Deep research

/deep-research <question> starts a sourced research task in the current conversation. It is a bundled instruction skill: Kin reads relevant sources, checks the evidence, and writes a cited answer you can discuss or resume later.

/deep-research What license is Textual released under, and who maintains it?

Search and context retrieval require hosted-web access or a local Brave key. Kin can also read an exact source with web_fetch. If evidence is unavailable, the skill instructs it to say what remains unverified.

For a question about the current codebase or repository material, use /research <question>. It grounds the answer in local source and names the files and symbols that support it.

Work that fits the question

Kin identifies the scope, prefers primary sources, checks dates and conflicting claims, and separates observed facts from inference. A focused question may need only a few reads. A broader question may benefit from independent researcher children; these use normal agent supervision. There is no fixed number of research lanes, mandatory reviewer panel, or script-owned phase sequence.

The answer belongs to the ordinary conversation. Tool calls follow normal permissions, and mutating actions still use the session’s approval policy.

Citation checks

Before the final report, the skill instructs Kin to pass every cited URL to cite_check, in batches when necessary:

• Keep a live URL after confirming that its page supports the associated claim.
• Read an archived substitute before citing it, and identify it as archived.
• Drop a dead, unarchived citation and explain any resulting evidence gap.
• If a probe fails, disclose that verification failed and mark the affected citations unverified.

cite_check remains a deterministic READ tool. It checks up to 32 URLs per call using HEAD, with a body-less GET fallback, and looks for a Wayback snapshot when a source is dead. It applies the same SSRF protections as web_fetch and returns each URL’s live, wayback, or dead status.

The probe establishes whether a URL resolves; reading the source establishes whether it supports the claim. The skill asks Kin to do both. This is ordinary tool use guided by instructions, so the final report should show its evidence and limits. There is no script that automatically filters every model answer.

Model loading and saved workflows

deep-research-inline remains a model-loadable research skill. /deep-research is a human-invoked procedure. Both work in the normal conversation; choose the slash command when you want to request the procedure explicitly.

For saved Python skills from an earlier release, see Workflow migration.