Merge pull request #5 from zander-raycraft/gh/issue-pr-filter

fixed prop cutoff issue for pr/issue filtering

.github/scripts/triage/embedding_utils.py

@@ -0,0 +1,178 @@
+"""Pure math utilities for triage sweep embedding analysis.
+
+All functions are stateless and perform no I/O (except model loading by FastEmbed).
+Each function operates on numpy arrays and returns numpy arrays or plain Python types.
+"""
+from __future__ import annotations
+
+import numpy as np
+from numpy.typing import NDArray
+from fastembed import TextEmbedding
+from sklearn.decomposition import PCA
+from sklearn.covariance import EllipticEnvelope
+from sklearn.metrics.pairwise import cosine_similarity
+
+# FastEmbed model — BAAI/bge-small-en-v1.5 produces 384-dimensional embeddings.
+# ~46MB quantized ONNX, runs on CPU in ~0.5s per batch of 32.
+EMBEDDING_MODEL: str = "BAAI/bge-small-en-v1.5"
+
+# Embedding dimensionality (determined by model choice).
+EMBEDDING_DIM: int = 384
+
+# Batch size for FastEmbed. 32 balances memory and throughput on
+# a 2-vCPU GitHub Actions runner with ~7GB RAM.
+EMBEDDING_BATCH_SIZE: int = 32
+
+
+def embed_texts(texts: list[str]) -> NDArray[np.float32]:
+    """Embed a list of texts into dense vectors using FastEmbed.
+
+    Returns an array of shape (len(texts), 384) with dtype float32.
+    Empty input returns a (0, 384) array.
+    """
+    if not texts:
+        return np.empty((0, EMBEDDING_DIM), dtype=np.float32)
+
+    model = TextEmbedding(model_name=EMBEDDING_MODEL)
+    vectors = list(model.embed(texts, batch_size=EMBEDDING_BATCH_SIZE))
+    return np.vstack(vectors).astype(np.float32)
+
+
+def normalize_rows(matrix: NDArray[np.float32]) -> NDArray[np.float32]:
+    """L2-normalize each row to unit length.
+
+    Zero-norm rows (e.g. from empty text) remain zero vectors.
+    Uses eps=1e-10 in the denominator to avoid division by zero.
+    """
+    if matrix.shape[0] == 0:
+        return matrix
+
+    norms = np.linalg.norm(matrix, axis=1, keepdims=True)
+    return matrix / (norms + 1e-10)
+
+
+def reduce_dimensions(
+    matrix: NDArray[np.float32],
+    variance_ratio: float,
+    max_components: int,
+) -> NDArray[np.float32]:
+    """Reduce dimensionality via PCA.
+
+    Computes n_components = min(max_components, n-1, d). If n_components < 1,
+    returns the matrix unchanged. Logs explained variance for observability.
+    The variance_ratio parameter documents intent but is not strictly enforced;
+    the actual retained variance depends on the data and component cap.
+    """
+    n, d = matrix.shape
+    if n <= 1:
+        return matrix
+
+    n_components = min(max_components, n - 1, d)
+    if n_components < 1:
+        return matrix
+
+    pca = PCA(n_components=n_components)
+    reduced = pca.fit_transform(matrix)
+    explained = pca.explained_variance_ratio_.sum()
+    print(f"PCA: {d}d -> {n_components}d, explained variance: {explained:.3f}")
+    return reduced.astype(np.float32)
+
+
+def detect_outliers(
+    matrix: NDArray[np.float32],
+    threshold: float,
+) -> list[int]:
+    """Flag items whose Mahalanobis distance exceeds the threshold.
+
+    Uses EllipticEnvelope (robust covariance via MCD) to estimate the
+    multivariate Gaussian, then computes sqrt(squared Mahalanobis distance)
+    for each sample. Returns indices of outliers sorted ascending.
+    """
+    n = matrix.shape[0]
+    if n < 2:
+        return []
+
+    envelope = EllipticEnvelope(contamination=0.1, random_state=42)
+    envelope.fit(matrix)
+
+    # .mahalanobis() returns squared Mahalanobis distances
+    distances = np.sqrt(envelope.mahalanobis(matrix))
+    outlier_mask = distances > threshold
+    return list(np.where(outlier_mask)[0])
+
+
+def find_duplicate_pairs(
+    matrix: NDArray[np.float32],
+    threshold: float,
+) -> list[tuple[int, int, float]]:
+    """Find pairs of items with cosine similarity above threshold.
+
+    Returns (i, j, similarity) tuples where i < j. The input should be
+    L2-normalized embeddings (full dimensionality, not PCA-reduced) so
+    cosine similarity equals the dot product.
+    """
+    n = matrix.shape[0]
+    if n <= 1:
+        return []
+
+    sim_matrix = cosine_similarity(matrix)
+    # Upper triangle indices (i < j), excluding diagonal
+    rows, cols = np.triu_indices(n, k=1)
+    similarities = sim_matrix[rows, cols]
+
+    mask = similarities > threshold
+    pairs: list[tuple[int, int, float]] = []
+    for idx in np.where(mask)[0]:
+        pairs.append((int(rows[idx]), int(cols[idx]), float(similarities[idx])))
+
+    return pairs
+
+
+# ── Label suggestion via embedding similarity ────────────────────────
+
+# Minimum similarity between an item and a label to suggest it.
+# 0.4 is intentionally permissive — the report is for human review.
+LABEL_SIMILARITY_THRESHOLD: float = 0.4
+
+# Maximum number of labels to suggest per item.
+MAX_LABELS_PER_ITEM: int = 3
+
+
+def suggest_labels(
+    item_embeddings: NDArray[np.float32],
+    label_embeddings: NDArray[np.float32],
+    label_names: list[str],
+    threshold: float = LABEL_SIMILARITY_THRESHOLD,
+    max_per_item: int = MAX_LABELS_PER_ITEM,
+) -> list[list[tuple[str, float]]]:
+    """Suggest labels for each item based on embedding similarity.
+
+    Computes cosine similarity between item embeddings (n, 384) and
+    label embeddings (m, 384). For each item, returns the top-k labels
+    whose similarity exceeds the threshold, sorted by similarity descending.
+
+    Returns a list of length n, where each element is a list of
+    (label_name, similarity) tuples. Empty list if no label exceeds threshold.
+    """
+    n = item_embeddings.shape[0]
+    m = label_embeddings.shape[0]
+    if n == 0 or m == 0:
+        return [[] for _ in range(n)]
+
+    # (n, m) similarity matrix: each row is one item vs all labels
+    sim_matrix = cosine_similarity(item_embeddings, label_embeddings)
+
+    suggestions: list[list[tuple[str, float]]] = []
+    for i in range(n):
+        row = sim_matrix[i]
+        # Indices sorted by similarity descending
+        ranked = np.argsort(row)[::-1]
+        item_labels: list[tuple[str, float]] = []
+        for idx in ranked[:max_per_item]:
+            score = float(row[idx])
+            if score < threshold:
+                break
+            item_labels.append((label_names[idx], score))
+        suggestions.append(item_labels)
+
+    return suggestions

