through the spawn-wrapper builder:

items: [
  { type: "skill", name: "cxc-search", path: "<cxc-search SKILL.md absolute path>" },
  { type: "text",  text: "TASK: one lane in a Luna search swarm. LANE: <source class / query family>. Run 5-10 distinct queries; open the source for every result that matters. Return 3-5 findings with URLs, dates, source type, primary-or-lead flag. No edits, no questions." }
]

(v2 deny_unknown_fields rejects items; the hook-inlined attachment applies only on plaintext V2 paths.)

Do not duplicate the Tier 1/2 tool list as inline prose — the attached skill already carries it. A subagent that cannot open pages must flag every finding as candidate — unverified snippet in its return.

Use Case

Use Luna search when breadth matters and each subtask can be narrow:

• release/news/changelog sweeps across many vendors
• competitor or ecosystem scans
• "find many sources first, judge later" research
• Korean requests such as 루나검색, luna로 5개 돌려봐, 병렬 웹검색, 싸게 많이 찾아봐
• workflows where a 5.5 main session should conserve quota by delegating source discovery to Luna

Do not use it for local repository grep, one-source latest/current facts, implementation work, or high-stakes final advice without primary-source proof.

Swarm Shape

Default to five explorer subagents. Use three for smaller research and two for corroboration only.

Assign one distinct lane per agent. Do not send duplicate prompts — rewrite the user request into query families, then give each agent one family:

• official docs/changelogs (site:<docs-domain> changelog)
• vendor blogs/release notes (site:<blog-domain> "release notes")
• GitHub releases/issues/discussions (site:github.com <topic>)
• standards/specs/API references (filetype:pdf <topic> spec or intitle:specification <topic>)
• independent reports, benchmarks, community findings (site:reddit.com OR site:news.ycombinator.com <topic> after:<date>)

Search English first — it is the largest authoritative corpus. Add a local-language sweep only when the topic is inherently local or the user asks for sources in a specific language.

Spawn Contract

Each Luna subagent gets: (1) the cxc-search mention in its message and (2) a short task naming its lane (see the attachment section above). V1 may use structured items when the caller supplies that channel manually. The skill carries the tool list and proof rules; the task carries only the lane assignment and return shape. No five-part hand-written message — the attached skill is the tooling contract.

Spawn all lanes in one turn — parallel, not sequential. Report the spawned agent ids/nicknames to the user. The runtime may choose nicknames; do not claim manual naming unless the spawn tool supports it.

Proof Handoff (to cxc-search)

Luna output is candidate evidence only. After the swarm returns, the main agent runs the cxc-search proof ladder on the strongest candidates:

1. Build a compact claim ledger:
   • claim, source URL, date, source type, Luna lane, status
   • status: candidate, verified, contradicted, or unreachable
2. Open primary sources (cxc-search Tier 2) before final synthesis. Prefer official docs, release notes, source repositories, specs, and original announcements.
3. When sources conflict, state which source wins and why. Do not average.
4. Mark snippet-only or unreachable items as unverified leads.

For a high-risk non-code claim (price, market share, dated, causal), require

=2 independent source domains plus a counter-search before promoting it to verified — the ultraresearch claim-ledger gate, applied lightly.

Final Report

Return compactly:

1. Spawn path: hardcoded Luna used, or serial-fallback after error.
2. Swarm: number of agents and lanes.
3. Verified findings: source-opened claims only.
4. Open leads: promising but unverified Luna results.

Never treat Luna snippets or subagent summaries as final proof.

Gap note (vs lazycodex ultraresearch)

This skill is intentionally lighter than lazycodex ultraresearch. It does not run the EXPAND convergence loop, keep a session journal, verify by executing code, or generate reports — those belong to cxc-search Tier 3. lunasearch is the cheap one-shot discovery fan-out; ultraresearch is the deep multi-wave research protocol. Use lunasearch when breadth-for-cost is the goal; escalate to cxc-search Tier 3 when the question needs iterative expansion and contested-claim verification.