clawfeedradar（clawfeedradar 搜索）

clawfeedradar (OpenClaw Skill)

clawfeedradar is a personal "news radar" skill built on top of an existing clawsqlite knowledge base and its interest clusters.

It answers one question:

"Given what I have been reading / saving into clawsqlite, what new articles on Hacker News / RSS / arXiv etc. am I likely to care about today?"

The heavy lifting (interest clusters, embeddings, hybrid search) lives in clawsqlite. This skill focuses on:

1. Fetching candidates from external sources (HN/RSS/etc.)
2. Scoring them against your long‑term interest clusters in clawsqlite
3. Generating per‑source RSS feeds (XML + JSON sidecar) with optional

small‑LLM summaries and bilingual bodies

1. Relationship to clawsqlite & clawsqlite-knowledge

• clawsqlite

- Maintains your knowledge base and interest clusters - Provides the interest_clusters + interest_cluster_members tables - Writes an interest_meta key interest_clusters_last_built_at after each cluster build

• clawsqlite-knowledge (skill)

- High‑level ingest/search/show skill for OpenClaw agents - Runs on the same DB / articles directory as clawsqlite

• clawfeedradar (this skill)

- Reads interest_clusters & interest_meta from the same DB - Never mutates the clawsqlite DB schema or articles - Uses the embedding config (EMBEDDING_* + CLAWSQLITE_VEC_DIM) shared with clawsqlite to embed candidate articles into the same interest space - Writes XML/JSON feeds to a separate CLAWFEEDRADAR_OUTPUT_DIR - Optionally publishes generated feeds to a git repository for GitHub/Gitee Pages

In short: clawsqlite knows what you like, clawfeedradar goes out and finds similar things, then packages them as RSS.

2. Installation & bootstrap

This skill expects a Python runtime with access to PyPI (for the clawfeedradar package) and a workspace where the skill lives under skills/clawfeedradar.

2.1 Install the skill shell

openclaw skills install clawfeedradar

This will create:

~/.openclaw/workspace/skills/clawfeedradar
  ├── SKILL.md
  ├── manifest.yaml
  ├── bootstrap_deps.sh
  ├── run_clawfeedradar.py
  ├── README.md
  ├── README_zh.md
  └── ENV_EXAMPLE.md

Note: this skill does not read an .env inside the skill directory. All configuration lives in the upstream clawfeedradar project (or at the agent/host level), and the skill simply reuses that environment.

The actual Python package clawfeedradar is installed by the bootstrap script, not vendored here.

2.2 Install / upgrade the clawfeedradar package

The manifest.yaml declares a bash bootstrap step:

install:
  - id: clawfeedradar_bootstrap
    kind: bash
    label: Install clawfeedradar Python package
    script: bootstrap_deps.sh

bootstrap_deps.sh is a tiny wrapper around python -m pip install with an optional workspace prefix fallback, similar to the pattern used by the clawsqlite-knowledge skill.

In simplified form, it:

• Tries to install clawfeedradar into the default Python environment

used by the skill;

• If that fails (read‑only env, no pip), tries again with

--prefix="$WORKSPACE/skills/clawfeedradar/.venv";

• Prints a NEXT: hint on success so you know where the package was

installed and which site-packages will be added to PYTHONPATH.

You normally do not need to run this script manually; openclaw skills install clawfeedradar (and a future openclaw skills update clawfeedradar) will run it for you.

3. Runtime contract

The skill runtime calls run_clawfeedradar.py with a JSON payload on stdin and expects a JSON response on stdout.

Common fields:

• root (optional): path to the clawsqlite knowledge root (if omitted,

CLAWSQLITE_ROOT / .env defaults are used)

• action: one of the supported actions below

All handlers return a JSON object with at least:

• ok: true|false
• data: ... on success, or error / exit_code / stdout /

stderr on failure

• next: [...] when the underlying CLI emits NEXT hints
• error_kind for coarse classification (e.g. missing embedding /

scraper / vec ext)

Internally, the runtime:

• Ensures the workspace‑local site‑packages prefix is on PYTHONPATH;
• Executes python -m clawfeedradar.cli ... with the right arguments;
• Captures stdout/stderr and parses JSON when possible.

4. Supported actions

4.1 run_once

Run the radar once for a single source URL (HN/RSS/etc.), score candidates against interest clusters, and write XML/JSON feeds.

Payload example:

{
  "action": "run_once",
  "source_url": "https://feeds.bbci.co.uk/news/technology/rss.xml",
  "max_source_items": 30,
  "max_items": 12,
  "score_threshold": 0.4,
  "source_lang": "en",
  "target_lang": "zh",
  "enable_preview": true,
  "preview_words": 512,
  "root": "/home/node/.openclaw/workspace/knowledge_data"
}

Behavior (high level):

• Resolves the source type (HN / RSS / etc.) from source_url;
• Fetches up to max_source_items raw items from that source;
• Normalizes URLs and deduplicates using a 7‑day seen_urls statefile;
• Embeds each candidate into the same interest space as clawsqlite using

the shared EMBEDDING_* + CLAWSQLITE_VEC_DIM config;

• Loads interest_clusters and warns if

interest_clusters_last_built_at is older than 7 days;

• Computes an interest_score_raw for each candidate based on:

  interest_raw = Σ_k (cluster_weight_k * max(0, sim_k))
  cluster_weight_k = size_k / Σ_j size_j

• Maps interest_raw into [0,1] via a sigmoid:

  interest_score = sigmoid(k*(interest_raw - 0.5))
  k = CLAWFEEDRADAR_INTEREST_SIGMOID_K (default 4.0)

• Adds a small bias from recency and popularity:

  interest_bias = w_recency * rec + w_popularity * pop

where rec/pop are in [0,1] and weights are controlled by CLAWFEEDRADAR_W_RECENCY / CLAWFEEDRADAR_W_POPULARITY.