.github/scripts/triage/requirements.txt

@@ -0,0 +1,3 @@
+fastembed>=0.5.0
+numpy>=1.26.0
+scikit-learn>=1.4.0

.github/scripts/triage/sweep.py

@@ -0,0 +1,442 @@
+"""Triage sweep: fetch open issues/PRs, detect outliers and duplicates, generate a report.
+
+Entrypoint script for the triage-sweep workflow. Fetches all open items via
+the GitHub REST API, delegates embedding and analysis to embedding_utils,
+generates a markdown report, and optionally creates a report issue.
+"""
+from __future__ import annotations
+
+import json
+import os
+import sys
+import urllib.request
+import urllib.parse
+from typing import TypedDict
+from datetime import datetime, timezone
+
+from embedding_utils import (
+    embed_texts,
+    normalize_rows,
+    reduce_dimensions,
+    detect_outliers,
+    find_duplicate_pairs,
+    suggest_labels,
+)
+
+# ── Thresholds (overridable via workflow_dispatch inputs) ──────────────
+
+# Mahalanobis distance beyond which an item is flagged as an outlier.
+# Default 3.0 ~ 99.7% of a Gaussian distribution (3-sigma rule).
+MAHALANOBIS_THRESHOLD: float = float(os.environ.get("INPUT_MAHALANOBIS_THRESHOLD", "3.0"))
+
+# Cosine similarity above which two items are flagged as duplicates.
+# 0.92 catches near-identical issues while tolerating paraphrasing.
+COSINE_THRESHOLD: float = float(os.environ.get("INPUT_COSINE_THRESHOLD", "0.92"))
+
+# Hard cap on items to process. Prevents runaway costs on very large repos.
+MAX_ITEMS: int = int(os.environ.get("INPUT_MAX_ITEMS", "500"))
+
+# When true, print report to stdout/file but do not create a GitHub issue.
+DRY_RUN: bool = os.environ.get("INPUT_DRY_RUN", "false").lower() == "true"
+
+# ── Fixed constants (not user-configurable) ───────────────────────────
+
+# Minimum number of samples required for EllipticEnvelope to fit
+# a Gaussian. Below this, outlier detection is skipped because
+# covariance estimation is unreliable.
+MIN_SAMPLES_FOR_OUTLIER_DETECTION: int = 10
+
+# PCA: retain components explaining this fraction of variance.
+# 0.95 keeps 95% of information while reducing dimensionality enough
+# for EllipticEnvelope to be numerically stable.
+PCA_VARIANCE_RATIO: float = 0.95
+
+# PCA: maximum number of components regardless of variance ratio.
+# Caps dimensionality for EllipticEnvelope's n_samples > n_features^2 rule.
+PCA_MAX_COMPONENTS: int = 50
+
+# GitHub REST API page size (max allowed is 100).
+API_PAGE_SIZE: int = 100
+
+# Report issue label.
+REPORT_LABEL: str = "triage-report"
+
+# Report file path (written for the summary step to pick up).
+REPORT_FILE: str = "/tmp/triage-report.md"
+
+
+class TriageItem(TypedDict):
+    """One open issue or PR, with only the fields we need."""
+    number: int
+    title: str
+    html_url: str
+    is_pr: bool
+    labels: list[str]
+    created_at: str
+    # title + body concatenated, used as embedding input
+    text: str
+
+
+def github_api_get(path: str) -> list[dict]:
+    """Make a single authenticated GET request to the GitHub REST API.
+
+    Reads GITHUB_TOKEN and GITHUB_REPOSITORY from env. Raises SystemExit
+    with the HTTP status and response body on any non-2xx response.
+    """
+    token = os.environ["GITHUB_TOKEN"]
+    repo = os.environ["GITHUB_REPOSITORY"]
+    url = f"https://api.github.com/repos/{repo}{path}"
+
+    req = urllib.request.Request(url)
+    req.add_header("Accept", "application/vnd.github+json")
+    req.add_header("Authorization", f"Bearer {token}")
+    req.add_header("X-GitHub-Api-Version", "2022-11-28")
+
+    try:
+        with urllib.request.urlopen(req, timeout=30) as resp:
+            return json.loads(resp.read().decode("utf-8"))
+    except urllib.error.HTTPError as e:
+        body = e.read().decode("utf-8", errors="replace")
+        print(f"::error::GitHub API {e.code}: {body}")
+        sys.exit(1)
+
+
+def fetch_all_open_items() -> list[TriageItem]:
+    """Paginate through all open issues and PRs.
+
+    Returns up to MAX_ITEMS TriageItem dicts. Items with a pull_request
+    key are marked is_pr=True. The text field is title + body concatenated.
+    """
+    items: list[TriageItem] = []
+    page = 1
+
+    while len(items) < MAX_ITEMS:
+        path = (
+            f"/issues?state=open&per_page={API_PAGE_SIZE}"
+            f"&sort=created&direction=desc&page={page}"
+        )
+        data = github_api_get(path)
+
+        if not data:
+            break
+
+        for raw in data:
+            if len(items) >= MAX_ITEMS:
+                break
+
+            body = raw.get("body", "") or ""
+            items.append(TriageItem(
+                number=raw["number"],
+                title=raw["title"],
+                html_url=raw["html_url"],
+                is_pr="pull_request" in raw,
+                labels=[lbl["name"] for lbl in raw.get("labels", [])],
+                created_at=raw["created_at"],
+                text=f"{raw['title']}\n\n{body}",
+            ))
+
+        if len(data) < API_PAGE_SIZE:
+            break
+
+        page += 1
+
+    return items
+
+
+class RepoLabel(TypedDict):
+    """A label from the repo with its embedding text."""
+    name: str
+    description: str
+    # "name: description" concatenated for embedding
+    text: str
+
+
+def fetch_repo_labels() -> list[RepoLabel]:
+    """Fetch all labels from the repository.
+
+    Returns labels with name, description, and a text field suitable
+    for embedding ("name: description"). Labels with no description
+    use just the name.
+    """
+    data = github_api_get("/labels?per_page=100")
+    labels: list[RepoLabel] = []
+    for raw in data:
+        name = raw["name"]
+        desc = raw.get("description", "") or ""
+        text = f"{name}: {desc}" if desc else name
+        labels.append(RepoLabel(name=name, description=desc, text=text))
+    return labels
+
+
+def apply_labels_to_item(item_number: int, labels: list[str]) -> None:
+    """Add labels to a single issue/PR via the GitHub API.
+
+    Skips silently if labels list is empty. Uses POST which adds labels
+    without removing existing ones.
+    """
+    if not labels:
+        return
+
+    token = os.environ["GITHUB_TOKEN"]
+    repo = os.environ["GITHUB_REPOSITORY"]
+    url = f"https://api.github.com/repos/{repo}/issues/{item_number}/labels"
+
+    payload = json.dumps({"labels": labels}).encode("utf-8")
+    req = urllib.request.Request(url, data=payload, method="POST")
+    req.add_header("Accept", "application/vnd.github+json")
+    req.add_header("Authorization", f"Bearer {token}")
+    req.add_header("X-GitHub-Api-Version", "2022-11-28")
+    req.add_header("Content-Type", "application/json")
+
+    try:
+        with urllib.request.urlopen(req, timeout=30) as resp:
+            resp.read()
+    except urllib.error.HTTPError as e:
+        body = e.read().decode("utf-8", errors="replace")
+        # Non-fatal: log warning but don't abort the sweep
+        print(f"::warning::Failed to label #{item_number}: {e.code} {body}")
+
+
+def generate_report(
+    items: list[TriageItem],
+    outlier_indices: list[int],
+    duplicate_pairs: list[tuple[int, int, float]],
+    label_suggestions: list[list[tuple[str, float]]] | None = None,
+) -> str:
+    """Generate a structured markdown triage report."""
+    now = datetime.now(timezone.utc).strftime("%Y-%m-%d %H:%M:%S")
+    repo = os.environ.get("GITHUB_REPOSITORY", "unknown/repo")
+
+    lines: list[str] = [
+        "## Triage Sweep Report",
+        "",
+        f"**Run:** {now} UTC",
+        f"**Items analyzed:** {len(items)}",
+        f"**Thresholds:** Mahalanobis > {MAHALANOBIS_THRESHOLD}, Cosine > {COSINE_THRESHOLD}",
+        "",
+        f"### Potential Outliers / Spam ({len(outlier_indices)})",
+        "",
+        "Items with unusually high Mahalanobis distance from the distribution center.",
+        "These may be spam, off-topic, or poorly described.",
+        "",
+    ]
+
+    if outlier_indices:
+        lines.append("| # | Type | Title | Distance |")
+        lines.append("|---|------|-------|----------|")
+        for idx in outlier_indices:
+            item = items[idx]
+            kind = "PR" if item["is_pr"] else "Issue"
