[Feature]: Support using OpenWiki without an API key by exporting it as Claude Code skills

[Kowshik4593]

Summary

Problem

Today, running OpenWiki on a repository requires configuring an OpenWiki model
provider and API key — OpenWiki's CLI is the LLM host and runs its own agent
loop. That's a barrier for a large group of users who already have a capable
coding agent (Claude Code) in their editor and simply want repository docs
generated and maintained, without setting up and paying for a separate
inference provider.

Proposal

Add a first-class, spec-kit-style subcommand:

openwiki integration claude [path]

It scaffolds two Claude Code skills — openwiki-init and openwiki-update —
into .claude/skills/ in the target repository (defaulting to the current
directory). The user then runs /openwiki-init / /openwiki-update inside
Claude Code, and Claude Code itself generates and maintains the openwiki/
docs. No OpenWiki model provider or API key is required — Claude Code is the
engine.

Scope

• In scope: code mode only (repository docs under openwiki/).
• Out of scope: personal mode and connectors (Gmail/Slack/Notion/X), which
  depend on OpenWiki's credentialed runtime.

Approach

• The skill bodies are bundled, maintained templates derived from the
  repository-mode instructions in src/agent/prompt.ts, with the tool
  vocabulary remapped to Claude Code natives (Read/Write/Edit/Glob/Grep/Bash),
  connector/CLI/local-wiki text removed, and virtual-root /openwiki/... paths
  converted to real relative openwiki/... paths.
• This never touches the live agent's prompt path, keeping regression risk off
  the running agent.
• The subcommand sits alongside auth/cron/ingest/ngrok and is built on a
  small integration registry so other host agents can be added later.

Acceptance criteria

• openwiki integration claude writes
  .claude/skills/openwiki-init/SKILL.md and .../openwiki-update/SKILL.md
  with valid Claude Code skill frontmatter.
• openwiki integration claude <path> targets that repository.
• The command requires no model provider or API key.
• Skill bodies use Claude-native tools and real relative openwiki/ paths;
  no connector-only text or virtual-root paths.
• Unknown/missing agent prints a clean error and exits 1.
• Existing OpenWiki commands are unaffected.
• Unit tests cover parsing, scaffolding, idempotency, and template content.

Notes / trade-off

The skill templates are a hand-maintained copy of the repository-mode prompt, so
they can drift when prompt.ts changes. A guard test pins the core discipline
sections to make an incomplete re-port fail loudly, but semantic sync remains a
manual, periodic responsibility.

Motivation

Running OpenWiki on a repository today requires configuring an OpenWiki model
provider and API key — OpenWiki's CLI is the LLM host and runs its own agent
loop. That's real friction for a large group of users who already pay for
inference through a capable coding agent (Claude Code Pro/Max/Team) sitting
right in their editor. They don't want a second key or a second bill; they just
want openwiki/ docs generated and kept fresh.

This is the same motivation raised in #156, and the same one #151 addresses
from the OpenAI/ChatGPT-subscription side. OpenWiki's real value is its
methodology and on-disk format (openwiki/quickstart.md, section pages,
.last-update.json, the AGENTS.md/CLAUDE.md pointer block) — not the
specific runtime that produces it. That format can be produced directly by a
host coding agent, keylessly, without going through OpenWiki's LangChain engine
at all.

Proposed Solution

Add a first-class subcommand:

openwiki integration claude [path]

This scaffolds two Claude Code skills — openwiki-init and openwiki-update
— into .claude/skills/ in the target repository (defaulting to the current
directory). The user then runs /openwiki-init / /openwiki-update inside
Claude Code, and Claude Code itself generates and maintains the openwiki/
docs. No OpenWiki model provider or API key is required.

Key design points:

• Bundled, maintained templates — the skill bodies are derived from the
  repository-mode instructions in src/agent/prompt.ts, with the tool
  vocabulary remapped to Claude Code natives (Read/Write/Edit/Glob/Grep/Bash),
  connector/CLI/local-wiki-only text removed, and virtual-root /openwiki/...
  paths converted to real relative openwiki/... paths.
• Zero regression risk on the live agent — this never touches prompt.ts
  or the running agent loop; it's a pure export step, gated behind its own
  subcommand.
• Extensible by construction — it sits alongside auth/cron/ingest/
  ngrok and is built on a small integration registry, so other host agents
  (Cursor, Copilot, etc.) can be added the same way later.
• Scoped to code mode — repository docs under openwiki/ only; personal
  mode and connectors stay on the credentialed runtime.

Alternatives Considered

No response

Additional Context

No response