convez.ai
Agente
Cheeryoh

drift-sentinel

Cheeryoh/drift-sentinel

Ver no GitHub
InicialPluginScore · de 0 a 100

HITL multi-agent orchestration that detects product↔course drift, prioritizes the fix backlog, and drafts patch content. Claude Code Max plan.

Instalação

$npx convez-ai Cheeryoh/drift-sentinel

1 comando no terminal: detecta o formato do repositório (marketplace ou plugin avulso) e instala via Claude Code. Requer Node 18+ e o Claude Code instalado.

Métricas

Estrelas
0
Forks
0
Acompanhando
0
Saúde
28%
Último commit
há 1 semana
Versão
—
Issues abertas
0
Contribuidores
1

README

A tradução para o português deste README ainda está na fila.

Drift Sentinel

A human-in-the-loop, multi-agent orchestration that catches where shipped product has outrun the course catalog — then prioritizes the fix backlog and drafts the patch content.

This is the answer to the hardest structural problem in customer education at a fast-moving product company: product velocity outruns content velocity. Features ship weekly; courses get updated quarterly. The gap between them is invisible until a customer hits outdated guidance. Drift Sentinel makes that gap visible, ranked, and partly pre-written — with a human approving every judgment.

Part of a three-product portfolio. See also Launch Content Studio (launch asset packages) and Signal Loop (customer signals → education priorities).


Why it exists

You can't manually diff every release against every course. So drift accumulates silently: a workshop keeps teaching a pattern the product just made obsolete; a breaking deprecation ships with no migration content and a hard deadline. Drift Sentinel runs that diff automatically, scores customer risk, and proposes the smallest effective fix for each gap — but stops at three gates so a person owns the severity calls, the priorities, and what actually gets drafted.

The pipeline

 ChangelogParser ─┐
 CatalogMapper  ──┴─▸ DriftDetector ─────────▸ [Gate 1 · Drift report]
 RemediationPlanner ──────────────────────────▸ [Gate 2 · Prioritized backlog]
 PatchWriter → PatchQA ────────────────────────▸ [Gate 3 · Drafted content]
                                                       │
                                                       ▼
                                          ranked gaps + ready-to-edit drafts
AgentRole
ChangelogParserAgentRelease notes → discrete product changes with honest customer-impact ratings
CatalogMapperAgentCourse catalog → structured coverage (topics, last-updated, owner)
DriftDetectorAgentCross-references every change against coverage → none/partial/stale/covered + severity + concrete customer risk
RemediationPlannerAgentDrift report → prioritized backlog (P0–P3, S/M/L effort, smallest effective action, suggested owner)
PatchWriterAgentDrafts ready-to-edit content for the P0/P1 gaps only
PatchQAAgentTechnical-accuracy review of each draft against the shipped change

The severity model is the interesting part: a change is "critical" not just because it's high-impact, but because it's high-impact and the catalog actively teaches something now wrong (stale) or covers it not at all (none). That's the difference between "new feature, no course yet" and "course is now lying to customers."

Run it

Requires Node 20+. No API key needed — agent calls run through your local claude CLI (Max plan) by default; zero API billing.

npm install
npm run dev          # http://localhost:3000

Paste a changelog (or click an example). Optionally append your own catalog after a ---CATALOG--- delimiter; omit it to use the bundled sample catalog.

LLM backends

BackendWhenCost
cli (default)local claude CLI presentnone (Max plan)
apiANTHROPIC_API_KEY setAPI usage
fixtureUSE_FIXTURES=truenone — canned outputs for offline/demo/test

Test

npm test            # full pipeline, fixture mode, auto-approve
npm run test:reject # rejects the backlog gate once to exercise the rewind path

Both assert all six artifacts are produced and the run completes.

Architecture

Shares the kernel (src/kernel/) verbatim with the other two portfolio products: promise-based HITL gates, dual-backend ClaudeClient, filesystem persistence under .data/, and SSE live progress. Only agents.ts, schemas.ts, Conductor.ts, and public/config.js are product-specific. See HANDOFF.md for the developer guide.

src/
  kernel/        shared orchestration engine (identical across products)
  agents.ts      6 product agents
  schemas.ts     Zod artifact schemas
  Conductor.ts   the drift pipeline + default sample catalog
  index.ts       wires the conductor into the kernel server
public/          dashboard
fixtures/        canned agent outputs for offline tests
scripts/         testPipeline.ts

MIT licensed.

Ver o original no GitHub

Bom para

Relacionados