hada-scout

3-phase in-session pipeline that monitors hada.io (via feedburner RSS) for relevant articles, uses a haiku LLM batch to pre-score relevance, and dispatches full /scout analysis only on high-scoring candidates.

Purpose

Replace the v1.0 keyword-regex approach with context-aware LLM pre-scoring. Haiku evaluates all feed titles in a single batch call, reducing false positives from ~30-40% to ~5-10% and eliminating the external CronJob dependency for user-invoked runs.

Architecture: 3-Phase In-Session Pipeline

Phase 1 — Fetch & Parse

1. WebFetch the hada.io RSS feed (https://feeds.feedburner.com/geeknews-feed)
2. Parse all items: title, URL, publication date
3. Default: latest 50 items (configurable via --limit or HADA_SCOUT_LIMIT)

Phase 2 — Pre-Scout (haiku batch)

1. Spawn 1 haiku agent with ALL item titles as a single batch input
2. Agent evaluates each title against oh-my-customcode's domain (see prompt template below)
3. Returns relevance score (0–100) and a 1-line reason for each item
4. Threshold: ≥ 60 passes to Phase 3 (configurable via --threshold or HADA_SCOUT_THRESHOLD)
5. Cost: ~$0.01–0.05 for 50 items

Phase 3 — Scout Dispatch

1. Dedup: check existing hada-scout labeled issues via gh issue list --label hada-scout
2. For each passing item (max 5 per run, configurable via MAX_SCOUT_PER_RUN):
   • Run full /scout analysis via Skill tool invocation
   • Scout creates GitHub issue with verdict label on baekenough/oh-my-customcode
   • Add hada-scout label to the created issue
3. Dispatch up to 4 scouts in parallel per R009

Pre-Scout Prompt Template

The haiku agent receives the following system prompt:

You are a relevance filter for the oh-my-customcode project — an AI agent harness/orchestration
system built on Claude Code CLI with 50 agents, 115 skills.

Project domains (HIGH relevance):
- AI agent orchestration, multi-agent systems, agent design patterns
- Harness, benchmark, evaluation frameworks for AI agents
- Claude Code, Anthropic ecosystem, MCP protocol
- Code review automation, development workflow automation
- Agent sandbox, isolation, security patterns
- LLM-assisted development tools and methodologies

Project domains (MEDIUM relevance):
- General AI/ML tooling that could be adapted for agent workflows
- DevOps automation patterns applicable to agent infrastructure
- New programming paradigms for AI-assisted development

NOT relevant:
- Pure frontend/UI frameworks without agent connection
- Business/management topics
- Hardware, networking, non-AI infrastructure
- Social media, marketing tools

For each item below, return: score (0-100) | reason (1 line)

Items:
{numbered_item_list}

Display Format

[hada-scout] Scanning hada.io feed...
├── Phase 1: Fetched {n} items
├── Phase 2: Pre-scout → {passed}/{total} items passed (threshold: {t}%)
│   ├── ✓ {title1} (score: {s1}%)
│   ├── ✓ {title2} (score: {s2}%)
│   └── ✗ {title3} (score: {s3}%) — skipped
├── Phase 3: Scout dispatch ({n} items, max 5)
│   ├── [1] /scout {url1} → {verdict}
│   └── [2] /scout {url2} → {verdict}
└── [Done] {created}/{dispatched} issues created

Label Scheme

Label	Purpose
hada-scout	Source identification — all hada-scout created issues
scout:internalize	/scout verdict: adopt into project
scout:integrate	/scout verdict: use as external dependency
scout:skip	/scout verdict: not relevant

Cost Controls

Stage	Model	Estimated Cost
Pre-scout (Phase 2)	haiku	~$0.01–0.05 per run (50 items)
Full scout (Phase 3)	sonnet	~$0.5–1.5 per item, max 5 per run
Total max per invocation	—	~$8

Environment Variables

Variable	Default	Description
FEED_URL	https://feeds.feedburner.com/geeknews-feed	RSS feed URL
HADA_SCOUT_THRESHOLD	60	Pre-scout score threshold (0–100)
HADA_SCOUT_LIMIT	50	Max feed items to fetch and score
MAX_SCOUT_PER_RUN	5	Max /scout executions per invocation
GH_TOKEN	(required)	GitHub PAT for issue creation and dedup

Integration

Rule	How
R009	Phase 3 scout dispatches run in parallel (up to 4 concurrent)
R010	Orchestrator manages phases; analysis delegated to haiku/sonnet agents
R015	Pre-scout scores and reasons displayed before dispatching full scouts
scout skill	Phase 3 invokes /scout via Skill tool for each candidate URL

Differences from v1.0

Aspect	v1.0 (keyword)	v2.0 (LLM pre-scout)
Filtering	Regex keyword match	LLM relevance scoring (haiku)
Invocation	External CronJob only	User-invocable /hada-scout + CronJob
Precision	Low (keyword false positives)	High (context-aware scoring)
Cost per scan	$0 (regex) + $2.5–7.5 (/scout)	$0.05 (pre-scout) + $2.5–7.5 (/scout)
False positive rate	~30–40%	~5–10%
Scope	package	core

Tracking

GitHub Issue #841