+            lines.append(
+                f"| [#{item['number']}]({item['html_url']}) "
+                f"| {kind} | {item['title']} | flagged |"
+            )
+    else:
+        lines.append("None found.")
+
+    lines.extend([
+        "",
+        f"### Potential Duplicates ({len(duplicate_pairs)} pairs)",
+        "",
+        "Pairs of items with cosine similarity above the threshold.",
+        "",
+    ])
+
+    if duplicate_pairs:
+        lines.append("| Item A | Item B | Similarity |")
+        lines.append("|--------|--------|------------|")
+        for i, j, sim in duplicate_pairs:
+            a = items[i]
+            b = items[j]
+            kind_a = "PR" if a["is_pr"] else "Issue"
+            kind_b = "PR" if b["is_pr"] else "Issue"
+            lines.append(
+                f"| [#{a['number']}]({a['html_url']}) {kind_a}: {a['title']} "
+                f"| [#{b['number']}]({b['html_url']}) {kind_b}: {b['title']} "
+                f"| {sim:.3f} |"
+            )
+    else:
+        lines.append("None found.")
+
+    # ── Label suggestions section ────────────────────────────────────
+    outlier_set = set(outlier_indices)
+    if label_suggestions is not None:
+        # Only unlabeled, non-outlier items — spam shouldn't get categorized
+        items_with_suggestions = [
+            (i, sugs) for i, sugs in enumerate(label_suggestions)
+            if sugs and not items[i]["labels"] and i not in outlier_set
+        ]
+        lines.extend([
+            "",
+            f"### Suggested Labels ({len(items_with_suggestions)} unlabeled items)",
+            "",
+            "Labels suggested by embedding similarity against repo label descriptions.",
+            "Only shown for unlabeled items that were not flagged as outliers.",
+            "",
+        ])
+
+        if items_with_suggestions:
+            lines.append("| # | Type | Title | Suggested Labels |")
+            lines.append("|---|------|-------|-----------------|")
+            for idx, sugs in items_with_suggestions:
+                item = items[idx]
+                kind = "PR" if item["is_pr"] else "Issue"
+                label_strs = [f"`{name}` ({score:.2f})" for name, score in sugs]
+                lines.append(
+                    f"| [#{item['number']}]({item['html_url']}) "
+                    f"| {kind} | {item['title']} | {', '.join(label_strs)} |"
+                )
+        else:
+            lines.append("No unlabeled items need suggestions.")
+
+    lines.extend([
+        "",
+        "### Summary",
+        "",
+        f"- {len(outlier_indices)} outliers flagged for review",
+        f"- {len(duplicate_pairs)} duplicate pairs found",
+        f"- {len(items)} items analyzed in total",
+    ])
+
+    if label_suggestions is not None:
+        applied = sum(
+            1 for i, s in enumerate(label_suggestions)
+            if s and not items[i]["labels"] and i not in outlier_set
+        )
+        lines.append(f"- {applied} items suggested for labeling")
+
+    lines.extend([
+        "",
+        "---",
+        f"*Generated by [triage-sweep](https://github.com/{repo}/actions) — no LLM was used.*",
+    ])
+
+    return "\n".join(lines)
+
+
+def create_report_issue(report_body: str) -> None:
+    """Create a GitHub issue with the triage report.
+
+    Posts to the issues API with the triage-report label.
+    Raises SystemExit on non-201 response.
+    """
+    token = os.environ["GITHUB_TOKEN"]
+    repo = os.environ["GITHUB_REPOSITORY"]
+    url = f"https://api.github.com/repos/{repo}/issues"
+
+    today = datetime.now(timezone.utc).strftime("%Y-%m-%d")
+    payload = json.dumps({
+        "title": f"Triage Sweep Report — {today}",
+        "body": report_body,
+        "labels": [REPORT_LABEL],
+    }).encode("utf-8")
+
+    req = urllib.request.Request(url, data=payload, method="POST")
+    req.add_header("Accept", "application/vnd.github+json")
+    req.add_header("Authorization", f"Bearer {token}")
+    req.add_header("X-GitHub-Api-Version", "2022-11-28")
+    req.add_header("Content-Type", "application/json")
+
+    try:
+        with urllib.request.urlopen(req, timeout=30) as resp:
+            resp_body = resp.read().decode("utf-8")
+            if resp.status != 201:
+                print(f"::error::Failed to create issue: {resp.status} {resp_body}")
+                sys.exit(1)
+            result = json.loads(resp_body)
+            print(f"Created issue: {result.get('html_url', 'unknown')}")
+    except urllib.error.HTTPError as e:
+        body = e.read().decode("utf-8", errors="replace")
+        print(f"::error::Failed to create issue: {e.code} {body}")
+        sys.exit(1)
+
+
+def write_report(report: str) -> None:
+    """Write the report to the file system for the summary step."""
+    with open(REPORT_FILE, "w", encoding="utf-8") as f:
+        f.write(report)
+
+
+def main() -> None:
+    """Orchestrate the full triage sweep."""
+    # 1. Validate environment
+    for var in ("GITHUB_TOKEN", "GITHUB_REPOSITORY"):
+        if not os.environ.get(var):
+            print(f"::error::Missing required environment variable: {var}")
+            sys.exit(1)
+
+    # 2. Fetch all open issues + PRs
+    items = fetch_all_open_items()
+    print(f"Fetched {len(items)} open items")
+
+    if len(items) == 0:
+        report = "## Triage Sweep Report\n\nNo open issues or PRs found."
+        write_report(report)
+        print("No items to analyze.")
+        return
+
+    # 3. Extract texts for embedding
+    texts: list[str] = [item["text"] for item in items]
+
+    # 4. Embed all texts (returns numpy float32 array of shape [n, 384])
+    embeddings = embed_texts(texts)
+
+    # 5. L2-normalize
+    embeddings = normalize_rows(embeddings)
+
+    # 6. Outlier detection (Mahalanobis via EllipticEnvelope)
+    outlier_indices: list[int] = []
+    if len(items) >= MIN_SAMPLES_FOR_OUTLIER_DETECTION:
+        reduced = reduce_dimensions(embeddings, PCA_VARIANCE_RATIO, PCA_MAX_COMPONENTS)
+        outlier_indices = detect_outliers(reduced, MAHALANOBIS_THRESHOLD)
+    else:
+        print(
+            f"Skipping outlier detection: {len(items)} items < "
+            f"{MIN_SAMPLES_FOR_OUTLIER_DETECTION} minimum"
+        )
+
+    # 7. Duplicate detection (pairwise cosine similarity)
+    duplicate_pairs = find_duplicate_pairs(embeddings, COSINE_THRESHOLD)
+
+    # 8. Label suggestion via embedding similarity
+    label_suggestions: list[list[tuple[str, float]]] | None = None
+    repo_labels = fetch_repo_labels()
+    if repo_labels:
+        label_texts = [lbl["text"] for lbl in repo_labels]
+        label_names = [lbl["name"] for lbl in repo_labels]
+        label_embeddings = embed_texts(label_texts)
+        label_embeddings = normalize_rows(label_embeddings)
+        label_suggestions = suggest_labels(embeddings, label_embeddings, label_names)
+        print(f"Computed label suggestions against {len(repo_labels)} repo labels")
+
+        # Apply top label to unlabeled items (unless dry run)
+        # Skip outliers — flagged items shouldn't get categorized
+        outlier_set = set(outlier_indices)
+        if not DRY_RUN:
+            applied_count = 0
+            for i, sugs in enumerate(label_suggestions):
+                if sugs and not items[i]["labels"] and i not in outlier_set:
+                    # Apply only the top-1 label (highest confidence)
+                    apply_labels_to_item(items[i]["number"], [sugs[0][0]])
+                    applied_count += 1
+            print(f"Applied labels to {applied_count} unlabeled items")
+    else:
+        print("No repo labels found — skipping label suggestions")
+
+    # 9. Generate report
+    report = generate_report(items, outlier_indices, duplicate_pairs, label_suggestions)
+
+    # 10. Write report to file (for summary step)
+    write_report(report)
+
+    # 11. Create report issue (unless dry run)
+    if DRY_RUN:
+        print("Dry run — skipping issue creation and label application.")
+        print(report)
+    else:
+        create_report_issue(report)
+        print("Report issue created.")
+
+
+if __name__ == "__main__":
+    main()

