agent-system

name

agent-system

description

Grida AI agent system work: `@grida/agent`, AgentHost, AgentSidecar server, loopback HTTP transport, sessions, workspaces/files, secrets/BYOK, providers, runtime/tool execution, skills discovery, prompts, tiers, and sandbox policy. Use for `packages/grida-ai-agent/**`, desktop sidecar protocol changes, agent chat transport, and bugs in agent state or streams. For pure Electron window, preload, menu, deep-link, or CDP work, use `desktop`.

Grida Agent System

This skill is for @grida/agent, the host-agnostic agent system behind Grida Desktop and future hosts. It owns AgentHost lifecycle, the loopback HTTP perimeter, sessions, workspaces/files, secrets, providers, tool execution, prompt composition, skills, and the AI-SDK stream contract.

Use desktop for Electron shell work: BrowserWindow, menus, preload, native dialogs, file associations, deep links, and CDP verification.

Adjacent: sdk-design and sdk-seam for exported contracts and host seams, security when a change touches a GRIDA-SEC boundary.

When to use this skill

• Editing packages/grida-ai-agent/**.
• Changing AgentHost lifecycle, server capabilities, HTTP routes, transport, Basic Auth, Referer policy, sessions, workspaces/files, recents, secrets, BYOK providers, model tiers, sandbox policy, shell policy, runtime, tools, prompts, skills, or message streams.
• Touching the desktop sidecar protocol because the AgentHost contract changed.
• Debugging empty file lists, missing sessions, broken stream parts, provider resolution, tool-call behavior, or agent state recovery.
• Editing renderer agent chat transport when the bug is in stream/session semantics rather than Electron bridge presence.

Skip this skill for BrowserWindow, menu, native dialog, file association, deep-link, packaging, or preload-only bugs. Use desktop for those.

---

Shape

Host adapter (Desktop today, other hosts later)
  - supplies auth material, workspace roots, entitlement hooks, sandbox wrapper
  - starts AgentHost or AgentSidecar
        |
        v
AgentHost (@grida/agent/server)
  - long-lived services
  - sessions, workspaces, files, secrets, providers, shell, run loop
        |
        v
HTTP perimeter (@grida/agent/server + @grida/agent/transport)
  - Hono on 127.0.0.1:<random> for Desktop sidecar
  - per-request Basic Auth
  - Referer guard as defense in depth
        |
        v
Runtime-agnostic agent core (@grida/agent)
  - protocol DTOs, prompts, tiers, skills, toolset, AI-SDK UI-message stream

The host owns process supervision and native capability exposure. The package owns agent-system semantics.

---

Package Boundaries

@grida/agent has intentional entrypoints:

• @grida/agent: neutral contracts, provider metadata, handshake/run/stream DTOs, createAgent, prompt composition, toolset, tiers, and session row types.
• @grida/agent/server: Node-only AgentHost and server capability contracts.
• @grida/agent/transport: Basic Auth signing, fetch helpers, and AgentTransport.Client.
• @grida/agent/fs: storage-agnostic virtual filesystem and AI-SDK file tools.
• @grida/agent/fs/backends/opfs: browser OPFS backend.
• @grida/agent/todos: plan store and todo_write.
• @grida/agent/tiers: model tier constants.

Keep browser-safe imports neutral. Node-only code must stay behind server, sandbox, or host adapter entrypoints.

Do not create a second desktop-specific agent implementation under desktop/src/**. Desktop supervises and adapts AgentHost; it does not own the agent semantics.

---

Ownership Rules

Core tests belong in @grida/agent first. If a bug is in files, workspaces, auth/session, providers, entitlements, runtime, tools, or stream formatting, add or fix package tests under packages/grida-ai-agent/src/**/*.test.ts.

Desktop tests should prove Electron imports, starts, and wires the core. They should not duplicate AgentHost behavior.

Anti-goals to preserve:

• Not a general model-provider router.
• Not a hosted model gateway.
• Not a billing or entitlement engine.
• Not a multi-agent orchestration graph.
• Not a UI framework.
• Not a desktop bridge package.
• Not a private chat-history IR; session rows are the contract.

---

Common Diagnostics

Empty list_files in a workspace-bound agent usually means the workspace filesystem was not hydrated. For design-agent workspace runs, @grida/agent must create AgentFs(NodeFsBackend(root)) and call await fs.hydrate() before tool calls.

