kokochat-deeply-research

OpenClaw side of the KokoChat Deeply mini-app's deep-research course path.

The user is in a deeply-course conversation created from the "定制课程" sheet

with kind = "research".

Two-phase pipeline (this skill is Phase A)

KokoChat splits research course generation into two model passes:

• Phase A — this turn (the agent run you are in now): research only.

Search the web, narrate the process to the user in Chinese prose, and emit

one koko.deeply.research.notes fenced block at the end containing a

300–1200 字 synthesis plus a flat list of 5–20 cited sources. You do

not decide on course sections, write section titles, or assign sources

per section — those are Phase B's job.

• **Phase B — separate stateless inferOnce the client triggers right after

you finish**: the same agent, but no web tools available. Phase B reads

your notes block, decides how many sections to use, and emits the final

koko.deeply.research.outline JSON. You will never see Phase B run; just

hand off clean notes and Phase B will handle the rest.

This split exists because asking one turn to plan tools, narrate prose, AND

emit a strict per-section JSON schema repeatedly burned attention budget on

the schema and led to "toolCallCount=0 but plausible-looking sources" — the

2026-05-28 hallucination regression.

When this skill fires

The first user message in the conversation looks exactly like:

> 请围绕「」做一份深度调研课程

(or with an explicit N 节 between 做一份 and 深度调研课程 when the user

chose a non-auto length in the sheet). When you see this shape, run the

procedure below. For any later user message in the same conversation, just

respond to the host's mainline prompt — KokoChat re-injects a fresh prompt on

every turn.

What the host sends with the kickoff

KokoChat injects a Phase A kickoff prompt right before the visible user line.

That prompt is the authoritative source of truth for **this turn's hard

constraints** (3 numbered rules), the koko.deeply.research.notes JSON

schema, and the prose sentinel 〔KP〕 requirement. **Always follow what the

kickoff prompt says.** Do not produce a koko.deeply.research.outline block

from this turn — that schema belongs to Phase B and emitting it here will be

ignored.

This skill file only adds background you need across all such turns: what

the search/fetch tools can / cannot do, and how to plan the research itself.

Tools

The deeply agent uses KokoChat's hosted search wrapper plus OpenClaw's built-in

web_fetch.

KokoChat hosted search (kokochat-search)

Run a web search via the local exec wrapper installed by KokoChat. The

KokoChat Runtime Contract in AGENTS.md gives the exact absolute path for

the wrapper; use that command shape:

<kokochat-search/bin/search.mjs absolute path> '{"query":"EN keywords","count":5}'

It calls KokoChat's hosted search proxy (Brave-backed in production), so the

user does not need to configure any search API key in their own OpenClaw.

Returns title / url / snippet per result.

Args:

• query (string, required): EN keywords. Translate Chinese topics into

English search keywords; keep user-facing prose Chinese.

• count (int 1–10, optional): default ~5; lower for targeted lookups,

higher when scoping broadly.

Pass only query and count. Provider-specific args like

domain_filter, date_after, freshness, country, language,

max_tokens will be rejected.

Tool quirks worth knowing:

• The wrapper may return { "ok": false, "error": "..." } if KokoChat's

hosted search is unavailable or rate-limited. Be honest; do not invent URLs.

• Don't batch multiple search calls in one assistant turn — issue them

sequentially across narration steps.

web_fetch

Pull one URL's main content as readable text. Useful when a search hit

looks unusually authoritative and you want the actual body instead of just

the snippet.

Args:

• url (string, required).
• maxChars (int, optional): prefer 60000 so the model sees enough body.

The Gateway may cap this lower.

Only fetch http:// / https:// URLs that came from a successful

KokoChat search. Never fetch file://, skill files, workspace docs, or URLs

you invented. If a fetch fails (network block, private-IP guard, challenge,

timeout), don't retry many other URLs — continue from search snippets.

Research Planning (generic; no hard-coded domain templates)

Before searching, infer a small research plan **from the user's topic

itself**. Don't insert fixed people, organisations, industries, or examples

that the user didn't mention.

Identify:

1. Key subjects: people / organisations / works / events / markets /

technologies / concepts named or implied by the topic.

1. Time and scope: year, geography, stage, version, comparison target,

or other limits.

1. Controversy structure: mainstream view, counter-view, boundary

conditions, common misreadings likely relevant.

1. Best evidence types: primary text, interview, transcript, filing,

paper, policy document, official data, high-quality secondary synthesis.

Queries should start with the exact topic and key subjects, then expand to

counterpoints / background / more authoritative original evidence. Avoid

filling sources with broad trend pages when narrower material exists. If a

topic is very narrow, prefer fewer real sources over more weak ones.

What is NOT in this skill any more

Schema, field rules, prose sentinel marker (〔KP〕), exact step-by-step

narration order, and the "3 hard rules" all live in the host-injected

Phase A kickoff prompt now. Treating those as the source of truth keeps this

file small and stops the model from being overwhelmed by duplicated, slightly

divergent instructions across two layers (which is exactly what caused the

2026-05-28 regression where the model emitted plausible-looking sources

without ever calling search).

The koko.deeply.research.outline schema (per-section sources, course

introduction, outline markdown) lives entirely in the Phase B prompt and is

not part of Phase A's output.