Skip to content

@robota-sdk/agent-framework SPEC

Overview

Robota SDK is a programming SDK built by assembling existing Robota packages. It is provider-neutral: the consumer (CLI, server, worker, etc.) creates the provider and passes it to the SDK. The primary entry point is InteractiveSession({ cwd, provider }). A createQuery({ provider }) factory is also provided for single-shot prompt use.

Core Principles

  1. Assembly first: All features are implemented using existing packages. Independent implementation is prohibited.
  2. No duplication: If the same functionality exists in an existing package, use it. Refactor the existing package if needed.
  3. Connection required: All features in agent-framework must be connected to the Robota package ecosystem.
  4. General/specialized separation: General-purpose features (permissions, hooks, tools) belong in their respective packages; only SDK-specific features (config, context) are kept in agent-framework.
  5. React-free: agent-framework is a pure TypeScript package. React hooks, React context, and React components must never be added to this package. React/Ink belongs in product shells (agent-cli) and command packages (agent-command-*). This keeps the SDK usable in any TypeScript context — CLI, web server, worker, test — without a React dependency.
  6. Assembly layer, not a re-export layer: The SDK composes sessions, runtime, tools, and core into a single SDK surface. Pass-through re-exports are only permitted through explicit SDK facade barrels (background-tasks/, subagents/). General-purpose symbols must be imported from their owner packages, not tunnelled through the SDK.

Architecture

Package Dependency Chain

agent-core           ← types, abstractions, utilities (unchanged)
agent-executor       ← background task + subagent lifecycle primitives (unchanged)
agent-session        ← Session, permissions, compaction (unchanged)
agent-tools          ← tool infrastructure + 8 built-in tools (unchanged)
agent-provider-*     ← provider implementations (unchanged)

agent-framework      ← InteractiveSession (single entry point)
  ├── embedded: SystemCommandExecutor (session.executeCommand())
  ├── embedded: CommandRegistry, BuiltinCommandSource, SkillCommandSource, PluginCommandSource
  ├── common API: command effects/interactions, lifecycle metadata, session replay validation, provider settings/profile helpers
  ├── common API: prompt file-reference parsing, resolution, diagnostics, and structured records
  ├── common API: skill discovery, skill metadata, and skill activation host context
  ├── extension: ICommandModule command/source/session-requirement injection
  ├── optional: agent runtime deps + AgentDefinitionLoader when a module requests agent-executor
  ├── composed: agent-executor BackgroundTaskManager, SubagentManager, runner ports
  ├── internal: createSession(), createDefaultTools(), loadConfig(), loadContext()
  ├── optional: sandboxClient injection for sandbox-aware built-in tool execution
  ├── optional: workspaceManifest application through agent-tools sandbox ports
  ├── optional: sandbox snapshot hydration through agent-tools sandbox ports
  ├── exposed: createQuery({ provider }) → (prompt) => result
  └── NO provider dependency (provider-neutral)

agent-command-*      ← built-in/optional command modules
  ├── consumes SDK command interfaces
  ├── consumes SDK common APIs like third-party modules
  └── NO dependency from agent-framework back to command modules

agent-cli            ← minimal TUI
  ├── creates provider (reads config, picks provider package)
  ├── selects product-default command modules such as @robota-sdk/agent-command-skills and @robota-sdk/agent-command-agent
  ├── creates InteractiveSession({ cwd, provider, commandModules })
  ├── subscribes to events → renders to terminal
  └── owns: slash prefix parsing, Ink components, paste handling, CJK input

SDK is provider-neutral. The consumer (CLI, server, etc.) creates the provider and passes it to the SDK. Assembly (wiring tools, provider, system prompt) happens inside the SDK, but the provider itself comes from the consumer.

SDK command code is split between generic infrastructure and command-facing common APIs. The SDK responsibility is the command contract layer: command contracts, registries/executors, lifecycle metadata, effects/interactions, reusable command-facing common APIs, and skill discovery/activation services consumed by command modules. User-visible internal commands, including /skills, must be implemented as command modules selected by composition roots.