.github/scripts/triage/test_embedding_utils.py

@@ -0,0 +1,311 @@
+"""Tests for embedding_utils.py — all embedding model calls are mocked."""
+from __future__ import annotations
+
+import sys
+from unittest.mock import patch, MagicMock
+import numpy as np
+import pytest
+
+# Mock fastembed before importing the module under test (persistent)
+if "fastembed" not in sys.modules:
+    sys.modules["fastembed"] = MagicMock()
+
+from embedding_utils import (
+    embed_texts,
+    normalize_rows,
+    reduce_dimensions,
+    detect_outliers,
+    find_duplicate_pairs,
+    suggest_labels,
+    EMBEDDING_DIM,
+    EMBEDDING_MODEL,
+    EMBEDDING_BATCH_SIZE,
+    LABEL_SIMILARITY_THRESHOLD,
+    MAX_LABELS_PER_ITEM,
+)
+
+
+class TestEmbedTexts:
+    """Tests for the embed_texts function."""
+
+    def test_empty_list_returns_empty_array(self):
+        result = embed_texts([])
+        assert result.shape == (0, EMBEDDING_DIM)
+        assert result.dtype == np.float32
+
+    @patch("embedding_utils.TextEmbedding")
+    def test_single_text(self, mock_cls):
+        mock_model = MagicMock()
+        mock_cls.return_value = mock_model
+        vec = np.random.randn(EMBEDDING_DIM).astype(np.float32)
+        mock_model.embed.return_value = iter([vec])
+
+        result = embed_texts(["hello world"])
+
+        mock_cls.assert_called_once_with(model_name=EMBEDDING_MODEL)
+        mock_model.embed.assert_called_once_with(
+            ["hello world"], batch_size=EMBEDDING_BATCH_SIZE
+        )
+        assert result.shape == (1, EMBEDDING_DIM)
+        assert result.dtype == np.float32
+        np.testing.assert_array_almost_equal(result[0], vec)
+
+    @patch("embedding_utils.TextEmbedding")
+    def test_multiple_texts(self, mock_cls):
+        mock_model = MagicMock()
+        mock_cls.return_value = mock_model
+        vecs = [
+            np.random.randn(EMBEDDING_DIM).astype(np.float32)
+            for _ in range(5)
+        ]
+        mock_model.embed.return_value = iter(vecs)
+
+        result = embed_texts(["a", "b", "c", "d", "e"])
+        assert result.shape == (5, EMBEDDING_DIM)
+        assert result.dtype == np.float32
+
+
+class TestNormalizeRows:
+    """Tests for L2 row normalization."""
+
+    def test_empty_matrix(self):
+        m = np.empty((0, 10), dtype=np.float32)
+        result = normalize_rows(m)
+        assert result.shape == (0, 10)
+
+    def test_single_row(self):
+        m = np.array([[3.0, 4.0]], dtype=np.float32)
+        result = normalize_rows(m)
+        # Norm should be ~1.0
+        norm = np.linalg.norm(result[0])
+        assert abs(norm - 1.0) < 1e-5
+
+    def test_multiple_rows(self):
+        rng = np.random.default_rng(42)
+        m = rng.standard_normal((10, 50)).astype(np.float32)
+        result = normalize_rows(m)
+        norms = np.linalg.norm(result, axis=1)
+        np.testing.assert_allclose(norms, 1.0, atol=1e-5)
+
+    def test_zero_row_stays_near_zero(self):
+        m = np.array([[0.0, 0.0, 0.0], [1.0, 0.0, 0.0]], dtype=np.float32)
+        result = normalize_rows(m)
+        # Zero row divided by eps -> very small values
+        assert np.linalg.norm(result[0]) < 1e-3
+        # Non-zero row should be unit norm
+        assert abs(np.linalg.norm(result[1]) - 1.0) < 1e-5
+
+    def test_preserves_direction(self):
+        m = np.array([[2.0, 0.0], [0.0, 3.0]], dtype=np.float32)
+        result = normalize_rows(m)
+        np.testing.assert_allclose(result[0], [1.0, 0.0], atol=1e-5)
+        np.testing.assert_allclose(result[1], [0.0, 1.0], atol=1e-5)
+
+
+class TestReduceDimensions:
+    """Tests for PCA dimensionality reduction."""
+
+    def test_single_sample_returns_unchanged(self):
+        m = np.random.randn(1, 50).astype(np.float32)
+        result = reduce_dimensions(m, 0.95, 10)
+        np.testing.assert_array_equal(result, m)
+
+    def test_reduces_dimensions(self):
+        rng = np.random.default_rng(42)
+        m = rng.standard_normal((100, 50)).astype(np.float32)
+        result = reduce_dimensions(m, 0.95, 10)
+        assert result.shape == (100, 10)
+        assert result.dtype == np.float32
+
+    def test_caps_at_n_minus_1(self):
+        rng = np.random.default_rng(42)
+        # 5 samples, 20 features -> max components = 4 (n-1)
+        m = rng.standard_normal((5, 20)).astype(np.float32)
+        result = reduce_dimensions(m, 0.95, 50)
+        assert result.shape == (5, 4)
+
+    def test_caps_at_d(self):
+        rng = np.random.default_rng(42)
+        # 100 samples, 3 features -> max components = 3
+        m = rng.standard_normal((100, 3)).astype(np.float32)
+        result = reduce_dimensions(m, 0.95, 50)
+        assert result.shape == (100, 3)
+
+    def test_max_components_respected(self):
+        rng = np.random.default_rng(42)
+        m = rng.standard_normal((50, 30)).astype(np.float32)
+        result = reduce_dimensions(m, 0.95, 5)
+        assert result.shape[1] == 5
+
+
+class TestDetectOutliers:
+    """Tests for Mahalanobis-based outlier detection."""
+
+    def test_single_sample_returns_empty(self):
+        m = np.random.randn(1, 5).astype(np.float32)
+        result = detect_outliers(m, 3.0)
+        assert result == []
+
+    def test_empty_returns_empty(self):
+        # n < 2 case
+        m = np.empty((0, 5), dtype=np.float32)
+        result = detect_outliers(m, 3.0)
+        assert result == []
+
+    def test_finds_outliers_in_synthetic_data(self):
+        rng = np.random.default_rng(42)
+        # Create a tight cluster with one obvious outlier
+        cluster = rng.standard_normal((50, 3)).astype(np.float32) * 0.1
+        outlier = np.array([[100.0, 100.0, 100.0]], dtype=np.float32)
+        m = np.vstack([cluster, outlier])
+        result = detect_outliers(m, 3.0)
+        # The outlier (index 50) should be detected
+        assert 50 in result
+
+    def test_returns_list_of_ints(self):
+        rng = np.random.default_rng(42)
+        m = rng.standard_normal((20, 3)).astype(np.float32)
+        result = detect_outliers(m, 3.0)
+        assert isinstance(result, list)
+        for idx in result:
+            assert isinstance(idx, (int, np.integer))
+
+    def test_low_threshold_flags_more(self):
+        rng = np.random.default_rng(42)
+        m = rng.standard_normal((30, 3)).astype(np.float32)
+        low = detect_outliers(m, 1.0)
+        high = detect_outliers(m, 10.0)
+        assert len(low) >= len(high)
+
+
+class TestFindDuplicatePairs:
+    """Tests for cosine similarity duplicate detection."""
+
+    def test_single_item_returns_empty(self):
+        m = np.random.randn(1, 10).astype(np.float32)
+        result = find_duplicate_pairs(m, 0.9)
+        assert result == []
+
+    def test_empty_returns_empty(self):
+        m = np.empty((0, 10), dtype=np.float32)
+        result = find_duplicate_pairs(m, 0.9)
+        assert result == []
+
+    def test_identical_vectors_detected(self):
+        vec = np.random.randn(10).astype(np.float32)
+        vec = vec / np.linalg.norm(vec)
+        m = np.vstack([vec, vec, np.random.randn(10).astype(np.float32)])
+        result = find_duplicate_pairs(m, 0.99)
+        # Items 0 and 1 are identical, should be found
+        assert any(i == 0 and j == 1 for i, j, _ in result)
+
+    def test_orthogonal_vectors_not_detected(self):
+        m = np.eye(5, dtype=np.float32)
+        result = find_duplicate_pairs(m, 0.5)
+        assert result == []
+
+    def test_returns_correct_format(self):
+        vec = np.random.randn(10).astype(np.float32)
+        vec = vec / np.linalg.norm(vec)
+        m = np.vstack([vec, vec])
+        result = find_duplicate_pairs(m, 0.5)
+        assert len(result) >= 1
+        for item in result:
+            assert len(item) == 3
+            i, j, sim = item
+            assert isinstance(i, int)
+            assert isinstance(j, int)
+            assert isinstance(sim, float)
+            assert i < j
+
+    def test_i_less_than_j(self):
+        rng = np.random.default_rng(42)
+        # Create some similar vectors
+        base = rng.standard_normal(10).astype(np.float32)
+        m = np.vstack([base + rng.standard_normal(10) * 0.01 for _ in range(5)])
+        result = find_duplicate_pairs(m, 0.5)
+        for i, j, _ in result:
+            assert i < j
+
+    def test_high_threshold_fewer_pairs(self):
+        rng = np.random.default_rng(42)
+        m = rng.standard_normal((10, 20)).astype(np.float32)
+        # Normalize for meaningful cosine similarities
+        norms = np.linalg.norm(m, axis=1, keepdims=True)
+        m = m / norms
+        low = find_duplicate_pairs(m, 0.3)
+        high = find_duplicate_pairs(m, 0.9)
+        assert len(low) >= len(high)
+
+
+class TestSuggestLabels:
+    """Tests for embedding-based label suggestion."""
+
+    def test_empty_items_returns_empty_lists(self):
+        items = np.empty((0, 10), dtype=np.float32)
+        labels = np.random.randn(3, 10).astype(np.float32)
+        result = suggest_labels(items, labels, ["a", "b", "c"])
+        assert result == []
+
+    def test_empty_labels_returns_empty_per_item(self):
+        items = np.random.randn(5, 10).astype(np.float32)
+        labels = np.empty((0, 10), dtype=np.float32)
+        result = suggest_labels(items, labels, [])
+        assert len(result) == 5
+        assert all(s == [] for s in result)
+
+    def test_identical_embedding_gets_that_label(self):
+        """If an item embedding equals a label embedding, it should suggest that label."""
+        vec = np.array([1.0, 0.0, 0.0], dtype=np.float32)
+        items = np.array([vec], dtype=np.float32)
+        labels = np.array([vec, [0, 1, 0], [0, 0, 1]], dtype=np.float32)
+        result = suggest_labels(items, labels, ["bug", "feature", "docs"], threshold=0.5)
+        assert len(result) == 1
+        assert result[0][0][0] == "bug"
+        assert result[0][0][1] > 0.99
+
+    def test_threshold_filters_low_similarity(self):
+        """With a high threshold, orthogonal vectors should get no suggestions."""
+        items = np.eye(3, dtype=np.float32)
+        labels = np.eye(3, dtype=np.float32)
+        # threshold=0.99 means only near-exact matches
+        result = suggest_labels(items, labels, ["a", "b", "c"], threshold=0.99)
+        # Each item should match exactly one label (itself)
+        for sugs in result:
+            assert len(sugs) == 1
+
+    def test_max_per_item_respected(self):
+        """Even if all labels are similar, max_per_item caps the results."""
+        rng = np.random.default_rng(42)
+        base = rng.standard_normal(10).astype(np.float32)
+        items = np.array([base])
+        # All labels very similar to item
+        labels = np.array([base + rng.standard_normal(10) * 0.01 for _ in range(10)])
+        names = [f"label-{i}" for i in range(10)]
+        result = suggest_labels(items, labels, names, threshold=0.1, max_per_item=2)
+        assert len(result[0]) <= 2
+
+    def test_returns_sorted_by_similarity_descending(self):
+        """Suggestions should be ordered highest similarity first."""
+        items = np.array([[1.0, 0.5, 0.0]], dtype=np.float32)
+        labels = np.array([
+            [1.0, 0.0, 0.0],  # decent match
+            [1.0, 0.5, 0.0],  # exact match
+            [0.0, 0.0, 1.0],  # poor match
+        ], dtype=np.float32)
+        result = suggest_labels(items, labels, ["a", "b", "c"], threshold=0.1)
+        scores = [s for _, s in result[0]]
+        assert scores == sorted(scores, reverse=True)
+
+    def test_returns_correct_format(self):
+        rng = np.random.default_rng(42)
+        items = rng.standard_normal((3, 10)).astype(np.float32)
+        labels = rng.standard_normal((5, 10)).astype(np.float32)
+        names = ["bug", "feature", "docs", "ci", "test"]
+        result = suggest_labels(items, labels, names, threshold=0.0)
+        assert len(result) == 3
+        for sugs in result:
+            for name, score in sugs:
+                assert isinstance(name, str)
+                assert isinstance(score, float)
+                assert name in names

