Dobby System

Dobby System

Purpose

Use this skill to orient before changing the Dobby system across repos. Keep the mental model centralized: each repo owns one layer, and cross-repo work should change the owning layer directly instead of scattering duplicate rules.

This skill is the canonical cross-repo orientation layer for Dobby. Future agents touching any Dobby-system repo should check it first, and should consider whether their change updates the shared mental model here before finishing.

Repo Ownership Map

Core Model

• Folder identity matters. An agent opened in adi is operating as Adi's Dobby workspace; an agent opened in angie is operating as Angie's Dobby workspace.
• Workspace repos are thin. They hold identity/data and call the shared engine through workspace-local ./bin/dobby.
• The shared engine must remain workspace-agnostic. Runtime person data comes from DOBBY_WORKSPACE or workspace marker discovery, not from hardcoded Adi/Angie paths.
• Document access is a domain bridge, not engine-owned data. dobby-engine exposes dobby documents status/search plus the explicit copy-only dobby documents ingest front door; the separate documents repo owns catalog, import, and search-index implementation. Each personal workspace must opt into a document root such as DOBBY_DOCUMENTS_DATA_ROOT; the shared engine should not guess a person's corpus from a machine-wide default.
• dobby-gateway and client repos should treat Dobby as a workspace-backed service boundary. They should call the workspace/engine contract, not read private memory files ad hoc.
• Health storage is engine-owned but local and person-bound. The combined SQLite DB lives on the internal SSD at ~/Library/Application Support/Dobby/health/health.sqlite; dobby-engine owns schema, migrations, sync/import, and queries. person_id separates Adi and Angie rows. dobby-gateway, dobby-ios, and other products must consume health through workspace-bound Dobby commands or gateway responses, never by reading health JSON files or SQLite directly. JSON health files in workspaces are import, audit, or backup material only.
• Shelf storage is engine-owned but local and person-bound. The SQLite DB lives on the internal SSD at ~/Library/Application Support/Dobby/shelf/shelf.sqlite; dobby-engine owns schema, migrations, query/projection, and mutation logic. person_id separates Adi and Angie rows. Products and dashboards must consume Shelf through workspace-bound dobby shelf commands or gateway responses, never by reading workspace JSON or SQLite directly. This database is per-machine and is not replicated by Git sync: a hosted dashboard reads its serving host's Shelf, so dashboard-visible writes must target that host or an explicit remote-authoritative API. Workspace state/shelf.json is legacy import, audit, or backup material only.
• agents distributes capabilities and lifecycle hooks. Put shared Dobby orientation here as this skill; do not turn agents into the Dobby architecture doc home.
• scripts may schedule Dobby work on a specific machine, but only as a thin machine wrapper around the owning repo entrypoint. Live launchd plists under ~/Library/LaunchAgents/ are machine-local runtime state; tracked installers/runners live in scripts.

Workspace Anatomy

Use this skill as the entry point, then read the current workspace's docs/body-map.md for exact routing before changing workspace shape or deciding where person data belongs. Do not duplicate body-map detail here.

Adi and Angie workspaces share the same organ layout:

bin/ dobby/ docs/ journal/ memory/ projects/ scripts/ state/ tmp/

Inside a person workspace:

• bin/dobby is the only Dobby CLI entry point; it pins the workspace and calls the shared engine.
• dobby/constitution.md is Dobby behavior and boundaries. It is declared once as model_instructions_file in ~/GitHub/agents/codex/config/repo-bootstrap.json; the optional model_instructions_clients selector controls which clients receive it. Codex loads it through .codex/config.toml; Claude and Copilot bridges remain available for workspaces that enable them. Adi is intentionally Codex-only, while Angie retains the default cross-client propagation. The session-start hook does not emit the constitution.
• Codex SessionStart is bounded and source-aware: startup/clear receive the wake card, compact receives a smaller reorientation card, and resume emits nothing. Exact contents and budgets belong in the engine lifecycle docs.
• memory/profile.md is durable person context; memory/now.md is current orientation.
• memory/areas/<area>/canon.md is the one durable area layer.
• memory/areas/<area>/area.json indexes area metadata, data dirs, and assets.
• memory/sessions/ and memory/dreams/ are continuity records.
• journal/ holds raw reflections/check-ins; state/ holds legacy/import or backup machine-readable material; projects/ holds active work trackers; tmp/ is disposable scratch.

Before writing Dobby memory, check the shared write contract at ~/GitHub/dobby-engine/docs/agent-write-recipes.md. Use ./bin/dobby for CLI-owned writes such as journal, Shelf, sessions, dreams, calendar, and mail. Direct file writes are for documented shapes such as area canon, area.json, and source-material JSON.