Model command common APIs are provider-aware but provider-neutral. They resolve the effective active provider profile from the provider settings document, read model catalog fallback metadata from injected IProviderDefinition records, can explicitly invoke provider-owned catalog refresh hooks, and produce command descriptors without hardcoding CLI/TUI provider branches. If a live refresh fails or a provider does not expose catalog metadata, /model remains manually invocable and the command result must surface stale/unavailable catalog state rather than showing another provider's models.

Client–SDK–Session Relationship

Any client (CLI, web, API server, worker)

    │  1. creates provider:  new AnthropicProvider({ apiKey })
    │  2. creates session:   new InteractiveSession({ cwd, provider })
    │  3. subscribes:        session.on('text_delta', ...)

InteractiveSession  (agent-framework — pure TypeScript, no React)
    │  submit(input, displayInput?, rawInput?)
    │  executeCommand(name, args)
    │  executeSkillCommandByName(name, args, request)  // host API used by /skills
    │  abort() / cancelQueue()
    │  getMessages() / getContextState() / getActiveTools()
    │  (config/context loaded internally from cwd)

Session  (agent-session — generic run loop)

Robota engine + Provider  (agent-core / agent-provider-*)

agent-cli (Ink TUI — thin bridge layer)
    creates provider → passes to InteractiveSession({ cwd, provider, commandModules })
    subscribes to InteractiveSession events → maps to React/Ink state
    routes /commands → session.executeCommand()

The SDK layer has no React dependency and no provider dependency. The CLI is a TUI-only layer that creates the provider and bridges InteractiveSession events to React state.

Package Roles

PackageRoleGeneral/Specialized
agent-coreRobota engine, execution loop, provider abstraction, permissions, hooksGeneral
agent-executorBackground task and subagent lifecycle primitives, runner ports, worktree runner decoratorGeneral
agent-toolsTool creation infrastructure, sandbox execution ports, and 8 built-in toolsGeneral
agent-sessionGeneric Session class, SessionStore (persistence)General
agent-frameworkAssembly layer: InteractiveSession (single entry point), command contracts/common APIs, createQuery(), config, contextSDK-specific
agent-command-*Built-in/optional command modules that consume SDK command interfaces/common APIs and can be selected by composition rootsCommand-specific
agent-cliInk TUI and product composition. Creates provider, selects command modules, passes both to InteractiveSession. No agent-session import.CLI-specific
agent-provider-*AI provider implementations. CLI depends on these directly; SDK does not.Provider-specific

Feature Layout (Current Implementation State)

agent-core
├── src/permissions/          ← permission-gate, permission-mode, types
├── src/hooks/                ← hook-runner, hook types
└── (existing) Robota, execution, providers, plugins

agent-executor (reusable runtime primitives — depends only on agent-core)
├── src/background-tasks/     ← BackgroundTaskManager, state machine, task runner ports
└── src/subagents/            ← SubagentManager, subagent runner port, worktree runner decorator

agent-tools
├── src/builtins/             ← bash, read, write, edit, glob, grep, web-fetch, web-search tools
├── src/sandbox/              ← ISandboxClient, workspace manifest contracts, snapshot ports, E2B structural adapter, and in-memory contract adapter
├── src/types/tool-result.ts  ← TToolResult
└── (existing) FunctionTool, createZodFunctionTool, schema conversion

agent-session (generic — depends only on agent-core)
├── src/session.ts                ← Session: orchestrates run loop, delegates to sub-components
├── src/permission-enforcer.ts    ← PermissionEnforcer: tool wrapping, permission checks, hooks, truncation
├── src/context-window-tracker.ts ← ContextWindowTracker: token usage, auto-compact threshold
├── src/compaction-orchestrator.ts ← CompactionOrchestrator: conversation summarization via LLM
├── src/session-logger.ts         ← ISessionLogger + FileSessionLogger / SilentSessionLogger
├── src/session-store.ts          ← SessionStore (JSON file persistence)
└── src/index.ts

