# RIFM

**Repository Path**: forantar/rifm

## Basic Information

- **Project Name**: RIFM
- **Description**: "Read It For Me" - per-topic news digests with local history and a browse UI
- **Primary Language**: Unknown
- **License**: MIT
- **Default Branch**: master
- **Homepage**: None
- **GVP Project**: No

## Statistics

- **Stars**: 0
- **Forks**: 0
- **Created**: 2026-09-09
- **Last Updated**: 2026-09-14

## Categories & Tags

**Categories**: Uncategorized

**Tags**: None

## README

# RIFM — Read It For Me

Per-topic news digests with local history and a browse UI.

## Usage

| Command | What it does |
|---|---|
| `/topics` | Numbered list of topics; reply with a number to read it |
| `/read <topic>` | Digest of **new** items since the last read |
| `/read <topic> full` | Full digest, ignoring read state (10-day window) |
| `/add` | Interview (topic → short name → sources) and register a new topic |
| `/remove <topic>` | Remove a topic (asks which one if no argument) |
| `/browse` | Open the digest browser page in your default browser |
| `/stop-server` | Stop the browse server (otherwise it exits after 30 min idle) |

A digest ends with a stats block (sources, duration, token usage measured
from the session transcript). Fetching and filtering are pure script
(`scripts/fetch.py`, no LLM); the LLM only selects and composes, so
incremental reads cost little. When nothing new appeared, you get a
one-liner followed by the newest stored digest. Every run that produced
items is stored dated, so history accumulates in
`.rifm/<topic>.<YYYYMMDD>.md`.

## Browse page

`/browse` builds a single self-contained page and serves it from a tiny
localhost server (`127.0.0.1:8642`, auto-started, exits after 30 min idle):

- home view lists every topic; a topic lists its digests by date,
  newest first; a digest shows its text with clickable URLs
- Home and Back buttons in the top-left corner of every view
- prev/next arrows above a digest page step through that topic's digests
  by date (each disabled at the end it can't go past)
- the "Read now" card on a topic runs `/read <topic>` headlessly
  (`claude -p`), shows live progress (sources N/M, then composing), and
  the tab refreshes itself with the fresh digest when it lands — no
  permission dialogs

The headless read uses `--permission-mode bypassPermissions` (it cannot
answer tool prompts) and only ever runs `/read` on a key that exists in
`topics.json`; its raw event log lands in `.rifm/last-read.log`.

## Layout

```
topics.json                 single registry: topics + their news providers
scripts/topics.sh           topic listing (with digest counts); add/remove
                            registry subcommands (atomic jq edits)
scripts/fetch.py            mechanical half of /read: fetch, filter, state
scripts/verify-feeds.py     feed health checks for /add (rss/html/fail)
scripts/browse.sh           builds the browse page; --build = rebuild only
scripts/browse-server.py    localhost server: page, /read trigger, /status
.claude/skills/read/        digest skill (fetch, filter, format, stats, state)
.claude/skills/topics/      topic-list skill (wraps scripts/topics.sh)
.claude/skills/browse/      browse-page skill (wraps scripts/browse.sh)
.rifm/                      gitignored, machine-local:
  <topic>.json              per-topic read state
  <topic>.<YYYYMMDD>.md     one stored digest per run that had items
  browse.html               the generated page
  last-read.log             raw event stream of the last headless read
```

Browse needs `jq`, `python3`, and `curl`; without python3 it falls back to
opening the static file with "Read now" disabled.

## Topics and sources

Each topic in `topics.json` hardcodes its own provider list:

```json
"ai-models": {
  "title": "AI models",
  "focus":   ["...what counts as relevant..."],
  "exclude": ["...what to drop..."],
  "sources": [{ "url": "...", "type": "rss", "note": "..." }]
}
```

- `type: "rss"` — feed URLs (preferred): dated entries, incremental reads,
  gists from the feed's own description
- `type: "html"` — page fallback: the page is fetched and summarized;
  newness is tracked by a seen-URL list