• Computes final_score ≈ interest_score + interest_bias + source_extras;
• Filters by score_threshold and picks up to max_items top

candidates;

• Fetches fulltext for the selected items (per‑host serial, cross‑host

parallel);

• Optionally calls a small LLM to generate:

- summary_preview (preview summary, controlled by preview_words / CLAWFEEDRADAR_LLM_CONTEXT_TOKENS) - body_bilingual (bilingual body paragraphs, controlled by CLAWFEEDRADAR_LLM_MAX_PARAGRAPH_CHARS and lang env);

• Writes:

- <OUTPUT_DIR>/<slug>.xml — RSS feed for your reader; - <OUTPUT_DIR>/<slug>.json — JSON sidecar with full debug info (scores, cluster matches, fulltext, bilingual body, etc.).

If publishing to git is configured (see ENV_EXAMPLE.md), the runtime will also update a remote git repo (GitHub/Gitee) so you can subscribe via Pages.

Return value: a JSON object with a summary of the run (number of candidates, selected items, output paths, warnings if any).

4.2 schedule_from_sources_json

Run the radar for multiple sources defined in a sources.json file.

This is a higher‑level wrapper around run_once that:

• Reads a sources JSON (same shape as the clawfeedradar CLI);
• Iterates over enabled sources, applying per‑source config

(max_items/threshold/lang/etc.);

• Aggregates results and returns a summary per source.

Payload example:

{
  "action": "schedule_from_sources_json",
  "sources_file": "/home/node/.openclaw/workspace/clawfeedradar/sources.json",
  "root": "/home/node/.openclaw/workspace/knowledge_data"
}

The exact JSON shape for sources.json is documented in the clawfeedradar README/SPEC; this skill does not invent a new format.

5. ENV configuration (summary)

This skill itself does not read a .env file; it relies on the underlying clawfeedradar and clawsqlite CLIs, which support project‑level .env files.

In OpenClaw deployments you typically configure env vars at the agent (or host) level so both skills and direct CLI usage share the same configuration.

Key envs (see ENV_EXAMPLE.md for a consolidated list):

• Knowledge base

- CLAWSQLITE_ROOT / CLAWSQLITE_DB

• Embedding service

- EMBEDDING_BASE_URL / EMBEDDING_MODEL / EMBEDDING_API_KEY - CLAWSQLITE_VEC_DIM

• Interest clusters (configured on clawsqlite side)

- CLAWSQLITE_INTEREST_* (algo, PCA, min_size, max_clusters, etc.)

• Fulltext fetch

- CLAWFEEDRADAR_SCRAPE_CMD - CLAWFEEDRADAR_SCRAPE_WORKERS

• LLM summaries / bilingual body

- SMALL_LLM_* - CLAWFEEDRADAR_LLM_* (context tokens, paragraph size, sleep, source/target lang)

• Scoring

- CLAWFEEDRADAR_INTEREST_SIGMOID_K - CLAWFEEDRADAR_W_RECENCY / CLAWFEEDRADAR_W_POPULARITY - CLAWFEEDRADAR_RECENCY_HALF_LIFE_DAYS

• Output & publish

- CLAWFEEDRADAR_OUTPUT_DIR - CLAWFEEDRADAR_PUBLISH_GIT_REPO / CLAWFEEDRADAR_PUBLISH_GIT_BRANCH / CLAWFEEDRADAR_PUBLISH_GIT_PATH

6. When to use this skill

Use clawfeedradar (skill) when:

• You already have a clawsqlite knowledge base with interest clusters;
• You want an agent‑friendly way to:

- run daily/weekly news radar jobs; - generate and publish personalized RSS feeds; - inspect scores / clusters per article via JSON.

Use clawfeedradar (CLI) when:

• You are developing or debugging the radar itself;
• You want full control over CLI flags / env without going through the

skill JSON API.

Both run against the same clawsqlite DB and can coexist.

7. Security & limits

• This skill does not modify the clawsqlite DB schema or articles;
• It only reads embeddings and interest cluster metadata;
• All external network requests (feeds, fulltext, LLM, git) are made via

the underlying clawfeedradar CLI and are fully auditable.

If you need stricter control (e.g. only allow certain sources, or completely disable publishing), you can wrap this skill with your own policy layer or restrict env/config at the agent level.

For full implementation details (candidate model, scoring formula, LLM pipelines), see the upstream clawfeedradar docs.

In particular:

• docs/SPEC_en.md / docs/SPEC_zh.md: formal spec for the pipeline
• README.md / README_zh.md: human‑oriented overview and usage

instructions