agent-framework (assembly layer — SDK-specific features only)
├── src/interactive/
│   ├── interactive-session.ts  ← InteractiveSession: event-driven wrapper over Session
│   ├── session-persistence.ts  ← SDK-owned session store facade and resumable-session summaries
│   └── types.ts                ← IToolState, IExecutionResult, IInteractiveSessionEvents
├── src/command-api/            ← Command module contracts, host context, effects/interactions, session/provider/model common APIs
│   ├── contracts.ts            ← ISystemCommand + lifecycle metadata
│   ├── command-module.ts       ← ICommandModule composition contract
│   ├── host-context.ts         ← ICommandHostContext narrow facade for command modules
│   ├── host-adapters.ts        ← generic host adapter contracts
│   ├── provider/               ← provider settings/profile/setup/probe common APIs
│   ├── model/                  ← provider-aware model catalog common APIs and refresh orchestration
│   ├── session/                ← session-history and replay-validation command common APIs
│   └── background/             ← background task command common APIs
├── src/commands/
│   ├── command-registry.ts     ← CommandRegistry: aggregates ICommandSource instances
│   ├── builtin-source.ts       ← BuiltinCommandSource: SDK core compatibility source; currently empty
│   ├── skill-source.ts         ← SkillCommandSource: discovers SKILL.md files
│   ├── plugin-source.ts        ← PluginCommandSource: discovers plugin commands (moved from agent-cli)
│   └── system-command.ts       ← SDK core command factory; currently empty because user-visible built-ins are command modules
├── src/assembly/               ← Session factory: createSession (internal), createDefaultTools (internal)
├── src/config/                 ← settings.json loading (6-layer merge, $ENV substitution)
├── src/context/                ← AGENTS.md/CLAUDE.md/memory discovery, project detection, system prompt
│   ├── context-reference-inventory.ts ← session context reference metadata, active/observed status, and bounded inventory policy
│   ├── prompt-file-reference-*.ts ← `@file` prompt reference parser/resolver, path policy, formatting, and diagnostics
│   └── task-context.ts         ← active `.agents/tasks/*.md` discovery, selection, formatting, and status updates
├── src/memory/                 ← project memory store, reusable capture policy, retrieval services
├── src/user-local/             ← user-local storage root validation, category projections, and baseline memory persistence
├── src/checkpoints/            ← edit checkpoint store + Write/Edit tool snapshot wrapper
├── src/self-hosting/           ← self-hosting verification planner + lifecycle state machine
├── src/tools/agent-tool.ts     ← Agent sub-session tool (SDK-specific: uses createSession)
├── src/subagents/              ← SDK in-process runner + explicit compatibility exports from agent-executor
├── src/background-tasks/       ← explicit compatibility exports from agent-executor
├── src/permissions/            ← permission-prompt.ts (terminal approval prompt)
├── src/paths.ts                ← projectPaths / userPaths helpers
├── src/types.ts                ← internal terminal type aliases; not a top-level public barrel
├── src/query.ts                ← createQuery() factory (provider-neutral; provider injected by consumer)
└── src/index.ts                ← SDK-owned APIs plus explicit SDK facade exports

agent-cli (Ink TUI — CLI-specific)
├── src/ui/                     ← App, MessageList, InputArea, StatusBar, PermissionPrompt,
│                                  SlashAutocomplete, CjkTextInput, WaveText, InkTerminal, render
├── src/permissions/            ← permission-prompt.ts (terminal arrow-key selection)
├── src/types.ts                ← ITerminalOutput, ISpinner (duplicate — SSOT is agent-session)
├── src/cli.ts                  ← CLI argument parsing, Ink render
└── src/bin.ts                  ← Binary entry point

Feature Details

Session Management

  • Package: agent-session (generic, depends only on agent-core)
  • Implementation: Session accepts pre-constructed tools, provider, and system message. Internal concerns are delegated to PermissionEnforcer, ContextWindowTracker, and CompactionOrchestrator.
  • Assembly: agent-framework/assembly/ provides createSession() (internal — not exported) which wires tools, provider, and system prompt from config/context. Consumers use InteractiveSession({ cwd, provider }) instead.
  • Persistence: SessionStore defaults to ~/.robota/sessions/{id}.json for generic session consumers. SDK exposes createProjectSessionStore(cwd) and resumable-session helpers so CLI composition can use project-local .robota/sessions without importing agent-session directly.
  • Replay validation common API: SDK command APIs expose validateCommandSessionReplayLog() and formatting helpers that load the current session's project-local .robota/logs/{sessionId}.jsonl file through agent-session replay validators. Command modules consume this API; agent-cli must not read replay logs or implement replay validation directly.

