gh-sentiment

name

gh-sentiment

description

Fetches GitHub issues and performs hybrid sentiment analysis using VADER (offline scoring) and the Cursor agent as the LLM (no API key required). Produces a sentiment-enriched JSON or markdown report per issue: tone label, tone detail, and a one-sentence summary. Use when the user asks to analyze sentiment of GitHub issues, score issue tone, classify issue mood, summarize issues with sentiment, or says things like "analyze sentiment", "score these issues", "what's the tone of this issue", or "run sentiment analysis without an API key".

GitHub Issue Sentiment Analysis (Agent-Guided)

Performs hybrid sentiment analysis on GitHub issues using VADER for a fast numeric score and the Cursor agent for nuanced tone classification — no external LLM API key needed.

Quick Start

# 1. Fetch issues as JSON
gh issue list --repo <owner/repo> --limit <N> --state open \
  --json number,title,state,createdAt,author,labels,url,body \
  > /tmp/issues.json

# 2. Score each issue with VADER
python .cursor/skills/gh-sentiment/scripts/vader_score.py /tmp/issues.json \
  > /tmp/issues_scored.json

# 3. Agent performs nuanced analysis (see Agent Workflow below)

Agent Workflow

When the user asks for sentiment analysis on GitHub issues, follow these steps:

Step 1 — Resolve the repo

Parse from whatever the user provides:

• Full URL: https://github.com/owner/repo → owner/repo
• Shorthand: owner/repo → use as-is
• If ambiguous, ask

Step 2 — Fetch issues

gh issue list --repo <owner/repo> --limit <N> --state open \
  --json number,title,state,createdAt,author,labels,url,body

If --limit not specified, default to 20.

Step 3 — Run VADER scoring

python .cursor/skills/gh-sentiment/scripts/vader_score.py - <<'JSON'
<paste issues JSON here>
JSON

Or pipe from a file:

python .cursor/skills/gh-sentiment/scripts/vader_score.py /tmp/issues.json

The script outputs the same issues array with a vader_score field added to each.

Step 4 — Agent performs nuanced analysis

For each issue in the VADER-scored output, reason about the issue text and produce:

Field	Description
tone_label	One of: neutral frustrated enthusiastic disappointed urgent constructive
tone_detail	1-3 word freeform sub-label (e.g. mildly critical, feature proposal)
summary	Single sentence — what the issue asks for or reports

Reasoning guidance:

• Use vader_score as a grounding signal: < -0.3 skews negative, > 0.3 skews positive
• Read the full title and body, not just the score
• Labels like bug suggest frustration/urgency; enhancement suggests constructive/enthusiastic
• Match the fixed tone_label vocabulary; use tone_detail for nuance
• Keep summary to one sentence; focus on the core ask or problem

Step 5 — Output

Merge the agent's analysis into each issue and produce the final output.

JSON shape per issue:

{
  "number": 3,
  "title": "...",
  "labels": [{"name": "enhancement"}],
  "sentiment": {
    "vader_score": 0.1027,
    "tone_label": "constructive",
    "tone_detail": "feature proposal",
    "summary": "Proposes adding LLM-based sentiment tagging to issues as a derived triage label.",
    "llm_provider": "cursor-agent"
  }
}

Markdown output: Render a summary table followed by per-issue sections matching the format used by the gh-issues skill, with a Sentiment block appended to each issue.

Step 6 — Optional export

If the user wants to save the output:

# Agent writes final JSON to disk

Ask: "Would you like me to export this to a file?" — if yes, write to the provided path.

---

Tone vocabulary reference

Label	When to use
neutral	Factual report, no strong emotion, routine request
frustrated	Expresses pain, blocker, "broken", anger
enthusiastic	Excitement about a feature, very positive framing
disappointed	Expectation not met, mild negative without anger
urgent	Time pressure, production impact, "ASAP", outage
constructive	Thoughtful improvement proposal, reasoned feature request

---

Requirements

• gh CLI installed and authenticated
• Python 3.10+ with vaderSentiment installed (pip install vaderSentiment)
• No LLM API key needed — the Cursor agent provides the nuanced analysis