An agent that sees SOME files but is missing whole subtrees (a .canvas deck, or most of the repo) is almost always hydrate-scan truncation, not a bug in those files. The walk stops at SCAN_MAX_FILES (10k) / SCAN_MAX_DEPTH and warns [agent-fs] … hydrate scan hit a cap … truncated (sidecar stderr). The usual cause is a large workspace_root containing heavy dirs that are NOT in IGNORED_SCAN_DIRS — vendored toolchains / git submodules (e.g. emsdk) or .claude/worktrees (full repo copies). .gitignore is NOT consulted, so a submodule slips through. Fixes: scope the workspace to the real project subdir, or add the offender to IGNORED_SCAN_DIRS in workspaces/scan.ts.

A client-resolved tool call that hangs at input-available (the turn just ends with no result; the assistant never continues) is the server-authoritative model view dropping it. The runtime rebuilds the model's input from the PERSISTED messages (buildModelMessages over listVisibleMessages), NOT the client's array — and it drops any tool call without a terminal result. For a workspace-less session (the desktop file-window sidebar, which resolves fs tools in the renderer over the live editor), the result lives only on the client's next-request assistant message. persistIncomingTail must persist those terminal tool-result parts (it does, as of the file-window sidebar fix) or the call stays input-available forever and the model never sees the answer. Diagnostic tell: workspace sessions show tool_state=output-available parts in sessions.db; a no-workspace session stuck at input-available for EVERY tool call is this class, not a tool bug. (Server-resolved tools are unaffected — the recorder writes their result straight from the stream.)

Broken desktop agent calls can still be package bugs. Check whether the same operation fails through AgentTransport.Client or package tests before debugging Electron.

If a method needs secrets, keep secret reads inside AgentHost. The renderer may check presence and set/delete BYOK keys, but it must not receive raw secret values.

If a change adds host-specific behavior, define the strict host capability contract first. Do not let Node, Electron, or renderer-only imports leak into neutral package entrypoints.

---

Live state on disk (inspecting a real session)

You CAN inspect a real running session — the host persists agent state under ~/.grida/agent/, separate from Electron's userData (the desktop supervisor passes it to the sidecar as --user-data; see home.join("agent") in desktop/src/main/agent-sidecar-supervisor.ts):

• sessions.db — SQLite (WAL): chat_sessions (incl. workspace_root, mode, parent_id), chat_messages, chat_parts. Schema in src/session/schema.ts.
• workspaces.json — the workspace registry (id → root).
• auth.json, recent.json.

Read it read-only (don't perturb the live WAL). The first thing to check for "why can't the agent see X" is workspace_root — the agent only sees files under it (and only up to the hydrate cap, above):

sqlite3 "file:$HOME/.grida/agent/sessions.db?mode=ro" \
  "SELECT id, workspace_root, mode FROM chat_sessions WHERE id='ses_…';"

---

Security Boundary

AgentHost's HTTP perimeter is one layer of GRIDA-SEC-004 when hosted by Desktop:

1. Per-launch Basic Auth credentials are supplied by the host adapter.
2. The preload signs requests; credentials stay in closure, not on window.grida.
3. AgentHost checks auth on every request.
4. Referer checks defend against same-origin XSS reaching the loopback server.
5. Secrets APIs expose presence/set/delete semantics, not raw secret reads.

If a route can mutate local files, start processes, read secrets, or execute tools, review both the transport contract and the host capability that grants it.

---

Verification

For package changes:

pnpm --filter @grida/agent typecheck
pnpm --filter @grida/agent test
pnpm --filter @grida/agent build

For session-store smoke checks against a real SQLite file:

pnpm --filter @grida/agent smoke:sessions:live

If the change crosses into Desktop sidecar wiring, also run the Electron owner checks from the desktop skill.

---

Pointers

• Package root: packages/grida-ai-agent/
• Package overview: packages/grida-ai-agent/README.md
• AgentHost: packages/grida-ai-agent/src/agent-host.ts
• HTTP server/routes: packages/grida-ai-agent/src/http/
• Transport: packages/grida-ai-agent/src/transport.ts
• Runtime: packages/grida-ai-agent/src/runtime/
• Sessions: packages/grida-ai-agent/src/session/
• Workspaces/files: packages/grida-ai-agent/src/workspaces.ts, packages/grida-ai-agent/src/files/, packages/grida-ai-agent/src/fs/
• Hydrate-scan policy (ignored dirs + caps): packages/grida-ai-agent/src/workspaces/scan.ts
• Live on-disk state (real sessions/workspaces): ~/.grida/agent/ (sessions.db, workspaces.json) — see "Live state on disk" above
• Providers/secrets: packages/grida-ai-agent/src/providers/, packages/grida-ai-agent/src/secrets.ts
• Skills/prompts/tools: packages/grida-ai-agent/src/skills/, packages/grida-ai-agent/src/agent/, packages/grida-ai-agent/src/tools/
• Desktop sidecar adapter: desktop/src/agent-sidecar.ts
• Renderer agent chat transport: editor/lib/agent-chat/