@zibby/workflow-templates 0.4.1 → 0.7.0

package/sentry-triage/nodes/dispatch-alerts-node.js

@@ -1,191 +0,0 @@
-/**
- * dispatch-alerts node — sub-graph fan-out to notify-slack/notify-lark.
- *
- * For each classified issue at or above severityThreshold:
- *   - Build a provider-neutral notification payload (severity, title,
- *     body, sentryLink, etc.) from the merged issue + classification
- *     records.
- *   - Add caller-supplied per-provider config (channel for Slack,
- *     receiveId for Lark, severity-conditional mentions).
- *   - dispatchSubgraph(state.notifyWorker, { input }) — SYNC mode so we
- *     get back the messageTs/messageId for the summary.
- *   - Continue on per-issue failure (notify failure shouldn't kill the
- *     whole triage run; we report `status: 'failed'` and move on).
- *
- * Sub-graph dispatch goes via in-process executor when the child is
- * bundled in the same Fargate task. The runtime threshold for severity
- * is enforced HERE, not in the LLM classifier, so an operator can
- * raise/lower the bar at deploy time without redeploying.
- */
-
-import { z } from 'zod';
-import { dispatchSubgraph } from '@zibby/agent-workflow';
-import { SEVERITY_LEVELS } from '../state.js';
-
-const DispatchAlertsOutputSchema = z.object({
-  dispatched: z.array(z.object({
-    issueId: z.string(),
-    severity: z.enum(SEVERITY_LEVELS),
-    status: z.enum(['sent', 'skipped', 'failed']),
-    detail: z.string().optional(),
-    messageTs: z.string().optional(),
-    messageId: z.string().optional(),
-  })),
-  summary: z.object({
-    total: z.number(),
-    sent: z.number(),
-    skipped: z.number(),
-    failed: z.number(),
-  }),
-});
-
-const SEVERITY_RANK = Object.freeze({
-  NOISE: 0, LOW: 1, MEDIUM: 2, HIGH: 3, CRITICAL: 4,
-});
-
-/**
- * Build the provider-neutral notification payload from issue + classification.
- * Pure function — exposed for tests so we can pin the wire shape.
- */
-export function buildNotifyPayload({ issue, classification, state }) {
-  const severityRaw = classification?.severity || 'LOW';
-  // notify-* workflows want lowercase severity (their inputSchema enum).
-  const severity = severityRaw.toLowerCase();
-  const reason = classification?.reasoning || '';
-  const userCount = typeof issue.userCount === 'number' ? issue.userCount : undefined;
-  const events = typeof issue.count === 'number' ? issue.count
-                : (typeof issue.count === 'string' && /^\d+$/.test(issue.count)) ? Number(issue.count)
-                : undefined;
-  const release = issue?.metadata?.release
-                || (issue.tags || []).find?.((t) => t.key === 'release')?.value
-                || undefined;
-  const firstSeen = issue.firstSeen || undefined;
-
-  const culpritLine = issue.culprit ? `\n📍 ${issue.culprit}` : '';
-  const reasonLine = reason ? `\n${reason}` : '';
-  const body = `${classification?.suggestedAction ? `*Action:* ${classification.suggestedAction}\n` : ''}` +
-               `${reasonLine}${culpritLine}`.trim();
-
-  // Code snippet: when fetch_issues populated metadata.filename, use it
-  // as a one-line hint. v2 will pull the actual context lines.
-  const codeSnippet = issue?.metadata?.filename
-    ? `${issue.metadata.filename}\n// ${issue.metadata.value || issue.title}`
-    : undefined;
-
-  // Mentions: only attach on CRITICAL. Lower severities get no @-blast.
-  const isCritical = severityRaw === 'CRITICAL';
-  const slackMentions = isCritical && Array.isArray(state.slackMentions) ? state.slackMentions : undefined;
-  const larkMentions  = isCritical && Array.isArray(state.larkMentions)  ? state.larkMentions  : undefined;
-
-  // Common fields both providers accept.