.github/scripts/triage/test_sweep.py

@@ -0,0 +1,616 @@
+"""Tests for sweep.py — all external calls (API, embedding) are mocked."""
+from __future__ import annotations
+
+import json
+import os
+import sys
+from io import BytesIO
+from unittest.mock import patch, MagicMock, mock_open
+from urllib.error import HTTPError
+
+import numpy as np
+import pytest
+
+# Mock fastembed before importing sweep (which imports embedding_utils)
+sys.modules["fastembed"] = MagicMock()
+
+# Set required env vars before importing sweep (module-level constants read env)
+os.environ.setdefault("GITHUB_TOKEN", "test-token")
+os.environ.setdefault("GITHUB_REPOSITORY", "owner/repo")
+
+from sweep import (
+    github_api_get,
+    fetch_all_open_items,
+    fetch_repo_labels,
+    apply_labels_to_item,
+    generate_report,
+    create_report_issue,
+    write_report,
+    main,
+    TriageItem,
+    RepoLabel,
+    REPORT_FILE,
+    REPORT_LABEL,
+    API_PAGE_SIZE,
+    MIN_SAMPLES_FOR_OUTLIER_DETECTION,
+)
+
+
+def _make_api_issue(number: int, title: str = "Test issue", is_pr: bool = False,
+                    body: str = "Issue body", labels: list[str] | None = None) -> dict:
+    """Helper to build a mock GitHub API issue response object."""
+    result: dict = {
+        "number": number,
+        "title": title,
+        "html_url": f"https://github.com/owner/repo/issues/{number}",
+        "body": body,
+        "created_at": "2026-03-21T00:00:00Z",
+        "labels": [{"name": lbl} for lbl in (labels or [])],
+    }
+    if is_pr:
+        result["pull_request"] = {"url": "..."}
+    return result
+
+
+class TestGithubApiGet:
+    """Tests for the github_api_get function."""
+
+    @patch("sweep.urllib.request.urlopen")
+    def test_successful_request(self, mock_urlopen):
+        mock_resp = MagicMock()
+        mock_resp.read.return_value = json.dumps([{"id": 1}]).encode()
+        mock_resp.__enter__ = lambda s: s
+        mock_resp.__exit__ = MagicMock(return_value=False)
+        mock_urlopen.return_value = mock_resp
+
+        result = github_api_get("/issues?state=open")
+        assert result == [{"id": 1}]
+
+    @patch("sweep.urllib.request.urlopen")
+    def test_http_error_exits(self, mock_urlopen):
+        error = HTTPError(
+            url="https://api.github.com/repos/owner/repo/issues",
+            code=403,
+            msg="Forbidden",
+            hdrs=None,  # type: ignore[arg-type]
+            fp=BytesIO(b'{"message": "rate limited"}'),
+        )
+        mock_urlopen.side_effect = error
+
+        with pytest.raises(SystemExit) as exc_info:
+            github_api_get("/issues")
+        assert exc_info.value.code == 1
+
+
+class TestFetchAllOpenItems:
+    """Tests for fetch_all_open_items."""
+
+    @patch("sweep.github_api_get")
+    def test_empty_repo(self, mock_get):
+        mock_get.return_value = []
+        items = fetch_all_open_items()
+        assert items == []
+
+    @patch("sweep.github_api_get")
+    def test_single_page(self, mock_get):
+        mock_get.return_value = [
+            _make_api_issue(1, "Bug report"),
+            _make_api_issue(2, "Feature request", is_pr=True),
+        ]
+        items = fetch_all_open_items()
+        assert len(items) == 2
+        assert items[0]["number"] == 1
+        assert items[0]["is_pr"] is False
+        assert items[1]["is_pr"] is True
+
+    @patch("sweep.github_api_get")
+    def test_text_field_constructed(self, mock_get):
+        mock_get.return_value = [
+            _make_api_issue(1, "My Title", body="My Body"),
+        ]
+        items = fetch_all_open_items()
+        assert items[0]["text"] == "My Title\n\nMy Body"
+
+    @patch("sweep.github_api_get")
+    def test_null_body_handled(self, mock_get):
+        issue = _make_api_issue(1, "No body")
+        issue["body"] = None
+        mock_get.return_value = [issue]
+        items = fetch_all_open_items()
+        assert items[0]["text"] == "No body\n\n"
+
+    @patch("sweep.github_api_get")
+    def test_labels_extracted(self, mock_get):
+        mock_get.return_value = [
+            _make_api_issue(1, "Labeled", labels=["bug", "high-priority"]),
+        ]
+        items = fetch_all_open_items()
+        assert items[0]["labels"] == ["bug", "high-priority"]
+
+    @patch("sweep.MAX_ITEMS", 3)
+    @patch("sweep.github_api_get")
+    def test_max_items_cap(self, mock_get):
+        mock_get.return_value = [_make_api_issue(i) for i in range(100)]
+        items = fetch_all_open_items()
+        assert len(items) == 3
+
+    @patch("sweep.API_PAGE_SIZE", 2)
+    @patch("sweep.github_api_get")
+    def test_pagination(self, mock_get):
+        # First page: 2 items (full page), second page: 1 item (partial -> stop)
+        mock_get.side_effect = [
+            [_make_api_issue(1), _make_api_issue(2)],
+            [_make_api_issue(3)],
+        ]
+        items = fetch_all_open_items()
+        assert len(items) == 3
+        assert mock_get.call_count == 2
+
+
+class TestGenerateReport:
+    """Tests for the markdown report generator."""
+
+    def test_no_findings(self):
+        items = [
+            TriageItem(
+                number=1, title="Test", html_url="https://example.com/1",
+                is_pr=False, labels=[], created_at="2026-01-01", text="Test",
+            ),
+        ]
+        report = generate_report(items, [], [])
+        assert "## Triage Sweep Report" in report
+        assert "Items analyzed:** 1" in report
+        assert "None found." in report
+        assert "0 outliers flagged" in report
+        assert "0 duplicate pairs found" in report
+
+    def test_with_outliers(self):
+        items = [
+            TriageItem(
+                number=10, title="Spam Issue", html_url="https://example.com/10",
+                is_pr=False, labels=[], created_at="2026-01-01", text="spam",
+            ),
+            TriageItem(
+                number=20, title="Good Issue", html_url="https://example.com/20",
+                is_pr=False, labels=[], created_at="2026-01-01", text="good",
+            ),
+        ]
+        report = generate_report(items, [0], [])
+        assert "#10" in report
+        assert "Spam Issue" in report
+        assert "1 outliers flagged" in report
+
+    def test_with_duplicates(self):
+        items = [
+            TriageItem(
+                number=1, title="First", html_url="https://example.com/1",
+                is_pr=False, labels=[], created_at="2026-01-01", text="a",
+            ),
+            TriageItem(
+                number=2, title="Second", html_url="https://example.com/2",
+                is_pr=True, labels=[], created_at="2026-01-01", text="b",
+            ),
+        ]
+        report = generate_report(items, [], [(0, 1, 0.954)])
+        assert "#1" in report
+        assert "#2" in report
+        assert "0.954" in report
+        assert "1 duplicate pairs found" in report
+
+    def test_pr_type_label(self):
+        items = [
+            TriageItem(
+                number=5, title="PR Title", html_url="https://example.com/5",
+                is_pr=True, labels=[], created_at="2026-01-01", text="pr",
+            ),
+        ]
+        report = generate_report(items, [0], [])
+        assert "| PR |" in report
+
+    def test_footer_present(self):
+        items = [
+            TriageItem(
+                number=1, title="T", html_url="u",
+                is_pr=False, labels=[], created_at="d", text="t",
+            ),
+        ]
+        report = generate_report(items, [], [])
+        assert "no LLM was used" in report
+
+
+class TestCreateReportIssue:
+    """Tests for creating the report GitHub issue."""
+
+    @patch("sweep.urllib.request.urlopen")
+    def test_successful_creation(self, mock_urlopen):
+        mock_resp = MagicMock()
+        mock_resp.status = 201
