The original site. Reads subagentjobs-dwh (Kimball warehouse: fact_job_posting, dim_board, dim_team, fact_match). Serves an A2A agent at POST /a2a.

Round-of-32 radial bracket, D1-backed, autonomous A2A score ingestion every 10 minutes via cron. Originated the family's most advanced techniques — :has(), CSS anchor positioning, desynchronized canvas — none yet ported elsewhere.

RAG over Cloudflare's own docs — D1 FTS5 across 859 pages, generation via Workers AI (@cf/zai-org/glm-5.2). Zero external API keys.

Validated, SELECT-only NL-to-SQL over three D1 databases (subagentjobs-dwh, subagentworkers-dwh, subagentdb), same Workers AI model, zero external API keys.

Read-only Kimball-style dashboard over subagentjobs-dwh — star-schema diagram, pipeline-run history, per-source freshness. No writes.

Static mirror of Anthropic's Cowork (30 files) and Claude Tag (52 files) docs, plus 5 original Cloudflare Managed Agents pages. Serves /integrity (sha256 drift-detection) and /design-system (the living token/component showcase for this whole family).

20 original platform-docs pages documenting this operator's own multi-agent coworkers platform — Design/Engineering/Data/Finance/Legal/HR/Product/Sales/Operations coworkers, the e2m envelope, DurableTask queues, approval gates — plus the /indexdb, /redis, and /postgres infra pages added 2026-07-10.

D1-backed durable Contract primitive — real GET/POST/PATCH /api/contracts API, grounded in Claude Tag's own "the thread is durable, the sandbox is not" and "definition of done" vocabulary. Dogfooded by cowork, coworkers, devices, and this site.

D1-backed Task primitive generalized directly from the real TaskCreate/TaskUpdate/TaskList tool schema powering this Cowork session's own Progress panel. Seeded with this exact session's 44 real tasks. contract_id FK joins to subagentcontracts-web.

Every iPhone 16 Pro / Chrome 150 canary progressive-enhancement technique the family ships — Display-P3 widening, safe-area insets, View Transitions, relative-color syntax — each with real code and a grep-verified per-site adoption matrix, honest about the gaps.

D1-backed catalog of real identity/access boundaries, grounded in Claude Tag's channel-service-account-vs-DM-personal-account duality. Dogfoods this repo's own macos_desktop_cowork vs cloud_docker_mcp naming ontology rather than inventing generic content.

D1-backed durable citation store, generalized directly from Anthropic's real Citations API schema (char_location / page_location / content_block_location). Seeded with real quotes already used to ground sibling sites this session.

Durable, provenance-tagged history of real Cloudflare cache/CDN analytics for this account (manual_dashboard_paste vs graphql_api). Seeded with a real dashboard pull, not synthetic numbers. Live GraphQL ingestion documented as a next step, not yet wired.

D1-backed catalog of gradeable scoring rubrics, generalized from the real Managed Agents define_outcome rubric schema. Seeded with the Worker Quality Rubric that subagentevaluations.com scores this session's shipped sites against.

D1-backed catalog of session outcomes, generalized from the real Managed Agents user.define_outcome event and result enum. rubric_id cross-references subagentrubrics.com by id.

D1-backed catalog of grader evaluations, generalized from the real span.outcome_evaluation_* event schema. Seeded with 9 real, evidence-based rows scoring this session's shipped sites.

Read-only aggregator: catalogs every real GET/POST/PATCH route across all 37 D1-backed primitives, and live-pings each one on /health. Holds no state of its own.

Docs and distribution for subagent-cli, a real zero-dependency Node.js script that queries every primitive's JSON API from the terminal.

D1-backed brand-guideline catalog adapting the operator's real brand-voice:discover-brand skill to this repository's own shipped pages instead of enterprise connectors.

D1-backed catalog of this repo's own typed Claude Code sub-agent system (AgentConfig in crates/schema/src/agent.rs). Seeded with the real 10 generated agents, including the 2026-07-01 upgrade of engineering-coworker to claude-fable-5 — the day Anthropic restored global Fable 5 access.

D1-backed catalog of real prompting techniques and guardrails, grounded in this repo's docs mirror plus live doc fetches. Dogfoods this session's real Sonnet-main/Fable-advisor pairing — Fable 5 reviewed this site's own draft before deploy.

D1-backed catalog of every real SKILL.md file this repo's plugins ship, cataloging the actual skills (not a hypothetical taxonomy) already running across the family.

D1-backed catalog generalizing Anthropic's real Memory Stores and Dreams APIs — versioned, path-addressed memories capped at 100kB/~25k tokens per memory.md's own limits, plus a deterministic dream that consolidates a store into a fresh output store without ever touching the input.