Permission System

  • Package: agent-core (general-purpose security layer)
  • Implementation: 3-step evaluation — deny list → allow list → mode policy
  • Modes: plan (read-only), default (write requires approval), acceptEdits (write auto-approved), bypassPermissions (all auto-approved)
  • Pattern syntax: Bash(pnpm *), Read(/src/**), Write(*) etc. with glob matching
  • Terminal prompt: agent-framework/src/permissions/permission-prompt.ts is the SSOT implementation of the terminal approval prompt. Used by both InteractiveSession/createQuery() and agent-cli (which imports from @robota-sdk/agent-framework).
  • Default allow patterns: createSession() automatically adds allow patterns for config folder access: Read(.agents/**), Read(.claude/**), Read(.robota/**), Glob(.agents/**), Glob(.claude/**), Glob(.robota/**). These are merged with user-configured permissions.

Hooks System

  • Package: agent-core (general-purpose extension points)
  • Events: PreToolUse, PostToolUse, PreCompact, PostCompact, SessionStart, Stop
  • Implementation: Executes shell commands, passes JSON via stdin, determines allow(0)/deny(2) by exit code
  • Matcher: Tool name regex pattern matching

Tool System

  • Infrastructure: agent-tools (createZodFunctionTool, FunctionTool, Zod→JSON conversion)
  • Built-in tools: agent-tools/builtins/ — Bash, Read, Write, Edit, Glob, Grep, WebFetch, WebSearch
  • Agent runtime deps: agent-framework/tools/agent-tool.ts stores reusable subagent runtime dependencies for /agent and context: fork skill execution when a composed command module requests agent-executor. createSession() does not register a separate model-visible Agent tool; model and user routing use the built-in command layer such as /agent.
  • Edit checkpoint wrapper: agent-framework/checkpoints/edit-checkpoint-tools.ts wraps Write and Edit at SDK session assembly time. The underlying tool package stays generic; the SDK wrapper snapshots the target file before the first mutation in each prompt turn.
  • Tool result type: TToolResult in agent-tools/types/tool-result.ts

Sandbox Execution

  • Port owner: agent-tools/sandbox/ owns ISandboxClient, ISandboxRunOptions, and ISandboxRunResult.
  • Workspace manifest owner: agent-tools/sandbox/ owns IWorkspaceManifest, manifest path validation, and generic manifest application. SDK and CLI must not redeclare the manifest shape or implement manifest application algorithms.
  • Adapter owner: agent-tools owns structural sandbox adapters such as E2BSandboxClient and InMemorySandboxClient. It does not install provider SDKs; application composition roots may install e2b or another provider and wrap its sandbox object.
  • SDK assembly: createSession() and InteractiveSession accept sandboxClient?: ISandboxClient. When present, SDK-created Bash, Read, Write, and Edit tools are created through sandbox-aware factories and route command/filesystem operations through the injected client.
  • Fresh workspace setup: InteractiveSession accepts workspaceManifest?: IWorkspaceManifest with sandboxClient. The async interactive construction path applies the manifest once before creating the Session, using the current project cwd as the host root for relative localFile and localDir sources. The internal synchronous createSession() factory does not apply manifests; direct low-level consumers call applyWorkspaceManifest() from agent-tools before session creation.
  • Snapshot hydration: If sessionStore and a sandbox client with snapshot() are provided, InteractiveSession.shutdown() captures a provider-owned sandboxSnapshotId and stores it in the session record. On non-fork resumeSessionId, SDK initialization loads the saved id and calls sandboxClient.restore(snapshotId) before constructing the underlying Session and before deferred saved-message injection. When a snapshot is restored, fresh workspaceManifest application is skipped to avoid overwriting hydrated state.
  • Reversible execution: If reversibleExecution.mode is enabled and no explicit isolation is set, a supplied sandboxClient makes the SDK classify Bash and sandbox-routed file mutations as provider-sandbox isolated instead of host checkpoint-backed mutations.
  • Boundary: agent-cli and other hosts only decide whether to provide a sandbox client and, if they parse manifest files, convert them into the agent-tools manifest contract. They must not implement sandbox command/file/manifest algorithms or provider-specific restore behavior in UI code.