+        mock_resp.read.return_value = json.dumps({
+            "html_url": "https://github.com/owner/repo/issues/99",
+        }).encode()
+        mock_resp.__enter__ = lambda s: s
+        mock_resp.__exit__ = MagicMock(return_value=False)
+        mock_urlopen.return_value = mock_resp
+
+        # Should not raise
+        create_report_issue("# Test Report")
+
+    @patch("sweep.urllib.request.urlopen")
+    def test_http_error_exits(self, mock_urlopen):
+        error = HTTPError(
+            url="https://api.github.com/repos/owner/repo/issues",
+            code=422,
+            msg="Unprocessable",
+            hdrs=None,  # type: ignore[arg-type]
+            fp=BytesIO(b'{"message": "validation failed"}'),
+        )
+        mock_urlopen.side_effect = error
+
+        with pytest.raises(SystemExit) as exc_info:
+            create_report_issue("# Test Report")
+        assert exc_info.value.code == 1
+
+
+class TestWriteReport:
+    """Tests for the write_report helper."""
+
+    @patch("builtins.open", mock_open())
+    def test_writes_to_file(self):
+        write_report("# Report Content")
+        from builtins import open as builtin_open  # noqa
+        # Verify open was called with the right path
+        from unittest.mock import call
+        open_mock = open  # The patched version
+        open_mock.assert_called_once_with(REPORT_FILE, "w", encoding="utf-8")  # type: ignore[attr-defined]
+        open_mock().write.assert_called_once_with("# Report Content")  # type: ignore[attr-defined]
+
+
+class TestFetchRepoLabels:
+    """Tests for fetch_repo_labels."""
+
+    @patch("sweep.github_api_get")
+    def test_fetches_and_constructs_labels(self, mock_get):
+        mock_get.return_value = [
+            {"name": "bug", "description": "Something isn't working"},
+            {"name": "enhancement", "description": "New feature or request"},
+            {"name": "docs", "description": ""},
+        ]
+        labels = fetch_repo_labels()
+        assert len(labels) == 3
+        assert labels[0]["name"] == "bug"
+        assert labels[0]["text"] == "bug: Something isn't working"
+        assert labels[2]["text"] == "docs"  # no description, just name
+
+    @patch("sweep.github_api_get")
+    def test_empty_repo_labels(self, mock_get):
+        mock_get.return_value = []
+        labels = fetch_repo_labels()
+        assert labels == []
+
+    @patch("sweep.github_api_get")
+    def test_null_description_handled(self, mock_get):
+        mock_get.return_value = [
+            {"name": "wontfix", "description": None},
+        ]
+        labels = fetch_repo_labels()
+        assert labels[0]["text"] == "wontfix"
+
+
+class TestApplyLabelsToItem:
+    """Tests for apply_labels_to_item."""
+
+    def test_empty_labels_skips(self):
+        # Should not make any API call
+        apply_labels_to_item(1, [])
+
+    @patch("sweep.urllib.request.urlopen")
+    def test_successful_label_application(self, mock_urlopen):
+        mock_resp = MagicMock()
+        mock_resp.read.return_value = b'[{"name": "bug"}]'
+        mock_resp.__enter__ = lambda s: s
+        mock_resp.__exit__ = MagicMock(return_value=False)
+        mock_urlopen.return_value = mock_resp
+
+        # Should not raise
+        apply_labels_to_item(42, ["bug", "enhancement"])
+
+    @patch("sweep.urllib.request.urlopen")
+    def test_http_error_is_non_fatal(self, mock_urlopen):
+        error = HTTPError(
+            url="https://api.github.com/repos/owner/repo/issues/1/labels",
+            code=404,
+            msg="Not Found",
+            hdrs=None,  # type: ignore[arg-type]
+            fp=BytesIO(b'{"message": "not found"}'),
+        )
+        mock_urlopen.side_effect = error
+
+        # Should NOT raise — labeling failures are warnings, not fatal
+        apply_labels_to_item(1, ["bug"])
+
+
+class TestGenerateReportWithLabels:
+    """Tests for label suggestions in the report."""
+
+    def test_report_includes_label_section(self):
+        items = [
+            TriageItem(
+                number=1, title="Fix crash", html_url="https://example.com/1",
+                is_pr=False, labels=[], created_at="2026-01-01", text="crash",
+            ),
+        ]
+        suggestions = [[("bug", 0.85), ("enhancement", 0.42)]]
+        report = generate_report(items, [], [], label_suggestions=suggestions)
+        assert "Suggested Labels" in report
+        assert "`bug` (0.85)" in report
+        assert "1 items suggested for labeling" in report
+
+    def test_report_skips_already_labeled_items(self):
+        items = [
+            TriageItem(
+                number=1, title="Already labeled", html_url="https://example.com/1",
+                is_pr=False, labels=["bug"], created_at="2026-01-01", text="bug",
+            ),
+        ]
+        suggestions = [[("bug", 0.95)]]
+        report = generate_report(items, [], [], label_suggestions=suggestions)
+        assert "0 items suggested for labeling" in report
+        assert "No unlabeled items" in report
+
+    def test_report_excludes_outliers_from_suggestions(self):
+        items = [
+            TriageItem(
+                number=1, title="Spam garbage", html_url="https://example.com/1",
+                is_pr=False, labels=[], created_at="2026-01-01", text="spam",
+            ),
+            TriageItem(
+                number=2, title="Real bug", html_url="https://example.com/2",
+                is_pr=False, labels=[], created_at="2026-01-01", text="bug",
+            ),
+        ]
+        suggestions = [[("bug", 0.85)], [("bug", 0.90)]]
+        # Item 0 is an outlier — should be excluded from label suggestions
+        report = generate_report(items, [0], [], label_suggestions=suggestions)
+        assert "1 unlabeled items" in report  # only item 2
+        assert "#2" in report
+        # Item 1 (outlier) should NOT be in the suggestions table
+        assert "Spam garbage" not in report.split("Suggested Labels")[1]
+
+    def test_report_without_label_suggestions(self):
+        items = [
+            TriageItem(
+                number=1, title="T", html_url="u",
+                is_pr=False, labels=[], created_at="d", text="t",
+            ),
+        ]
+        report = generate_report(items, [], [], label_suggestions=None)
+        assert "Suggested Labels" not in report
+
+
+class TestMain:
+    """Tests for the main orchestration function."""
+
+    @patch.dict(os.environ, {"GITHUB_TOKEN": "", "GITHUB_REPOSITORY": "owner/repo"})
+    def test_missing_token_exits(self):
+        with pytest.raises(SystemExit) as exc_info:
+            main()
+        assert exc_info.value.code == 1
+
+    @patch.dict(os.environ, {"GITHUB_TOKEN": "tok", "GITHUB_REPOSITORY": ""})
+    def test_missing_repo_exits(self):
+        with pytest.raises(SystemExit) as exc_info:
+            main()
+        assert exc_info.value.code == 1
+
+    @patch("sweep.write_report")
+    @patch("sweep.fetch_all_open_items", return_value=[])
+    def test_no_items(self, mock_fetch, mock_write):
+        main()
+        mock_write.assert_called_once()
+        report = mock_write.call_args[0][0]
+        assert "No open issues or PRs found" in report
+
+    @patch("sweep.create_report_issue")
+    @patch("sweep.write_report")
+    @patch("sweep.suggest_labels", return_value=[])
+    @patch("sweep.find_duplicate_pairs", return_value=[])
+    @patch("sweep.detect_outliers", return_value=[])
+    @patch("sweep.reduce_dimensions")
+    @patch("sweep.normalize_rows")
+    @patch("sweep.embed_texts")
+    @patch("sweep.fetch_repo_labels")
+    @patch("sweep.fetch_all_open_items")
+    def test_full_flow_with_enough_items(
+        self, mock_fetch, mock_labels, mock_embed, mock_norm, mock_reduce,
+        mock_outliers, mock_dupes, mock_suggest, mock_write, mock_create,