D1-backed catalog of the real Claude models this repo's own code depends on (crates/schema/src/agent.rs's AgentModel enum, this session's own Sonnet 5, and claude-mythos-5 correctly documented as unused) — scoped and cited, not a re-derivation of every model string mentioned incidentally in the mirrored docs.

No-D1 read-side aggregator (like this site) that server-side-fetches subagentapi.com's live catalog and renders a deterministic SVG node-link graph of every family site's routes. Known gap, sharpened 2026-07-02: the domain is not just zoneless — it is UNREGISTERED (Verisign .com registry returns no match), so registering it is an operator purchase decision, not an API call. Reachable today only at its workers.dev URL, which this directory links directly rather than pointing at a dead apex.

D1-backed catalog of the real serde-deriving structs/enums across this repo's own crates/ — grepped live from the repo, not a hypothetical style guide. Cross-references the Pydantic and Zod mirrors of crates/schema/src/agent.rs.

D1-backed catalog grounded in scripts/agents/schema.py, a real Pydantic v2 mirror of agent.rs added 2026-07-01 — this repo had no prior in-repo Pydantic code, honestly scoped rather than padded with vendored/docs mentions.

D1-backed catalog of scripts/agents/schema.ts's real Zod schemas. Documents a real drift bug found and fixed the same session: its AgentModel enum was missing claude-fable-5 relative to agent.rs.

D1-backed catalog grounded in permission-policies.md — always_allow/always_ask defaults per toolset, per-tool config overrides, and the fact that custom tools aren't governed by policies at all.

D1-backed catalog grounded in vaults.md — mcp_oauth/static_bearer/environment_variable credential types, the 20-per-vault cap, and egress-only secret substitution.

D1-backed catalog of the real 4-state session lifecycle (idle/running/rescheduling/terminated) from session-operations.md, complementing subagentmemories.com's session_attachments.

D1-backed catalog of the real 8-tool agent_toolset_20260401 from tools.md — a deliberately different, smaller scope than subagentcode.com's 21-tool Claude Code AgentTool enum.

D1-backed catalog of the 11 real session/vault webhook event types from webhooks.md.

D1-backed catalog of the job-function/responsibility axis — distinct from Agent Definition (subagentcode.com): roles describe what a coworker is responsible for, not which model/config runs it.

D1-backed catalog of this repo's real session-handoff JSON docs (sessions/handoff-*.json), seeded verbatim rather than summarized from memory.

D1-backed catalog of every real Cloudflare zone this repo's own cloudflare.toml [dns] table manages — domain, zone ID, bound Worker, route status. Complements this site's taglines and subagentdatamodels.com's API/CLI edges with the DNS layer itself.

D1-backed catalog of the real GraphQL projection contract (parse/stringify/injectCitationType/extractCitations/injectCitations/roundtrip) and the canonical Citation SDL type — a strict subset of CSL-JSON v1.0.2 — shipped by this repo's style-graphql skill.

D1-backed catalog of the real citation conventions the style-html skill documents: semantic <cite data-csl-id> elements, schema.org JSON-LD citation blocks, and the rehype/unified ecosystem for type-safe HTML5 generation/parsing.

D1-backed catalog of the real API surface documented in two skills this repo carries for Apple-platform markdown handling — style-swift-markdown (parsing/AST) and style-swift-markdown-ui (SwiftUI rendering) — kept as distinct pipeline layers via a category column.

D1-backed catalog of real facts about the engineering-coworker role, grounded in two distinct sources: the abstract engineering-coworker Cowork skill, and the concrete crates/engineering-coworker Rust MCP server in this repo that executes cargo/wrangler/git/D1 ops on the developer's Mac.

D1-backed catalog of the legal-coworker's 9 real plugin skills, identity facts, and hard invariants (never provides legal advice, never sends external correspondence directly), grounded in the legal-coworker SKILL.md from the knowledge-engineering sister repo.

D1-backed catalog grounded in this repo's installed finance-coworker skill — cost tracking, vendor spend, budget approval gates, and the 8 real finance plugin skills it routes to (journal-entry, reconciliation, variance-analysis, financial-statements, sox-testing, audit-support, close-management, journal-entry-prep).

D1-backed catalog of the layer above the coworker tier (subagentroles.com) — a manager owns OUTCOMES, never executes, dispatches down to exactly one coworker, and enforces a durability gate before accepting completion.

Independent, durable, D1-backed catalog of facts about Anthropic's real, public, open-source customer-support knowledge-work plugin — does not install it, embed its code, or claim to be a copy of it; every row cites its exact source URL.

Two deliberately separate D1 tables: this repo's own 20 real cwc-* Cowork plugins, and Anthropic's real public cowork-plugin-management plugin (cowork-plugin-customizer, create-cowork-plugin) — cataloged for reference/grounding only, never conflated with the 20 cwc-* plugins.