Edit Checkpointing

  • Package: agent-framework/checkpoints/ (SDK-specific session safety layer)
  • Storage: Project-local .robota/checkpoints/{session-id}/{turn-id}/manifest.json plus copied pre-image files under files/.
  • Turn model: Every cwd-backed InteractiveSession.submit() prompt starts a turn-level checkpoint. The checkpoint is finalized after the run finishes, even when no file was edited, so prompt turns can be listed consistently. Injected sessions without cwd do not implicitly create project checkpoints; they must provide cwd or use explicit checkpoint APIs.
  • Capture model: Write and Edit tools are wrapped during createSession() assembly when an IEditCheckpointRecorder is present. A file is captured once per turn before the first tool mutation. Repeated edits to the same file in the same turn reuse the first pre-image.
  • Inspection model: inspect(sessionId, checkpointId) returns captured files, workspace-relative display paths, snapshot availability, and the restore/rollback checkpoint ranges before a caller mutates the workspace.
  • Restore model: restoreToCheckpoint(sessionId, checkpointId) rolls back later checkpoints in reverse sequence order, restores copied pre-images, deletes files that did not exist at capture time, and removes later checkpoint directories. This provides code-only rewind to the selected prompt turn.
  • Boundary: agent-tools does not know about sessions, prompts, .robota, or checkpoints. CLI/TUI does not implement checkpoint algorithms; it only exposes SDK command output and future picker UI.
  • Current scope: Write and Edit mutations are tracked. Shell-side filesystem changes from Bash are not tracked by this layer.

Reversible Execution Mode

  • Package: agent-framework/reversible-execution/ (SDK-specific safety classification and opt-in tool wrapper)
  • Mode: createSession({ reversibleExecution: { mode: 'local-first' } }) enables local-first reversible execution enforcement. The mode is opt-in while provider sandbox snapshots are still future work.
  • File mutations: Write and Edit are reversible only when an edit checkpoint recorder is present. Without a checkpoint, local-first mode blocks the tool before mutation.
  • Shell/process side effects: Bash and BackgroundProcess are not checkpoint-restorable in the parent workspace. Local-first mode requires worktree or provider-sandbox isolation before allowing those side effects.
  • Agent jobs: Agent jobs are reversible through the worktree layer only when the request or all batch jobs specify isolation: 'worktree', or when the outer execution context is already isolated.
  • Read-only tools: Read, Glob, Grep, WebFetch, and WebSearch are classified as read-only and do not require rollback.
  • Boundary: The policy is SDK-owned and provider-neutral. It does not execute Git commands, manage provider sandboxes, or render UI. Runtime shells and provider adapters supply actual worktree/sandbox isolation.

Self-Hosting Verification

  • Package: agent-framework/self-hosting/ (SDK-specific planning layer)
  • Purpose: Describes the safe edit/build/verify loop for Robota modifying its own source tree without replacing the currently running process.
  • Planner: planSelfHostingVerification() returns ordered steps for checkpoint creation, atomic file mutation, external process handoff, targeted package verification, harness verification, and rollback recovery.
  • State machine: transitionSelfHostingLoop() enforces deterministic lifecycle transitions from idle through checkpoint/edit/verify success or failure recovery.
  • Handoff model: The current process remains the old runtime and keeps already-loaded modules. Verification commands run in child processes against the new on-disk tree.
  • Boundaries: The SDK planner does not implement file writing, checkpoint storage, CLI rendering, or provider behavior. Atomic write behavior belongs to agent-tools; checkpoint storage belongs to agent-framework/checkpoints; CLI/TUI only invokes SDK APIs and renders results.
  • Verification defaults: For supplied package scopes, the default plan includes test, typecheck, and build commands before pnpm harness:verify -- --base-ref <ref> --skip-record-check. The harness verification step is always present.
  • Local tools: WebSearch and WebFetch are ordinary Robota function tools from @robota-sdk/agent-tools. They are available to CLI sessions as explicit local tools and are separate from provider-native hosted web features.
  • Provider-native tools: Anthropic server web search and Qwen Responses web search/fetch are exposed through provider-owned capability reports. Provider capability text must come from the provider adapter or tool descriptor; the system prompt composer must not inject hardcoded web-search behavior instructions.
  • Activation: Session layers use the provider-neutral configureNativeWebTools() hook when a provider chooses to expose automatic hosted web behavior. SDK must not branch on concrete provider names or mutate provider-specific fields directly.
  • Callback: onServerToolUse fires during streaming when a provider-owned server tool executes, allowing the UI to display search status.