-  const common = {
-    severity,
-    title: issue.title || issue.shortId || `Sentry ${issue.id}`,
-    body: body || undefined,
-    sentryLink: issue.permalink || undefined,
-    affectedUsers: userCount,
-    events,
-    release,
-    firstSeen,
-    codeSnippet,
-  };
-
-  if (state.notifyWorker === 'notify-lark') {
-    return {
-      ...common,
-      receiveId: state.larkReceiveId,
-      ...(larkMentions ? { mentions: larkMentions } : {}),
-    };
-  }
-  // default → notify-slack
-  return {
-    ...common,
-    channel: state.slackChannel,
-    ...(slackMentions ? { mentions: slackMentions } : {}),
-  };
-}
-
-/**
- * Convenience for unit tests + the node body. Decides whether a given
- * classification meets the threshold.
- */
-export function meetsSeverityThreshold(severity, threshold) {
-  const s = SEVERITY_RANK[severity] ?? 0;
-  const t = SEVERITY_RANK[threshold] ?? SEVERITY_RANK.MEDIUM;
-  return s >= t;
-}
-
-export const dispatchAlertsNode = {
-  name: 'dispatch_alerts',
-  outputSchema: DispatchAlertsOutputSchema,
-  timeout: 5 * 60 * 1000, // 5min — generous for fan-out across 20 children
-  execute: async (context) => {
-    const state = (context?.state && typeof context.state.getAll === 'function')
-      ? context.state.getAll()
-      : context;
-
-    const issues = state?.filter_noise?.kept || state?.fetch_issues?.issues || [];
-    const classifications = state?.classify?.classifications || [];
-    const classMap = new Map(classifications.map((c) => [c.issueId, c]));
-    const threshold = state.severityThreshold || 'MEDIUM';
-    const worker = state.notifyWorker || 'notify-slack';
-
-    // Validate the right per-provider config field is set.
-    if (worker === 'notify-slack' && !state.slackChannel) {
-      throw new Error('sentry-triage: slackChannel is required when notifyWorker=notify-slack');
-    }
-    if (worker === 'notify-lark' && !state.larkReceiveId) {
-      throw new Error('sentry-triage: larkReceiveId is required when notifyWorker=notify-lark');
-    }
-
-    const dispatched = [];
-    for (const issue of issues) {
-      const classification = classMap.get(String(issue.id));
-      const severity = classification?.severity || 'LOW';
-
-      if (!meetsSeverityThreshold(severity, threshold)) {
-        dispatched.push({
-          issueId: String(issue.id),
-          severity,
-          status: 'skipped',
-          detail: `severity ${severity} below threshold ${threshold}`,
-        });
-        continue;
-      }
-
-      const payload = buildNotifyPayload({ issue, classification, state });
-      try {
-        // Sync sub-graph dispatch. The in-process executor returns the
-        // child's finalState (or the extracted `output`). We pull
-        // messageTs / messageId by checking both shapes since the parent
-        // is provider-agnostic at this layer.
-        const result = await dispatchSubgraph(worker, {
-          input: payload,
-          async: false,
-        });
-        dispatched.push({
-          issueId: String(issue.id),
-          severity,
-          status: 'sent',
-          messageTs: result?.notify_slack?.messageTs || result?.messageTs,
-          messageId: result?.notify_lark?.messageId  || result?.messageId,
-        });
-      } catch (err) {
-        dispatched.push({
-          issueId: String(issue.id),
-          severity,
-          status: 'failed',
-          detail: err?.message || String(err),
-        });
-      }
-    }
-
-    const summary = {
-      total: dispatched.length,
-      sent: dispatched.filter((d) => d.status === 'sent').length,
-      skipped: dispatched.filter((d) => d.status === 'skipped').length,
-      failed: dispatched.filter((d) => d.status === 'failed').length,
-    };
-    return { dispatched, summary };
-  },
-};

package/sentry-triage/nodes/filter-noise-node.js

@@ -1,112 +0,0 @@
-/**
- * filter-noise node — deterministic regex-based pre-LLM filter.
