kitsdeals-river

Python SDK for consuming the Kit's Deals river — a real-time, structured stream of approved deals built for agent consumers.

The interaction model is agent-mediated. You're not expected to edit YAML by hand. You tell your Claude/Code agent something like "set up github.com/kitsdeals/kitsdeals-river for me — I'm looking for TVs and a laptop refresh", and the agent runs the onboarding skill (PR 37) to translate the conversation into a ~/.kitsdeals/profile.yaml plus a running watcher daemon. The CLI commands and library API exist so the agent has reliable surfaces to call; humans rarely use them directly.

Architecture

┌──────────────────────────────────────────────────────────────────┐
│  RiverWatcher                                                    │
│    SSE consumer (Kit's Deals river v2)                           │
│    ↓                                                             │
│    Reconnect with exponential backoff                            │
│    Cursor persistence (~/.kitsdeals/cursor.json)                 │
│    Two-step catch-up: REST for gaps, SSE for forward             │
│    ↓                                                             │
│  Handler  (your code or the spawn_claude_evaluator from PR 37)   │
│    ↓                                                             │
│  Profile           DecisionsLog        Notifier                  │
│  ~/.kitsdeals/     ~/.kitsdeals/       telegram / stdout / …     │
│   profile.yaml      decisions.jsonl                              │
└──────────────────────────────────────────────────────────────────┘

Status

Install

pip install kitsdeals-river

(Pre-publish: pip install -e . from this directory.)

Agent-driven setup (the expected install flow)

Paste this repo URL to your Claude Code agent and say:

Set up Kit's Deals for me — I'm looking for X and Y.

The agent loads skills/onboarding.md, runs a short interview (4–6 questions max), translates your answers to a profile, and starts the watcher. You don't see the YAML.

For ongoing changes — "stop showing me TVs", "I bought one of those", "only show me clear wins" — your agent loads skills/update-profile.md and translates the natural-language request into a structured CLI edit. Live reload via SIGHUP, no restart required.

CLI

The CLI is the agent-facing API to the profile and the watcher daemon. Humans rarely run these directly; the skills tell the agent how to use them.

kitsdeals-river setup --json '<full profile JSON>'      # write profile to disk
kitsdeals-river run                                     # foreground watcher (start here, or via systemd)
kitsdeals-river run --initial-backfill                  # also surface currently-live matching deals on first start
kitsdeals-river status                                  # daemon health snapshot
kitsdeals-river reload                                  # SIGHUP a running watcher
kitsdeals-river profile show [--json]                   # inspect current profile
kitsdeals-river profile update --remove-watch "TVs"     # programmatic edits
kitsdeals-river profile validate                        # well-formedness check

setup writes the profile only — it does not start the watcher. run (or a systemd unit invoking run) is what actually opens the SSE connection. Treat them as separate concerns: one configures, the other keeps the daemon alive.

See skills/onboarding.md and skills/update-profile.md for the agent- facing recipes.

Multiple watches

Profiles can declare multiple watches (profile.watches[]). When more than one is present, kitsdeals-river run opens one SSE connection per watch concurrently in a single process — each with its own filter, its own cursor file (<cursor>-<sanitized-label>.json), and its own notify_threshold. A deal that matches multiple watches is delivered on each matching connection; the default handler dedupes by deal id and routes to the first matching watch's threshold.

Notify thresholds

Each watch declares a notify_threshold ∈ {anything, good, clear_win} that gates which matched deals actually fire a notification. Mapping against the server's deal-quality score (deal_score, 0–100):

Deals lacking deal_score only notify when notify_threshold is anything — refuse-to-guess posture.

Telegram delivery setup

The default notifier reads TELEGRAM_BOT_TOKEN and TELEGRAM_CHAT_ID from the environment. Two prerequisites:

1. Create a bot via Telegram's @BotFather: send /newbot, pick a display name and username, save the token BotFather replies with.
2. The user must DM the bot first. Telegram bots cannot send the first message — the user has to open a chat with their new bot and send /start (or any message). Skip this step and the watcher runs fine but sendMessage returns Forbidden: bot can't initiate conversation with a user.

Then retrieve the chat_id (GET https://api.telegram.org/bot<TOKEN>/getUpdates returns result[*].message.chat.id) and put both values in the environment file your service reads. For full step-by-step including verification and forum-topic routing, see skills/onboarding.md — the agent-facing version walks through every failure mode.

The default notification format includes brand, product name, current

• original price, discount %, merchant, condition, deal score, optional quality verdict, the product URL (Telegram unfurls into a card), and an expiry countdown when known.

Minimal example (without an agent in the loop)

from kitsdeals_river import RiverWatcher, RiverEvent, StdoutNotifier
from kitsdeals_river.profile import WatchFilter
import asyncio

async def main():
    notifier = StdoutNotifier()
    watcher = RiverWatcher(
        filter=WatchFilter(
            category="tv",
            brand_in=["LG", "Sony"],
            min_discount_pct=25,
        ),
    )

    @watcher.on_event
    async def handle(event: RiverEvent):
        if event.is_deal:
            deal = event.data
            await notifier.send(
                f"{deal['brand']} — ${deal['current_price_cents']/100:.2f}",
                deal,
            )

    await watcher.run_async()

asyncio.run(main())

See examples/minimal.py for a runnable version.

Testing

pip install -e ".[dev]"
pytest

License

MIT — see LICENSE. Use freely, build on top, fork it for your own personal-event runtime.

Roadmap

This package is one channel of a broader pattern: decoupled summarization — a centralized service does expensive synthesis work once, emits a structured event stream, and agents do cheap per-event filtering against a user profile. Same architecture applies to news, jobs, regulatory filings, GitHub releases, real-estate listings. Kit's Deals is the first concrete implementation; the SDK's profile + watcher + decisions-log shapes are designed to translate.