fluffyjaws

name	fluffyjaws
description	Interact with Adobe FluffyJaws (Mr. FluffyJaws) — Adobe's internal AI assistant — via its public API. Use when the user mentions FluffyJaws, fj, FluffyPacks, the Adobe internal AI chatbot, the FluffyJaws CLI, fj-mcp, or wants to ask questions, stream chat completions, list/inspect FluffyPacks, list conversations, or talk to the FluffyJaws MCP endpoint without clicking through https://fluffyjaws.adobe.com. Activate on mentions of "fluffyjaws", "fluffy jaws", "fluffypack", "fluffypacks", "fj ask", "fj chat", "fj mcp", "ask FluffyJaws", "internal Adobe ChatGPT", or related Adobe internal AI workflows.
allowed-tools	bash

FluffyJaws

Use the bundled fj shell command instead of opening https://fluffyjaws.adobe.com.

Quick start

# Confirm you are signed in (uses the existing fluffyjaws.adobe.com browser tab).
fj me
# Verify the identity printed before issuing further commands.

# One-shot question. Streams to stdout, prints final answer + responseId.
fj ask "Summarize the last AEM Cloud Service release notes."

# Continue the same model thread with previousResponseId.
fj ask --continue resp_0fa5f24d3c... "Now turn that into three bullets."

# Use a specific FluffyPack (use `fj packs` to discover UUIDs).
fj ask --pack 2a0c713c-0767-4514-83af-93c9373f2f0a "What changed in Express last week?"

# Browse FluffyPacks.
fj packs --scope recommended
fj packs --q "incident" --scope discover --limit 10
fj pack 2a0c713c-0767-4514-83af-93c9373f2f0a

# Conversations.
fj convs --limit 20
fj conv <uuid>
fj search "release notes"

# Talk to the FluffyJaws MCP endpoint (JSON-RPC over HTTP).
fj mcp init
# Verify: confirm a session id was stored before continuing.
fj mcp tools
fj mcp call google_web_search '{"query":"AEM release notes"}'

# Local cached docs (no network).
fj docs            # list pages
fj docs api        # show full API doc

Authentication

fj authenticates via your existing fluffyjaws.adobe.com browser session cookie. It finds (or opens) a tab on that host and proxies every API call through it so the fjv3_session cookie is sent automatically.

If the session is expired you will see a redirect to adobe.okta.com. Re-sign-in once in the FluffyJaws tab (in the foreground browser) and retry.

Available commands

All commands print structured output. JSON commands print pretty-printed JSON; streaming commands print incremental text and a final summary line.

fj me

Prints the signed-in identity from /api/auth/me. Confirm the identity is correct before issuing fj ask or other commands.

fj ask <prompt> [options]

Stream a single chat turn. Prints assistant text as it arrives, then prints a final line responseId=resp_... you can pass back as --continue later.

Options:

• --model <id> — model to use (default gpt-5.4)
• --pack <uuid> — route through a FluffyPack
• --continue <responseId> — continue an earlier response (multi-turn)
• --reasoning <low|medium|high> — reasoning effort
• --no-search — disable webSearchEnabled
• --no-canvas — disable canvasMode
• --raw — print raw SSE events instead of pretty text
• --json — emit one JSON line per SSE event

fj chat <prompt> [options]

Alias for fj ask.

fj packs [options]

List FluffyPacks via /api/v1/fluffypack/list.

Options:

• --scope <all|builtin|discover|mine|recent|recommended> (default all)
• --q <query>
• --limit <1-100> (default 24)
• --offset <n> (default 0)
• --json — full response JSON instead of compact one-line summary

fj pack <uuid>

Fetch a pack's detail document via /api/v1/fluffypack/:uuid.

fj explore [--q <query>]

Returns the four grouped pack lists (recommended, recent, builtin, discover) the UI uses for the picker.

fj convs [--limit <n>] [--offset <n>]

List conversations.

fj conv <uuid>

Get a conversation envelope.

fj search <query> [--limit <n>]

Search conversations.

fj mcp <subcommand>

Direct JSON-RPC client for /api/v1/mcp. Stores the negotiated Mcp-Session-Id in ~/.config/fj/mcp-session.txt so subcommands chain.

Subcommands:

• fj mcp init — call initialize, store session id. Verify the session id was stored (fj mcp session) before calling tools.
• fj mcp ping — call ping
• fj mcp tools — call tools/list
• fj mcp call <name> <args-json> — call a tool, prints JSON result
• fj mcp shutdown — end the MCP session
• fj mcp session — print the stored session id

fj docs [page]

Print the bundled, locally cached FluffyJaws documentation (no network).

Pages: index, api, mcp, python, register-app, fluffypacks, fluffypack-builder, slack-channels. With no argument, lists the available pages.

Streaming output format

fj ask reads the SSE stream and renders friendly output by default:

• Each response.output_text.delta event is written to stdout as the delta.
• Tool lifecycle events (tool_executing, tool_complete) print a single status line to stderr, e.g. [tool] full_documentation_search ….
• On response.completed, a trailing \n and responseId=resp_... line are printed to stderr.
• On response.failed or error, the error is printed to stderr and fj exits with status 1.

--raw skips parsing and dumps every SSE line. --json emits one parsed JSON event per line for piping into jq.

Troubleshooting

Symptom	Likely cause	Fix
Redirect to adobe.okta.com	Expired browser session	Re-sign-in at fluffyjaws.adobe.com, then retry
fj mcp call returns session error	MCP session timed out or never initialised	Run fj mcp init, confirm session id with fj mcp session, then retry
fj pack <uuid> returns 404	Invalid or inaccessible pack UUID	Run fj packs or fj explore to find valid UUIDs
fj ask exits with status 1	response.failed or stream error	Check stderr for the error message; retry or adjust options

Endpoints reference

For the full endpoint catalog with request/response schemas, see references/endpoints.md.

Caveats

• All API calls go through the UI host (https://fluffyjaws.adobe.com), not api.fluffyjaws.adobe.com. Only the UI host accepts the cookie session you already have.
• Inline file uploads (input_file) and reusable fileIds are not currently exposed by fj ask.
• The /api/v1/conversation/* surface exposes only metadata reads and create. There is no public message-save route, so streamed turns via fj ask are not persisted in the FluffyJaws web UI's conversation history. Store the responseId chain yourself if you need durable history.
• Service-token / on-behalf-of-user auth is documented in references/endpoints.md but is not exercised by fj because the SLICC fetch sandbox cannot present the right Origin to api.fluffyjaws.adobe.com.