D1-backed catalog of the real MCP + A2A SDK dependencies this repo actually vendors/uses, verified directly from Cargo.toml/package.json and the vendors/ clone tree — not invented version numbers.

D1-backed catalog of the 3 real rmcp-based MCP servers this repo ships — crates/mcp-server, crates/a2a-bridge, crates/engineering-coworker — every binary name, transport, tool name, and dependency list copied verbatim from that crate's own Cargo.toml/src.

Narrow-scope catalog of style.yaml, rubric.md, and exactly the 7 rubric criteria the terse YAML style is graded against, plus its 4 structural elements (inputs/outputs/steps/eval). Deliberately does not duplicate subagentprotocols.com's broader protocol overview.

D1-backed comparison catalog of exactly 3 real inter-agent protocols this ecosystem uses — A2A, MCP, and e2m — covering purpose, transport, and real usage in this repo, not wire syntax (that narrower job belongs to subagentspec.com).

Python Worker (Pyodide) serving the R2-mirrored docs content, accepting authenticated pushes from the local scrapyd pipeline, and running a daily human-out-of-the-loop incremental crawl on its own Cron Trigger. Headless — no HTML shell of its own, so the swatch here is the family's neutral gray, not a site accent.

Live progress view of the operator's xAI Data Engineer take-home solution — pipeline architecture, sample-data findings, and the verified near-dedup performance fix. Every diagram node is a real pipeline.py function name; sample-data results only, the full 40GB corpus run has not happened yet.

D1-backed catalog of real, citable facts about Claude.ai's Artifacts-based Apps surface, grounded only in live-fetched Anthropic pages — every row's source_url was actually fetched, nothing invented.

D1-backed catalog of real Bedrock model IDs, cross-region inference profile prefixes, Guardrails, provisioned throughput, and auth differences vs. the direct Anthropic API — every row cites a real source_url from live 2026-07-01/02 fetches.

D1-backed catalog of the brief across disciplines — legal IRAC case briefs, marketing creative briefs, and Anthropic's own briefing-Claude guidance — which converge on fixing objective, audience, and done-criteria before execution starts. Honestly scoped: 3 externally-cited rows plus 1 in-repo artifact.

D1-backed catalog of web-search-verified facts about Anthropic's own governance/values framework (Claude's Constitution), its Economic Futures program, and Claude's enterprise positioning — where research was thin, the gap is stated honestly rather than papered over.

D1-backed catalog of real, sourced facts about agentic-coding verification gates, sandboxing, and runbook practices — grounded in this repo's own installed agentic-dev-practices skill plus a fresh web-search pass; every row cites a source_url that was actually fetched.

D1-backed catalog of MCP concepts, common pre-built connectors (Google Drive, Slack, GitHub, Notion, and peers), and this repo's own dogfooded naming ontology for distinguishing device/client surfaces when connecting MCP servers.

D1-backed catalog of real Claude context-management concepts — context windows, prompt caching (grounded in this repo's mirrored prompt-caching.md), and context editing/compaction. Exact token-limit figures are deliberately phrased generally unless independently verified.

D1-backed catalog of real Anthropic API billing concepts. No specific dollar figures are asserted anywhere — cost rows are phrased generally rather than faked to a false level of precision.

D1-backed catalog of prompt-engineering best practices specific to writing Claude tool/function descriptions and schemas — a focused subset, not a duplicate, of subagentprompts.com's general techniques.

Meta-catalog of this repo's own docs/docs/<domain> mirrors — real directory lists, real file counts per mirror, honest last-synced notes. An index of which docs live where and how fresh, not a docs viewer.

Catalogs crates/agent-gen, the tool that generates this repo's agent definitions from crates/schema/src/agent.rs. No product literally named Foundry exists elsewhere in this repo — verified by full-repo grep before building.

Honestly scoped: this repo had zero Java code before 2026-07-01. The catalog is grounded in the one real artifact added the same day — scripts/agents/schema.java, a Java 17 mirror of crates/schema/src/agent.rs (records + enums), written directly from the Rust source.

A curated, synthesized changelog — one entry per dated marker in CLAUDE.md plus entries drawn from real session handoff files, sorted newest first. Deliberately not a raw dump of handoff JSON.

D1-backed catalog of real Anthropic Console org-admin concepts — organizations, workspaces, API-key scoping, member roles — deliberately distinct from subagentpermissions.com (agent toolset policies) and subagentidentities.com (agent identity/attribution).

Infrastructure-layer catalog of the real Cloudflare account backing this family — 103 Workers, 48 D1 databases, 10 KV namespaces, 10 R2 buckets, live-queried 2026-07-02; the fleet has since grown — see the live count above — plus a per-site storage-primitive table.

D1-backed catalog of the real Anthropic product lineup — Claude.ai, Claude Code, Cowork, Platform/API, and the for-Chrome/Excel/Slack/M365 surfaces — each row distinct from a model (subagentmodels.com): a surface is what you touch, a model is what runs behind it.