- *
- * Cuts LLM cost ~80% on a typical Sentry stream by dropping issues
- * that are obviously noise BEFORE we pay for classification. The
- * patterns are deliberately conservative — anything ambiguous goes
- * through to the LLM classifier rather than being killed here.
- *
- * Noise categories (matched in order; first hit wins):
- *
- *   1. Cross-origin / opaque errors:
- *      - "Script error." (literal, with period) — useless without CORS
- *      - "Non-Error promise rejection captured"
- *
- *   2. Browser-internal benign loops:
- *      - "ResizeObserver loop limit exceeded"
- *      - "ResizeObserver loop completed with undelivered notifications"
- *
- *   3. Browser-extension noise — any frame URL with extension scheme:
- *      - chrome-extension://, safari-extension://, moz-extension://
- *
- *   4. Cancelled/aborted requests (user navigated away):
- *      - "AbortError", title containing "cancelled" / "Failed to fetch"
- *        AND empty stack frames
- *
- *   5. Bot/crawler traffic — usually surfaces via specific tag patterns;
- *      v1 doesn't have the tags in our payload shape so we skip. Add
- *      when the event-detail path is wired in.
- *
- * Anything not matching falls through to the LLM. The output carries
- * BOTH the kept list and the dropped list (with reason) so the
- * downstream dispatcher can report on filter activity ("auto-ignored
- * 14 noise issues this hour").
- */
-
-import { z } from 'zod';
-
-const FilterNoiseOutputSchema = z.object({
-  kept: z.array(z.any()).describe('Issues that survive the noise filter — passed to LLM classifier.'),
-  dropped: z.array(z.object({
-    id: z.string(),
-    reason: z.string(),
-  })).describe('Issues filtered out and why — surfaced in the final summary.'),
-});
-
-/** Test hook — exposes the rule table so unit tests can assert
- *  per-pattern matching without re-typing them. */
-export const NOISE_RULES = Object.freeze([
-  { reason: 'cross-origin opaque error',     test: (issue) => /^Script error\.?$/i.test((issue.title || '').trim()) },
-  { reason: 'non-Error promise rejection',   test: (issue) => /Non-Error promise rejection captured/i.test(issue.title || '') },
-  { reason: 'ResizeObserver loop',           test: (issue) => /ResizeObserver loop (limit exceeded|completed)/i.test(issue.title || '') },
-  { reason: 'browser-extension frame',       test: (issue) => isExtensionUrl(issue?.metadata?.filename) || isExtensionUrl(issue?.culprit) },
-  { reason: 'analytics SDK',                 test: (issue) => /\b(gtag|fbq|_paq|dataLayer|googletagmanager|piwik)\b/i.test(`${issue.title || ''} ${issue.culprit || ''}`) },
-  { reason: 'aborted/cancelled request',     test: (issue) => /AbortError|cancelled|Load failed/i.test(issue.title || '') && (!issue.userCount || issue.userCount < 3) },
-]);
-
-function isExtensionUrl(url) {
-  if (!url || typeof url !== 'string') return false;
-  return /^(?:chrome-extension|safari-extension|moz-extension|webkit-masked-url):\/\//.test(url)
-    || /chrome-extension:\/\//.test(url)
-    || /safari-extension:\/\//.test(url);
-}
-
-/**
- * Run the noise filter on an array of issues. Returns
- * `{ kept, dropped }`. Pure function — exposed so tests can call it
- * without going through the node's execute wrapper.
- */
-export function filterNoiseIssues(issues) {
-  const kept = [];
-  const dropped = [];
-  for (const issue of issues || []) {
-    if (!issue || typeof issue !== 'object' || !issue.id) {
-      // Defensive: malformed entries go to dropped rather than crashing
-      // the whole filter run.