For social archives, corpora, imported documents, and other source material:

• Keep raw exports as inputs/backups or repo-local temp/source assets, not as canon.
• Store normalized supporting captures under the relevant person's workspace, usually memory/areas/<area>/<dataDir>/... as source-material JSON when the area declares that data dir.
• Put durable conclusions in canon.md; put personal open loops in Shelf; put raw reflection in journal/.
• Keep person-private data out of dobby-engine, agents, dobby-gateway, dobby-ios, and public/content repos such as blog-personal. Those repos may be sources, not homes for private corpus data.
• Area log.jsonl files are retired; do not create them.

Choosing Where a Change Belongs

1. If it changes Dobby commands, hooks, validation, memory processing, dashboard code, or shared behavior for both people, change dobby-engine.
2. If it changes document inventory, copy-only ingest/import, extraction, canonical cleanup, document search-index construction, or the documents CLI contract consumed by Dobby, change documents.
3. If it changes Adi's or Angie's constitution, memory, journal, legacy Shelf import/audit material, person-specific prompt behavior, or workspace-specific document-root opt-in, change only that person's workspace repo.
4. If it changes the mobile/web gateway, assistant runtime selection, gateway HTTP contracts, Cloudflare-facing local service behavior, or product routing in front of Dobby, change dobby-gateway.
5. If it changes the iOS app, SwiftUI surfaces, iOS build/deploy/TestFlight scripts, or iOS-only docs, change dobby-ios.
6. If it changes which skills/tools/hooks/configs agents receive, change agents.
7. If it changes machine bootstrap, launchd scheduling, scheduler profiles, or a thin machine-local wrapper around a Dobby command, change scripts; keep the real Dobby behavior in dobby-engine or the person workspace.
8. If more than one repo is touched, keep each repo's change scoped to its ownership layer and validate each repo separately.

Documentation Boundary

Because Dobby is intentionally multi-repo and agent-native, documentation has two layers:

• This skill owns cross-repo orientation: repo ownership, privacy boundaries, shared engine/workspace model, routing rules, and rules that future agents must know before deciding where a change belongs.
• The owning repo owns implementation detail: command contracts, schemas, env vars, tests, local workflows, and feature-specific architecture docs. Put those in that repo's AGENTS.md, docs/architecture/, or docs/references/ according to the repo's guidance.

Do not duplicate repo-specific contracts into this skill. Do not bury cross-repo Dobby rules only inside one repo. When a Dobby change affects more than one repo, or changes how agents should reason about ownership/routing, update this skill in the same change or explicitly record why it did not need an update.

Operational Rules

• Prefer workspace-local ./bin/dobby for Dobby operations from adi or angie; do not call engine internals from a workspace unless debugging the engine boundary.
• Do not copy personal data into dobby-engine, dobby-gateway, dobby-ios, or agents.
• Do not merge Adi and Angie workspaces into one repo or one branch history. Separate repos are the privacy and writer-boundary.
• Do not recreate duplicate engine code in workspaces. Shared code belongs in dobby-engine.
• Visible workspace artifacts are a built-in system capability, not a skill. Each artifact lives at memory/areas/<area>/artifacts/<slug>/ (artifact.json + index.html dashboard); the workspace body map routes them. The dashboard (index.html) and the book (book.json) are independent expressions — never render one from the other. To read an artifact on the Kindle Scribe, the agent authors a book.json on demand (fresh prose, its own chapters) and runs bin/dobby artifact export --dir <dir> [--kindle] for a native book PDF. Person-private artifact content belongs in the person workspace; the minimal book toolkit, rendering, and the export command belong in dobby-engine (contract: docs/artifacts.md).
• Do not add README surfaces for this system. Use AGENTS.md for repo-local routing, this skill for cross-repo orientation, and repo docs for durable contracts.
• When a boundary changes, update the owning repo's AGENTS.md or docs in the same change, and check whether this skill's cross-repo orientation also changed. Update this skill for ownership, routing, privacy, or operating-model changes; leave feature-level contracts in the owning repo.

Validation Expectations

Run the owner repo's fast check before finishing:

cd ~/GitHub/adi && scripts/check-fast.sh
cd ~/GitHub/angie && scripts/check-fast.sh
cd ~/GitHub/dobby-engine && scripts/check-fast.sh
cd ~/GitHub/dobby-gateway && scripts/check-fast.sh
cd ~/GitHub/dobby-ios && scripts/check-ios-fast.sh
cd ~/GitHub/agents && scripts/check-fast.sh
cd ~/GitHub/scripts && ops/check-fast.sh

For cross-repo work, run checks for every touched repo. If changing shared engine behavior, also smoke at least one real workspace path when relevant.