Siftly boundary

Type: agentic-system-analyses/types/agentic-system-boundary.md

Boundary and evidence

Siftly is in scope as a memory/knowledge/context-engineering system: it imports Twitter/X bookmarks and likes, retains their content and model-generated metadata in SQLite, and reloads that retained content for browsing, search, visualization, export and a JSON CLI usable by agents. It is not classified as an enclosing general-purpose agent runtime. Its shipped application owns the knowledge-management loop and issues model requests for vision, semantic tagging, categorization, category suggestions and search reranking. Evidence: SRC-1, prisma/schema.prisma, app/api/import/route.ts, app/api/categorize/route.ts, app/api/search/ai/route.ts, lib/category-suggester.ts, cli/siftly.ts.

The intended use is a self-hosted personal bookmark knowledge base. The target's declared purpose and retained-state contract are explicit (SRC-1):

Siftly turns your Twitter/X bookmarks into a searchable, categorized, visual knowledge base — running entirely on your machine. No cloud, no subscriptions, no browser extensions required. Everything stays local except the AI API calls you configure. --- README.md

semanticTags String? // JSON array of AI-generated searchable tags entities String? // JSON: {hashtags[], urls[], mentions[], tools[], tweetType} enrichedAt DateTime? // when AI enrichment last ran — drives incremental processing enrichmentMeta String? // JSON: {sentiment, people[], companies[]} from batch enrichment --- prisma/schema.prisma

The README passage is declared intent, not proof of network isolation. The frozen code also includes X API access, media and link-preview retrieval, and optional Cloudflare tunnel setup. These shipped integrations remain inside the functional account; the remote services' implementations do not (SRC-1, lib/twitter-api.ts, app/api/import/x-oauth/fetch/route.ts, app/api/media/route.ts, app/api/link-preview/route.ts, scripts/setup-tunnel.sh, start.sh).

Functional coverage

The boundary is whole-system, not just the AI helper library. It includes the Next.js application and its browser-side import drivers, HTTP routes, model adapters and CLI bridges, SQLite state and FTS index maintenance, category maintenance, export writers, direct database CLI and shipped setup/maintenance instructions. The agent-facing CLI is a same-repository consumer, not an excluded host application. package.json ships it as the siftly npm script; cli/siftly.ts reads the same database and calls ftsSearch. The import UI also belongs to the target: its upload flow advances to a categorization step that starts or attaches to the server pipeline. This matters because the file-import endpoint itself persists data but does not start categorization (SRC-1, package.json, cli/siftly.ts, app/import/page.tsx, app/api/import/route.ts).

Minimum implementation passages establishing production, maintenance and later consumption of retained state (SRC-1):

const res = await fetch('/api/import', { method: 'POST', body: formData }) --- app/import/page.tsx

const results = await categorizeBatch(batch, client, categoryDescriptions, allSlugs) await writeCategoryResults(results) --- app/api/categorize/route.ts

if (!shouldAbort()) { await rebuildFts().catch((err) => console.error('FTS rebuild error:', err)) } --- app/api/categorize/route.ts

const ftsIds = await ftsSearch(keywords) --- cli/siftly.ts

Search also reloads semantic tags, entities, media tags and category assignments before composing the reranking prompt. Obsidian export reloads bookmarks with media and categories, writes notes, and writes category and author indexes. These are material consumption paths, not evidence that an external agent or Obsidian installation successfully used the output (SRC-1, app/api/search/ai/route.ts, lib/obsidian-exporter.ts).

The following table covers every top-level tracked entry from the frozen Git tree. All rows refer to SRC-1. “Included” describes functional scope, not a claim that every line was audited. Full reads and bounded excerpts established the central loop; bounded source searches established additional route callers and integration paths. Entries marked inventory-only were available but not substantively inspected; they are not access gaps.

Frozen evidence and limits

The only frozen source is SRC-1: https://github.com/ViperrCrypto/Siftly at b25daa45b858f4be096b5c368671c34db4407d8e. The supplied checkout's HEAD matched that revision; source inspection was read-only. The applicability cutoff is 2026-10-18, the supplied run date. It is not a claim that the commit was authored that day or that later revisions behave the same way. No supplementary repository, web capture, execution trace or experiment is registered.

Overall evidence is code-grounded. Executable source supports implementation findings; README.md and CLAUDE.md support attributed doctrine/design. Tests are inspectable code with mocked SDK responses, not observed runs. In particular (SRC-1):

• These test end-to-end scenarios like categorization and search
• using the MiniMaxAIClient with mocked SDK responses. --- __tests__/minimax-integration.test.ts

External dependencies are the Next.js/React and Prisma/SQLite stacks, provider SDKs and remote model endpoints, locally installed Claude/Codex CLIs and credentials, X/browser sessions and APIs, media/link hosts, optional Cloudflare infrastructure, and an optional destination Obsidian vault. Siftly's adapter code and calls are included; dependency internals, actual accounts, user databases, external host applications and remote service implementations are not frozen. The CLI bridges invoke external executables; they do not give Siftly ownership of those runtimes' internal loops (SRC-1, package.json, lib/db.ts, lib/ai-client.ts, lib/claude-cli.ts, lib/codex-cli.ts, lib/openai-auth.ts, lib/twitter-api.ts, lib/obsidian-exporter.ts, scripts/setup-tunnel.sh).

There was no access failure for a material shipped source path. Available but uninspected support files are identified above, not described as unavailable. The absence of frozen provider responses, deployed databases, browser sessions and runtime traces prevents conclusions about successful authentication, actual image understanding, search precision/recall, operational throughput, restart behavior in a real deployment or downstream agent task improvement. Excluding remote dependency internals prevents attributing their scheduling, tool execution, durability or safety properties to Siftly. None of these limits removes the shipped import, maintenance or consumption paths from this whole-system boundary.

Source register