-      dropped.push({ id: String(issue?.id || '<unknown>'), reason: 'malformed-issue' });
-      continue;
-    }
-    let matched = null;
-    for (const rule of NOISE_RULES) {
-      try {
-        if (rule.test(issue)) {
-          matched = rule.reason;
-          break;
-        }
-      } catch {
-        // A buggy rule shouldn't break the whole filter; just skip it.
-      }
-    }
-    if (matched) {
-      dropped.push({ id: String(issue.id), reason: matched });
-    } else {
-      kept.push(issue);
-    }
-  }
-  return { kept, dropped };
-}
-
-export const filterNoiseNode = {
-  name: 'filter_noise',
-  outputSchema: FilterNoiseOutputSchema,
-  // 5s is generous for a pure-CPU filter even on 100 issues — the
-  // input is already bounded by maxIssues (default 20).
-  timeout: 5 * 1000,
-  execute: async (context) => {
-    const state = (context?.state && typeof context.state.getAll === 'function')
-      ? context.state.getAll()
-      : context;
-    const issues = state?.fetch_issues?.issues || [];
-    return filterNoiseIssues(issues);
-  },
-};

package/sentry-triage/prompts/classify.md

@@ -1,76 +0,0 @@
-You are the **classify** node of a Sentry triage workflow. Your job is to classify each kept issue (after the `filter_noise` node removed obvious garbage) into a severity bucket and explain WHY.
-
-## Inputs (read from `state`)
-
-- `state.filter_noise.kept` — array of Sentry issue objects that passed the regex noise filter
-- `state.fetch_issues.issues` — (alias) the original list, for cross-referencing if needed
-
-## Severity rubric (apply IN ORDER, stop at first match)
-
-1. **CRITICAL** if ANY of:
-   - `userCount >= 20` (≥ 20 users affected — real prod impact)
-   - `culprit` or `metadata.filename` matches `/payment|billing|checkout|auth|login|signup|session/i` (security/revenue path)
-   - `level === "fatal"` and `count >= 10`
-   - `count >= 100` AND the first-seen-to-last-seen window is < 30 min (active spike)
-
-2. **HIGH** if ANY of:
-   - `userCount >= 5` AND `count >= 50`
-   - `level === "fatal"` (any count)
-   - `level === "error"` AND `userCount >= 3` AND `count >= 20`
-   - Errors in non-critical-but-important paths: settings, profile, search, dashboard, admin
-
-3. **MEDIUM** if ANY of:
-   - `count >= 20` AND `userCount >= 2`
-   - `count >= 50` regardless of userCount
-   - `level === "error"` AND `count >= 10`
-
-4. **LOW** if ANY of:
-   - `count < 20` AND `userCount < 5`
-   - `level === "warning"` or `level === "info"`
-
-5. **NOISE** — only if the issue clearly slipped past the regex filter. Reasons:
-   - Title says "Test ", "Demo ", "[STAGING]" — wrong environment
-   - Stack trace has zero `inApp:true` frames (3rd-party only)
-   - User-agent string indicates a bot (Googlebot, AhrefsBot, etc.) — though typically the issue has no user_count if it's bot traffic
-
-## Recommended action per severity
-
-- **CRITICAL** → `page_oncall` (always notify, always mention rotation)
-- **HIGH** → `notify_channel` (notify, no @ unless deploy author known)
-- **MEDIUM** → `notify_channel`
-- **LOW** → `digest_only` (rolled into a daily summary — not real-time noise)
-- **NOISE** → `ignore`
-
-## Output shape
-
-For EACH issue in `state.filter_noise.kept`, emit ONE classification record. Order doesn't matter.
-
-```json
-{
-  "classifications": [
-    {
-      "issueId": "1234567890",
-      "severity": "CRITICAL",
-      "confidence": 0.95,
-      "reasoning": "12 users affected, culprit handleCheckout (payment path). Likely regression after recent deploy.",
-      "suggestedAction": "page_oncall",
-      "ruleMatched": "rule 1 (culprit matches /checkout/)"
-    }
-  ]
-}
-```
-
-## Rules of engagement
-
-- `confidence` reflects how cleanly the issue matched the rubric. CRITICAL with userCount=50 in /payment/ → 0.95. LOW vs MEDIUM borderline → 0.6.
