/adapter

description: "Route to, register, inspect, validate, and run explicitly trusted local private adapters without putting their code or bulk data in the wiki."
argument-hint: "route --intent <effect> --resource <url>|add <path>|list|show <id>|doctor <id>|run <id> --request <json>|remove <id> --yes"
allowed-tools: Read, Write, Edit, Glob, Grep, Bash(ls:*), Bash(date:*), Bash(python3:*), Bash(scripts/llm-wiki:*), Bash(${CLAUDE_PLUGIN_ROOT}/bin/llm-wiki:*)

Your task

Manage or invoke a private llm-wiki adapter through the deterministic bundled CLI. Read `skills/wiki-manager/references/adapters.md` before acting.

Private-adapter management is wiki-neutral. Do not put executable registrations or absolute machine paths in `wikis.json` or a topic wiki. Resolve a topic wiki only when the user asks to promote a reviewed result after execution.

Declarative intent routing

Before treating an external URL as an ingestion source, normalize the requested effect to a lowercase intent token and run:

${CLAUDE_PLUGIN_ROOT}/bin/llm-wiki adapter route \
  --intent <effect> --resource '<external-url>' --json

On `matched`, read the returned adapter-owned guide, then run `show` and `doctor` for the returned adapter id before following that workflow. The private adapter owns all provider-specific authentication, transport, planning, recovery, and verification instructions. On `no-match`, continue with normal wiki routing. On `ambiguous` or `unavailable`, fail closed and repair the registration instead of guessing an adapter or falling back to an ungoverned write path. Route output does not echo the resource.

A URL by itself never authorizes invented edits. A bounded imperative approves only a faithful plan; generic remote-write plan hashes, revision locks, idempotency keys, private receipts, and read-back verification remain required.

Locate the CLI

For Claude Code, use:

"${CLAUDE_PLUGIN_ROOT}/bin/llm-wiki" adapter <subcommand>

In a source checkout, use `scripts/llm-wiki`. Other runtimes should use the bundled `bin/llm-wiki` relative to the installed plugin root described in the adapter reference. Do not assume the command is globally installed.

Management operations

• `add <path>`: require an existing local checkout and explicit read/write

roots. Add exact `--remote-resource` values only when the user authorized them. Never clone a repo or request credentials. Use `--replace` only when the user intends to trust a changed manifest.

• `list`: show compact id, version, capability, network, and local-root data.
• `show <id>`: show the machine-local registration; do not print environment

variable values.

• `route --intent <effect> --resource <url>`: match only registered declarative

routes and return the adapter-owned guide without echoing the resource.

• `doctor <id>`: verify manifest hash, executable, and handshake before a run.
• `remove <id> --yes`: remove only the registration. Never delete adapter code,

inputs, or outputs.

Run operation

1. Inspect the request JSON without opening referenced corpora or secrets. 2. Confirm the adapter id, operation, declared paths, and requested output directory match the user's intent. 3. Run `doctor`; stop on any issue. 4. For `remote-write`, confirm the user explicitly approved the exact plan, expected revision, target resource, and idempotency key. Invoke with `--approve-remote-write <plan-sha256>` and a private `--response` path inside the registered write root. Never infer approval from a request file alone. 5. Invoke `adapter run <id> --request <absolute-path> --json`. 6. Report the run id, bounded summary, and artifact counts by class. 7. Do not import anything automatically.

If the user asks to update the wiki, inspect only `wiki-safe` artifacts. Review them as candidates, keep `private` and `bulk` artifacts external, write the smallest useful raw note/evidence packet, compile bounded conclusions, update indexes, and append the normal topic and hub logs.

Private visibility is an access control, not permission to redistribute data or perform protected-content collection, deanonymization, sensitive inference, targeting, harassment, or other prohibited analysis.