subscope v0.2.0

Second release. Everything since v0.1.0.

Live subreddit discovery (onboarding)

Onboarding no longer seeds target subreddits from a generic founder template. It now searches Reddit live against your own pain phrasing + competitor brands, recommends only subreddits with a real buyer thread in the last 7 days (clickable link + absolute timestamp), explains why each was picked in plain English, and has Claude review every candidate to drop false positives a keyword search cannot catch (career questions, self-promo, brand-name collisions). Each recommendation carries a 0-100 confidence score. See "How subscope finds your subreddits" in the README.

DataForSEO + Firecrawl enrichment (optional)

Conditional consumers: when keys are present, DataForSEO seeds your competitor list and Firecrawl scrapes your homepage for positioning. Silent no-op without keys. Results cached in SQLite; daily scans are cache-read only.

Reddit OAuth removed

Public JSON only. The account-creation captcha was blocking too many users from the value path; the reply-tracking feature that needed OAuth went with it. Zero keys required on day 1.

Skills renamed

All 15 skills now carry the subscope- prefix (/subscope-run, /subscope-onboard, ...) so they are unambiguous in the slash-command menu.

Scan quality (live-feedback fixes)

Clearer /subscope-tune prompts, a grouped plain-English "what was filtered" footer on /subscope-run, a 24h freshness floor, author-quality thresholds externalized to weights.yml, and backfill so a run reliably returns 5 to 12 surfaces.

274 tests. MIT.

subscope v0.1.0 — first public release

A Claude Code plugin that classifies Reddit posts into 8 buyer-intent patterns and surfaces up to 12 per day, ranked. Runs locally inside your Claude Code session.

Highlights

• 16 invocable skills for daily Reddit pain-pattern surfacing
• 8 intent classes with distinct scoring paths (pricing-rage, churn, build-vs-buy, rfp-bait, stack-audit, alternative-seeking, resurrect, rivals)
• 60-second `/subscope:onboard` or 12-minute deep `/subscope:profile` for targeted config building (ICP, competitor anchor, pain language, subreddit tiers, few-shot examples)
• Provider-agnostic LLM support (Anthropic, OpenAI, Groq, OpenRouter, Ollama)
• Postmortem learning loop, auto-detects your sent replies, scores 7-day outcomes, feeds back into the ranker
• Multi-surface output: inline Claude Code chat (default), Notion 14-column triage DB, Slack webhook, Obsidian weekly digest
• Permanent SQLite dedup so a post never reappears
• Author vet pre-gate filters throwaway and karma-farmer OPs before scoring
• Cooling queue holds new surfaces 15 minutes before promotion to avoid bot-pattern fingerprints
• SSRF-guarded LLM endpoint (blocks RFC-1918, AWS metadata, http:// to non-localhost)
• Atomic 0o600 file writes for all credentials

Install

```bash
/plugin install dancolta/subscope
/subscope:setup
```

Runs on day 1 with zero API keys. Each layer (OAuth, LLM, Notion, Slack, Obsidian) is opt-in via the setup wizard.

Stats

• 374-LOC Reddit fetcher (merged OAuth + public JSON paths)
• 122 unit tests, all passing
• 0 telemetry, 100% local data
• Plugin manifest validates clean against Anthropic's current schema

Known limitations

• Manual invocation only (no cron / scheduler). Wrap `subscope fetch-score` in cron or launchd if you want recurring delivery.
• Reddit's public-JSON endpoint rate-limits without OAuth. Configure OAuth via `/subscope:setup` for 10x rate budget.

Next

• v0.1.1: `cli.py` unit tests, expanded preset library, Anthropic prompt caching
• v0.2.0: HackerNews adapter (`/subscope:hn`)
• v0.3.0: LinkedIn pulse adapter (public posts only)

See the README for full docs.