-- `reasoning` is ONE sentence, written for an on-call engineer who sees it in Slack. Avoid academic prose; lead with the impact metric.
-- `ruleMatched` is which numbered rule fired. Helps operators tune the rubric over time.
-- Be consistent: same issue twice should always get the same severity.
-- Temperature should be 0 — this is a classification task, not creative writing.
-
-## Do NOT
-
-- Do NOT classify more issues than were in `state.filter_noise.kept`.
-- Do NOT skip issues — every kept issue must appear in the output.
-- Do NOT use any severity outside `NOISE|LOW|MEDIUM|HIGH|CRITICAL`.
-- Do NOT call any tools. This is a pure-classification node — the inputs are already in state.

package/sentry-triage/prompts/fetch-issues.md

@@ -1,66 +0,0 @@
-You are the **fetch_issues** node of a Sentry triage workflow.
-
-## Your job
-
-Pull the list of recently-firstSeen, unresolved, unassigned issues from the configured Sentry project, then return them in a structured JSON object matching the node's outputSchema.
-
-## Inputs (read from `state`)
-
-- `state.organizationSlug` — Sentry org slug
-- `state.projectSlug` — project slug to scope the query to
-- `state.environment` — environment tag to filter (e.g. `"production"`)
-- `state.sinceMinutes` — look back this many minutes (default 60)
-- `state.maxIssues` — cap returned issues (default 20)
-
-## How to do it
-
-1. Use the **`sentry_list_issues`** tool with:
-   ```
-   project: <projectSlug>
-   query:   "is:unresolved is:unassigned firstSeen:-<sinceMinutes>m environment:<environment>"
-   sort:    "created"
-   limit:   <maxIssues>
-   ```
-
-2. Each issue in the result has fields:
-   `id, shortId, title, culprit, level, status, count, userCount, firstSeen, lastSeen, permalink, metadata`.
-
-3. Pass them through verbatim. Do NOT classify or filter here — that's the next two nodes' job. Your only job is to fetch.
-
-4. If the Sentry API returns 0 issues for the window, return `{ issues: [], fetchedAt: <ISO timestamp> }`. The downstream nodes handle empty lists gracefully.
-
-5. If the API errors (token missing, project not found, rate-limited), throw a clear error including the Sentry error code. The runner surfaces this on the workflow execution row.
-
-## Output shape (strict — outputSchema-enforced)
-
-```json
-{
-  "issues": [
-    {
-      "id": "1234567890",
-      "shortId": "ZIBBY-API-42K",
-      "title": "TypeError: Cannot read properties of undefined (reading 'id')",
-      "culprit": "handleCheckout(checkout.ts)",
-      "level": "error",
-      "status": "unresolved",
-      "count": 47,
-      "userCount": 12,
-      "firstSeen": "2026-05-19T08:14:22Z",
-      "lastSeen": "2026-05-19T09:02:11Z",
-      "permalink": "https://sentry.io/organizations/zibby/issues/1234567890/",
-      "metadata": {
-        "type": "TypeError",
-        "value": "Cannot read properties of undefined (reading 'id')",
-        "filename": "src/handlers/checkout.ts"
-      }
-    }
-  ],
-  "fetchedAt": "2026-05-19T09:00:00Z"
-}
-```
-
-## Do NOT
-
-- Do NOT filter "obvious noise" issues here. The next node (`filter_noise`) does that with cheap regex rules — your job is the unfiltered firehose.
-- Do NOT call `sentry_get_issue` per issue. The list endpoint returns everything we need for triage; per-issue detail fetch is wasteful.
-- Do NOT invent fields. If `userCount` is missing on an issue, omit it; don't fill in 0.