Feed Agent

Feed Agent

An autonomous intelligence aggregator that discovers, filters, and analyzes topic-specific content with self-evolution capabilities.

Commands

Workflow: /feed-agent [topic]

Step 1: Load Configuration

python3 scripts/pipeline.py --project-root . --action init --topic "{topic}"

Reads config/feed-agent.yaml for:

• Active topic
• Enabled search providers
• Scoring thresholds
• Evolution settings

Step 2: Scout (Discovery)

Run source discovery using enabled search providers:

python3 scripts/scout.py --project-root . --topic "{topic}"

Uses the SearchProvider interface to:

1. Search for "{topic} RSS feeds" and "{topic} news sources"
2. Validate candidate URLs (must return valid RSS/HTML)
3. Score candidates by relevance (LLM-based)
4. Store candidates in sources/feeds/{topic}/candidates.json

Step 3: Fetch (Content Collection)

Fetch full content from all active sources:

python3 scripts/fetcher.py --project-root . --topic "{topic}"

For each source:

1. Fetch RSS feed or scrape web page
2. Extract full article content using webfetch
3. Normalize content (strip HTML, extract text)
4. Store raw items in database

Step 4: Analyze (AI Scoring)

Score and filter articles:

python3 scripts/analyzer.py --project-root . --topic "{topic}"

For each article:

1. Calculate relevance score (0-10) against topic
2. Compare against historical reports for deduplication
3. Extract core points and insights
4. Mark as "new insight" or "continuation"
5. Store analysis in database

Articles scoring < 7 are filtered out.

Step 5: Generate Report

Create markdown intelligence report:

python3 scripts/reporter.py --project-root . --topic "{topic}" --output reports/{topic}/{date}.md --as-of-date {date} --window-days 7

Output structure:

• Core Insights (Top insights from promoted articles)
• Detailed Analysis (Full analysis of each article)
• Self-Evolution (Source quality updates and keyword adjustments)

Report generation rules:

• Reports are date-anchored to the requested --as-of-date.
• The default coverage window is the previous 7 days, inclusive.
• Article freshness is determined from published_at when available, with fetched_at as fallback.
• Stories already shown in earlier reports are hidden by default unless analysis marks the new coverage as a meaningful continuation.

Step 6: Self-Evolve

Update source quality and keywords:

python3 scripts/evolution.py --project-root . --topic "{topic}"

Based on average scores:

• High-quality sources (> 0.7 avg): Promote in priority
• Low-quality sources (< 0.3 avg): Flag for pruning after N consecutive runs
• Extract suggested keywords from high-scoring articles

Workflow: /feed-agent set-topic [topic]

Initialize or update topic configuration:

python3 scripts/pipeline.py --project-root . --action config --topic "{topic}"

Creates/updates:

• config/feed-agent.yaml - Topic and provider settings
• sources/feeds/{topic}/active.yaml - Active source list
• Database entry for the topic

Workflow: /feed-agent scout

Run discovery phase only:

python3 scripts/scout.py --project-root . --topic "{topic}" --verbose

Outputs discovered candidates to console and saves to candidates.json.

Workflow: /feed-agent report [--as-of-date YYYY-MM-DD]

Generate a fresh daily report for the target date:

python3 scripts/pipeline.py --project-root . --action report --topic "{topic}" --as-of-date {date} --window-days 7

This workflow refreshes sources, analyzes newly fetched items, and produces a report covering the inclusive 7-day window ending on {date}.

Workflow: /feed-agent sources

List source status:

python3 scripts/db.py --project-root . --topic "{topic}" list-sources

Shows:

• Active sources with quality scores
• Candidate sources pending validation
• Pruned sources (if any)

Workflow: /feed-agent evolve

Force evolution check:

python3 scripts/evolution.py --project-root . --topic "{topic}" --force

Search Providers

Feed Agent uses a pluggable search provider system. See references/search-providers.md for implementation guide.

Available Providers

Provider Configuration

providers:
  enabled:
    - exa
    - browser
  
  exa:
    max_results: 10
    use_autosuggestions: true
  
  browser:
    headless: true
    timeout: 30

AI Scoring

Each article is scored 0-10 for relevance to the configured topic. See references/scoring-prompts.md for the complete scoring prompt template.

