Historical PagerDuty triage demo

Legacy TypeScript website content. Shared public website, blog, protocol, guide, and language-neutral docs ownership now lives in ~/src/graphrefly under D563. This page is retained here only as migration/reference material while the TS API generator still lives in website/.

These notes describe the historical pre-CSP-9 PagerDuty triage browser demo. Its source was retired from the active tree during CSP-9/B66 closeout because it depended on retired root/pure-ts demo surfaces such as old AI utilities, agentMemory, and utils/demo-shell, so it is not an active clean-slate demo or workspace package.

The historical demo streamed 60 synthetic alerts through a reactive classify-and-route pipeline. Run twice, once in Baseline mode and once in GraphReFly mode, it showed the auto-classify counter climb as old agentMemory recognised decision patterns.

Historical pipeline modes

Baseline — every alert went through the LLM classifier, was either auto-routed at >=80% confidence or queued for the user, and learned nothing between sessions.

GraphReFly — same old classifier, plus the retired agentMemory helper watching decisions. After 2 consistent decisions for the same (service, error-category) pair, the demo wrote a LearnedPattern to patternsState; matching alerts were then routed programmatically.

Graph topology

currentAlert
    │
    ▼
classifyNode (promptNode — LLM)
    │
    ▼
routeEffect ──► binsState (actionable / escalated / resolved / deferred)
    │           userQueueState (awaiting human decision)
    │
decisionLog ──► agentMemory.compact ──► patternsState
                                            │
                              (snapshot read in classifyNode callback)

In the historical design, classifyNode depended only on currentAlert. patternsState was read as a cold snapshot inside the callback so pattern updates never re-triggered the LLM on the current alert.

routeEffect depended only on classifyNode and used a pendingAlerts map keyed by alertId to guard against stale in-flight results when alerts arrived faster than the classifier responded.

Deferred alerts used old source helpers rather than raw setTimeout; this is historical provenance, not current source guidance.

Option 5 pattern matching

Classification is two-tier:

1. LLM — the old promptNode demo utility classified the alert and returned {alertId, disposition, confidence, brief}.
2. Programmatic — if the retired agentMemory helper had a matching LearnedPattern with confidence >=70%, routeEffect routed without invoking the LLM.

The matching idea may still be useful later, but a current implementation should be designed over @graphrefly/ts orchestration/pattern subpaths.

Token accounting

The token bar at the bottom of the main pane shows cumulative inputTokens, outputTokens, LLM calls, and (GraphReFly only) localCacheHits. All counts are updated atomically via batch(). Token estimates are intentionally approximate — a real deployment would read response.usage from the adapter.

LLM adapters

Mode	Notes
Dry-run	Deterministic mock — keyword rules, no latency. Full graph topology and UX visible without an API key.
Chrome Nano	window.LanguageModel (Chrome 138+ with Prompt API flag). On-device Gemini Nano, zero cost.
BYOK	Any OpenAI-compatible API. Enter base URL, key, and model on the setup screen.

Alert generator

60 synthetic alerts are generated at page load from a seeded PRNG (mulberry32). Emission rate ramps: alerts 0–19 at 4 s each, 20–39 at 2 s, 40–59 at 1 s. The 3-minute hard cap stops emission; any queued alerts remain actionable. Click Randomize Alerts on the setup screen to reseed.

Source

The historical browser source was removed from the active workspace rather than migrated through compatibility shims. Re-activating this concept should be a new slice over current @graphrefly/ts orchestration/pattern public subpaths, not a compatibility shim.