Static, browsable catalog of real D1 table shapes — every table/column queried live via the Cloudflare D1 MCP and seeded verbatim, no invented columns. Reference documentation, not a query tool (that is subagentdb.com's job).

D1-backed catalog of real OpenTelemetry concepts the Claude Agent SDK/CLI actually emits — grounded in the mirrored observability.md plus this repo's own Rust tracing usage and the [observability] block every Worker in this family ships.

D1-backed catalog of real facts about running Claude via Google Cloud's Vertex AI Model Garden — same models as the direct API and Bedrock, GCP-native billing/IAM auth, publisher-model-ID addressing. Cross-references subagentbedrock.com rather than duplicating it.

Live view over subagentsubdomains.com's real dns_zones registry — every domain/worker registration in the family, fetched server-side at request time. Holds no state of its own.

The family's real crawl surface: the sanctioned cross-host sitemap.xml on subagentsubdomains.com plus the per-site robots.txt/llms.txt conventions, fetched and rendered live.

The agent*.com-side lens on the same live dns_zones registry subagentsubdomains.com serves — one registry, two brand views.

Dependency-tree view over subagenttasks.com's real task rows — subtasks are task_dependencies, rendered as trees from the live API.

A todo-list lens over subagenttasks.com's real task rows — the same TaskCreate/TaskUpdate schema this repo's own Cowork sessions run on, grouped by status.

The fleet grouped as portfolios — by kind, by accent, by launch wave — computed at request time from the live dns_zones registry. The fleet is the portfolio.

Quickstart docs for subagentoutcomes.com's real outcomes API — curl and Node examples in the subagentcli.com style, with a live response sample fetched at request time. The SDK is the API, documented honestly.

Catalog of diagrams the family actually ships — build-graph.json dependency edges (baked at build time, no invention), the /architecture node-link map on subagentdevices.com, and the data-shape SVG work — each entry citing its source.

What agent 'blogging' actually looks like in this fleet: session-authored launch notes, handoffs, and announcements — grounded in the repo's real session history, not an invented blog.

The family's real change history: launch dates and notable changes per site distilled from the repo's actual git/session records, linking each site's /api/changes feed where one exists.

Real Co-Authored-By records from this repo's git history — which commits carry Claude co-author trailers, by model, with counts generated from git itself. Nothing invented.

The real data models this repo runs on: D1 migration schemas, crates/schema serde types, and the three language mirrors (Rust/serde, Python/Pydantic, TypeScript/Zod) already cataloged by the subagentrust/python/typescript trio.

The real warehouses behind the family: subagentjobs-dwh (Kimball star schema), the local WBR Postgres star, and the ETL pipeline feeding them — with pointers to subagentetl.com's live dashboard.

How this fleet actually logs: deploy-runner step markers (STEP=n EXIT=rc, RUN_EXIT), the kit drift-gate output, [observability] on every worker, and crawl_log — real formats quoted from the real tools.

The family's real ML surfaces: Workers AI (@cf/zai-org/glm-5.2) behind subagentworkers.com and subagentdb.com, and the corpus-cleaning pipeline in cwc-data-engineering. Honestly scoped: no model training happens here.

Publishing here means deploying: the real pipeline every site ships through (typecheck, wrangler deploy, curl-verify, build-lock record) and the sites shipped through it — including this expansion's own wave record.

The fleet's real readers: D1 FTS5 RAG over 859 Cloudflare docs pages (subagentworkers.com), the 4,391-file local docs mirror with sha256 drift manifests, and the crawlers that built both.

One card per real worker in this repo — name, domain, accent, pattern, data backing, routes — system-card format at worker scale, from the same metadata the directory sites serve.

The real 108-zone brand inventory: names, accents, and the naming ontology across the fleet. Informational only — not legal advice; no trademark filings implied.

opencoworkers on GitHub — home of subagentjobs, this fleet's monorepo: ~97 workers, 20 cwc-* coworker plugins, the shared kit, and the build system. The org homepage the repo never had.

The repo's real typed sub-agent system in the open: crates/schema agent.rs -> agent-gen -> .claude/agents/*.yaml (10 generated agents), cataloged live at subagentcode.com.

Rebuild of the old fleet index as the MCP map: which family sites expose /mcp (Streamable HTTP + server cards), what tools they serve, and this repo's real MCP servers (mcp-server, engineering-coworker, subagentswift-mcp).

The fleet's real system-design decisions: the shared-kit extraction, the build-graph/drift-gate design, and per-site PRD/ARCHITECTURE records — every entry cited to its in-repo source.

The live portfolio index you're reading. Holds no state of its own — every count on this page is a server-side fetch to subagentcontracts.com and subagenttasks.com, grouped by source_site. Lists itself, last, on purpose.