Scoring Guide

Historical Comparison

The analyzer checks if an article's core points have been covered in previous reports:

• New insight: Not covered in last 30 days
• Continuation: Updates or extends previous coverage
• Duplicate: Same content from different source (filtered)

Self-Evolution

The agent continuously improves its source list:

Source Quality Tracking

-- Every run updates source quality scores
UPDATE feed_agent_sources 
SET quality_score = promoted_articles / total_articles,
    relevance_avg = AVG(recent_scores)
WHERE topic = ?

Auto-Promotion

Candidates with > 70% promoted articles after 3 validation runs are automatically added to the active source list.

Auto-Pruning

Sources with < 30% relevance average over 3 consecutive runs are flagged for removal. User must confirm pruning.

Keyword Evolution

High-scoring articles are analyzed for frequently occurring terms not in the current keyword list. These are stored in the database as candidate keywords, promoted to active when they repeatedly appear in strong articles, and then reused to drive feed discovery queries and relevance scoring.

Keyword States

• active - Used by scout query generation and analysis prompts
• candidate - Suggested by analysis, waiting for more evidence
• rejected - Suppressed from future reuse

Report Format

Generated reports follow the template in references/report-template.md:

# {Topic} Intelligence Report
**Date:** {date}  
**Sources Scanned:** {count} | **Articles Found:** {count} | **Promoted (>=7):** {count}

## Core Insights

> Top insights from all promoted articles

| # | Insight | Source | New? |
|---|---------|--------|------|
| 1 | {insight} | {source} | Yes |

## Detailed Analysis

### {Article Title}
**Score:** {score}/10 | **Source:** {source_name} | **New Insight:** {is_new}

**Core Points:**
- {point_1}
- {point_2}

**Link:** {url}

---

## Self-Evolution

### Source Quality Update
| Source | Avg Score | Status | Action |
|--------|-----------|--------|--------|
| {name} | 0.82 | Active | Maintained |

### Keyword Adjustments
- **Added:** {keywords}
- **Reasoning:** {reason}

Database Schema

See scripts/db.py for schema definitions. Key tables:

• feed_agent_sources - Source registry with quality tracking
• feed_agent_articles - Article storage with analysis
• feed_agent_evolution_log - Track all evolution actions

Dependencies

pip install feedparser beautifulsoup4 aiohttp pyyaml

Example Usage

User: /feed-agent "AI Agents"

Action:
1. Check if topic "AI Agents" is configured
2. Run scout to discover new sources
3. Fetch content from all active sources
4. Analyze and score articles (0-10)
5. Filter (keep ≥ 7)
6. Generate report at reports/ai-agents/2026-03-23.md
7. Run evolution to update source quality
8. Display summary to user

Resource Requirements

When invoking this skill, AI agents should be aware of the following runtime characteristics:

• Typical time: 3-10 minutes depending on source count and network conditions
• Network: Heavy (multiple RSS fetches, web searches, and content extraction)
• Recommended timeout: timeoutSeconds: 600 (10 minutes) minimum
• Token usage: ~500k-1M tokens per full pipeline run

This is a long-running batch task. Do not expect synchronous results; use appropriate async patterns (e.g., sessions_yield, background sessions) when calling.

Error Handling

Changelog

2026-04-21 - Bug Fixes

Fixed:

• Database schema migration - _ensure_article_columns() now properly adds missing columns (published_at, story_key, story_status) on database init
• Analyzer article retrieval - Added include_unanalyzed parameter to get_articles() to properly fetch articles pending analysis (with NULL or 0 relevance_score)
• Date parsing fallback - parse_date() in fetcher now defaults to current time if no date found in feed entry, ensuring published_at is always set
• Article insertion - insert_article() now ensures published_at is always set with fallback to current time
• Extended analysis window - Changed from 1 day to 7 days for finding unanalyzed articles to capture more content

Impact: These fixes resolve the "no such column" errors and ensure articles are properly fetched, analyzed, and included in reports even when RSS feeds lack proper date metadata.