+    ):
+        """Test the full flow with >= MIN_SAMPLES items (outlier detection runs)."""
+        items = [
+            TriageItem(
+                number=i, title=f"Item {i}", html_url=f"https://example.com/{i}",
+                is_pr=False, labels=[], created_at="2026-01-01", text=f"text {i}",
+            )
+            for i in range(15)
+        ]
+        mock_fetch.return_value = items
+        mock_labels.return_value = [
+            RepoLabel(name="bug", description="Something broken", text="bug: Something broken"),
+        ]
+
+        embeddings = np.random.randn(15, 384).astype(np.float32)
+        mock_embed.return_value = embeddings
+        mock_norm.return_value = embeddings
+        mock_reduce.return_value = np.random.randn(15, 10).astype(np.float32)
+
+        main()
+
+        mock_fetch.assert_called_once()
+        mock_labels.assert_called_once()
+        # embed_texts called twice: once for items, once for labels
+        assert mock_embed.call_count == 2
+        mock_norm.assert_called()
+        mock_reduce.assert_called_once()
+        mock_outliers.assert_called_once()
+        mock_dupes.assert_called_once()
+        mock_suggest.assert_called_once()
+        mock_write.assert_called_once()
+        mock_create.assert_called_once()
+
+    @patch("sweep.create_report_issue")
+    @patch("sweep.write_report")
+    @patch("sweep.suggest_labels", return_value=[])
+    @patch("sweep.find_duplicate_pairs", return_value=[])
+    @patch("sweep.detect_outliers")
+    @patch("sweep.reduce_dimensions")
+    @patch("sweep.normalize_rows")
+    @patch("sweep.embed_texts")
+    @patch("sweep.fetch_repo_labels", return_value=[])
+    @patch("sweep.fetch_all_open_items")
+    def test_skips_outlier_detection_for_few_items(
+        self, mock_fetch, mock_labels, mock_embed, mock_norm, mock_reduce,
+        mock_outliers, mock_dupes, mock_suggest, mock_write, mock_create,
+    ):
+        """With < MIN_SAMPLES items, outlier detection should be skipped."""
+        items = [
+            TriageItem(
+                number=i, title=f"Item {i}", html_url=f"https://example.com/{i}",
+                is_pr=False, labels=[], created_at="2026-01-01", text=f"text {i}",
+            )
+            for i in range(5)
+        ]
+        mock_fetch.return_value = items
+
+        embeddings = np.random.randn(5, 384).astype(np.float32)
+        mock_embed.return_value = embeddings
+        mock_norm.return_value = embeddings
+
+        main()
+
+        # Outlier detection should not have been called
+        mock_reduce.assert_not_called()
+        mock_outliers.assert_not_called()
+        # But duplicates should still be checked
+        mock_dupes.assert_called_once()
+
+    @patch.dict(os.environ, {"INPUT_DRY_RUN": "true"})
+    @patch("sweep.DRY_RUN", True)
+    @patch("sweep.write_report")
+    @patch("sweep.create_report_issue")
+    @patch("sweep.apply_labels_to_item")
+    @patch("sweep.suggest_labels", return_value=[[("bug", 0.85)]])
+    @patch("sweep.find_duplicate_pairs", return_value=[])
+    @patch("sweep.normalize_rows")
+    @patch("sweep.embed_texts")
+    @patch("sweep.fetch_repo_labels")
+    @patch("sweep.fetch_all_open_items")
+    def test_dry_run_skips_issue_creation_and_labeling(
+        self, mock_fetch, mock_labels, mock_embed, mock_norm,
+        mock_dupes, mock_suggest, mock_apply, mock_create, mock_write,
+    ):
+        items = [
+            TriageItem(
+                number=1, title="Item", html_url="https://example.com/1",
+                is_pr=False, labels=[], created_at="2026-01-01", text="text",
+            )
+        ]
+        mock_fetch.return_value = items
+        mock_labels.return_value = [
+            RepoLabel(name="bug", description="Broken", text="bug: Broken"),
+        ]
+        embeddings = np.random.randn(1, 384).astype(np.float32)
+        mock_embed.return_value = embeddings
+        mock_norm.return_value = embeddings
+
+        main()
+
+        mock_create.assert_not_called()
+        mock_apply.assert_not_called()
+        mock_write.assert_called_once()
+
+    @patch("sweep.create_report_issue")
+    @patch("sweep.write_report")
+    @patch("sweep.apply_labels_to_item")
+    @patch("sweep.suggest_labels")
+    @patch("sweep.find_duplicate_pairs", return_value=[])
+    @patch("sweep.normalize_rows")
+    @patch("sweep.embed_texts")
+    @patch("sweep.fetch_repo_labels")
+    @patch("sweep.fetch_all_open_items")
+    def test_applies_labels_to_unlabeled_items(
+        self, mock_fetch, mock_labels, mock_embed, mock_norm,
+        mock_dupes, mock_suggest, mock_apply, mock_write, mock_create,
+    ):
+        """When not dry run, top-1 label should be applied to unlabeled items."""
+        items = [
+            TriageItem(
+                number=1, title="Crash bug", html_url="https://example.com/1",
+                is_pr=False, labels=[], created_at="2026-01-01", text="crash",
+            ),
+            TriageItem(
+                number=2, title="Already labeled", html_url="https://example.com/2",
+                is_pr=False, labels=["enhancement"], created_at="2026-01-01", text="feat",
+            ),
+        ]
+        mock_fetch.return_value = items
+        mock_labels.return_value = [
+            RepoLabel(name="bug", description="Broken", text="bug: Broken"),
+        ]
+        mock_suggest.return_value = [
+            [("bug", 0.90)],       # item 1: unlabeled, should get labeled
+            [("bug", 0.45)],       # item 2: already labeled, skip
+        ]
+
+        embeddings = np.random.randn(2, 384).astype(np.float32)
+        mock_embed.return_value = embeddings
+        mock_norm.return_value = embeddings
+
+        main()
+
+        # Only item 1 (unlabeled) should get a label applied
+        mock_apply.assert_called_once_with(1, ["bug"])
+
+    @patch("sweep.create_report_issue")
+    @patch("sweep.write_report")
+    @patch("sweep.apply_labels_to_item")
+    @patch("sweep.suggest_labels")
+    @patch("sweep.find_duplicate_pairs", return_value=[])
+    @patch("sweep.detect_outliers")
+    @patch("sweep.reduce_dimensions")
+    @patch("sweep.normalize_rows")
+    @patch("sweep.embed_texts")
+    @patch("sweep.fetch_repo_labels")
+    @patch("sweep.fetch_all_open_items")
+    def test_outliers_do_not_get_labeled(
+        self, mock_fetch, mock_labels, mock_embed, mock_norm, mock_reduce,
+        mock_outliers, mock_dupes, mock_suggest, mock_apply, mock_write, mock_create,
+    ):
+        """Items flagged as outliers should not receive label suggestions."""
+        items = [
+            TriageItem(
+                number=i, title=f"Item {i}", html_url=f"https://example.com/{i}",
+                is_pr=False, labels=[], created_at="2026-01-01", text=f"text {i}",
+            )
+            for i in range(15)
+        ]
+        mock_fetch.return_value = items
+        mock_labels.return_value = [
+            RepoLabel(name="bug", description="Broken", text="bug: Broken"),
+        ]
+        # Outlier detection flags items 0 and 5
+        mock_outliers.return_value = [0, 5]
+        # Every item gets a suggestion
+        mock_suggest.return_value = [[("bug", 0.85)] for _ in range(15)]
+
+        embeddings = np.random.randn(15, 384).astype(np.float32)
+        mock_embed.return_value = embeddings
+        mock_norm.return_value = embeddings
+        mock_reduce.return_value = np.random.randn(15, 10).astype(np.float32)
+
+        main()
+
+        # Items 0 and 5 are outliers — should NOT be labeled
+        labeled_numbers = [call.args[0] for call in mock_apply.call_args_list]
+        assert 0 not in labeled_numbers
+        assert 5 not in labeled_numbers
+        # Other items should be labeled (13 items: 15 total - 2 outliers)
+        assert mock_apply.call_count == 13

.github/workflows/pr-description-check.yml

@@ -0,0 +1,94 @@
+name: PR Description Check
+
+on:
+  pull_request:
+    types: [opened, edited, reopened]
+    branches: [main]
+
+permissions:
+  pull-requests: write
+
+concurrency:
+  group: pr-desc-${{ github.event.pull_request.number }}
+  cancel-in-progress: true
+
+jobs:
+  check-description:
+    runs-on: ubuntu-latest
+    timeout-minutes: 5
+    steps:
+      - name: Check PR description quality
+        uses: actions/github-script@60a0d83039c74a4aee543508d2ffcb1c3799cdea # v7
+        with:
+          script: |
+            const MIN_BODY_LENGTH = 50;
+            const LABEL = 'needs-description';
+
+            const pr = context.payload.pull_request;
+            const body = (pr.body || '').trim();
+            const owner = context.repo.owner;
+            const repo = context.repo.repo;
+            const number = pr.number;
+
+            const hasLabel = pr.labels.some(l => l.name === LABEL);
+
+            if (body.length < MIN_BODY_LENGTH) {
+              // Add label if not already present
+              if (!hasLabel) {
+                await github.rest.issues.addLabels({
+                  owner, repo, issue_number: number,
+                  labels: [LABEL],
+                });
+              }
+
+              // Post or update a comment
+              const marker = '<!-- pr-desc-check -->';
+              const message = [
+                marker,
+                `### PR description is too short`,
+                '',
+                `This PR's description is **${body.length}** characters, ` +
+                `but the minimum is **${MIN_BODY_LENGTH}**.`,
+                '',
+                'Please update the PR description to explain:',
+                '- **What** this PR changes',
+                '- **Why** the change is needed',
+                '',
+                'Use the PR template as a guide. This check will re-run when you edit the description.',
+              ].join('\n');
+
+              // Find existing bot comment to update (avoid spam)
+              const comments = await github.rest.issues.listComments({
+                owner, repo, issue_number: number,
+              });
+              const existing = comments.data.find(c =>
+                c.body && c.body.includes(marker)
+              );
+
+              if (existing) {
+                await github.rest.issues.updateComment({
+                  owner, repo, comment_id: existing.id,
+                  body: message,
+                });
+              } else {
+                await github.rest.issues.createComment({
+                  owner, repo, issue_number: number,
+                  body: message,
+                });
+              }
+
+              core.setFailed(
+                `PR description is ${body.length} chars (minimum: ${MIN_BODY_LENGTH})`
+              );
+            } else {
+              // Description is acceptable — remove the label if present
+              if (hasLabel) {
+                await github.rest.issues.removeLabel({
+                  owner, repo, issue_number: number,
+                  name: LABEL,
+                }).catch(() => {});
+                // .catch: label may have been removed manually
+              }
+
+              core.info(`PR description OK (${body.length} chars)`);
+            }

.github/workflows/triage-sweep.yml

@@ -0,0 +1,87 @@
+name: Triage Sweep
+
+on:
+  workflow_dispatch:
+    inputs:
+      mahalanobis_threshold:
+        description: >-
+          Mahalanobis distance threshold for outlier detection.
+          Items with distance above this are flagged as potential spam/noise.
+          Lower = more aggressive flagging.
+        type: number
+        default: 3.0
+      cosine_threshold:
+        description: >-
+          Cosine similarity threshold for duplicate detection.
+          Pairs with similarity above this are flagged as potential duplicates.
+          Higher = only very similar pairs flagged.
+        type: number
+        default: 0.92
+      max_items:
+        description: >-
+          Maximum number of open issues + PRs to process.
+          Hard cap to prevent runaway costs on very large repos.
+        type: number
+        default: 500
+      dry_run:
+        description: >-
+          If true, print the report to workflow logs but do not create
+          a GitHub issue.
+        type: boolean
+        default: false
+
+permissions:
+  contents: read
+  issues: write
+
+concurrency:
+  group: triage-sweep
+  cancel-in-progress: true
+
+jobs:
+  sweep:
+    runs-on: ubuntu-latest
+    timeout-minutes: 30
+    steps:
+      - name: Checkout repository
+        uses: actions/checkout@34e114876b0b11c390a56381ad16ebd13914f8d5 # v4
+        with:
+          sparse-checkout: .github/scripts/triage
+          sparse-checkout-cone-mode: false
+          fetch-depth: 1
+
+      - name: Set up Python
+        uses: actions/setup-python@a26af69be951a213d495a4c3e4e4022e16d87065 # v5
+        with:
+          python-version: '3.12'
+          cache: pip
+          cache-dependency-path: .github/scripts/triage/requirements.txt
+
+      - name: Install dependencies
+        run: pip install -r .github/scripts/triage/requirements.txt
+
+      - name: Cache FastEmbed model weights
+        uses: actions/cache@5a3ec84eff668545956fd18022155c47e93e2684 # v4
+        with:
+          path: ~/.cache/fastembed_cache
+          key: fastembed-bge-small-en-v1.5
+
+      - name: Run triage sweep
+        id: sweep
+        env:
+          GITHUB_TOKEN: ${{ secrets.GITHUB_TOKEN }}
+          GITHUB_REPOSITORY: ${{ github.repository }}
+          INPUT_MAHALANOBIS_THRESHOLD: ${{ inputs.mahalanobis_threshold }}
+          INPUT_COSINE_THRESHOLD: ${{ inputs.cosine_threshold }}
+          INPUT_MAX_ITEMS: ${{ inputs.max_items }}
+          INPUT_DRY_RUN: ${{ inputs.dry_run }}
+        run: python .github/scripts/triage/sweep.py
+
+      - name: Post summary
+        if: always()
+        run: |
+          if [ -f /tmp/triage-report.md ]; then
+            cat /tmp/triage-report.md >> "$GITHUB_STEP_SUMMARY"
+          else
+            echo "No report generated." >> "$GITHUB_STEP_SUMMARY"
+          fi

.gitignore

@@ -70,3 +70,5 @@ gitnexus/test/fixtures/lang-resolution/**/bin
 GitNexus.sln
 # Git worktrees
 .worktrees/
+
+/github/scripts/triage/__pycache__/