Streaming

  • Implementation: TTextDeltaCallback type (IChatOptions in agent-core)
  • Behavior: AnthropicProvider uses the streaming API, returning the completed message while calling the callback for each text delta
  • UI connection: Session → onTextDelta → InteractiveSession text_delta event → client

InteractiveSession (SDK-Specific)

  • Package: agent-framework/interactive/
  • Pattern: Composition over Session (holds a Session instance, does not extend it)
  • Constructor: Accepts { cwd, provider } plus optional composition inputs such as commandModules. Config and context are loaded internally from cwd.
  • Responsibility: Streaming accumulation, tool state tracking, prompt queue (max 1), abort orchestration, full history management (IHistoryEntry[]), embedded command execution
  • Tool execution history: Each tool_start and tool_end event is recorded as an individual IHistoryEntry with category: 'event' and type: 'tool-start' or type: 'tool-end'. Data includes toolName, firstArg, isRunning, and result. For completed Edit tools, IToolState also carries diffFile and diffLines derived from the Edit tool arguments plus the tool result startLine. For completed command tools, IToolState carries toolResultData so transports can render bounded command output previews while raw tool messages remain persisted. The tool-summary entry (aggregated) is still pushed at execution completion and preserves the same per-tool metadata for persisted UI rendering.
  • Events: text_delta, tool_start, tool_end, thinking, complete, error, context_update, interrupted
  • submit() signature: submit(input, displayInput?, rawInput?)displayInput overrides what appears in the client's message list; rawInput is passed to Session.run() for hook matching
  • Prompt file references: Before a non-command prompt reaches Session.run(), InteractiveSession delegates to the SDK-owned prompt file-reference resolver. Path-like tokens such as @AGENTS.md, @./Makefile, and @docs/spec.md are resolved relative to the session cwd, constrained to the workspace root, bounded by explicit file/total byte limits, and expanded into model-only prompt context blocks. The user-visible history keeps the original prompt and records a prompt-file-reference event with structured records (sourcePath, relativePath, originalReference, reason, depth, byteLength) without storing file contents in the event. Missing, outside-root, directory, circular, max-depth, and size-limit failures are blocking diagnostics and the prompt is not sent to the provider.
  • executeCommand(): executeCommand(name, args) — executes a named system command via the embedded SystemCommandExecutor. Product composition roots inject command modules such as /compact; SDK-default user-visible commands are intentionally empty.
  • Edit checkpoints: listEditCheckpoints() returns checkpoint summaries for the active session. inspectEditCheckpoint(id) returns captured files and restore/rollback plans. restoreEditCheckpoint(id) restores code to a prior checkpoint and records a system history entry. It is rejected while a prompt is running.
  • listCommands(): listCommands() — returns Array<{ name, description }> of all registered system commands. Used by transport adapters (e.g., MCP) to expose commands as tools.
  • Queue behavior: If executing is true, the incoming prompt is queued. The queued prompt auto-executes after the current one completes. Only one prompt can be queued at a time.
  • Abort: abort() clears the queue and delegates to session.abort(). An interrupted event fires when the abort completes.
  • No-op terminal: Uses a built-in NOOP_TERMINAL so no ITerminalOutput implementation is required by callers
  • Session persistence: When an SDK-owned sessionStore facade is provided in options, auto-persists session state (messages, history, cwd, timestamps, system prompt, tool schemas, memory events, used memory references, and provider sandbox snapshot ids when available) after each submit() completion and on shutdown. The SDK facade delegates to the concrete SessionStore implementation from agent-session internally and exposes resumable-session summaries for hosts such as the CLI. Session JSON is the fast snapshot, while append-only JSONL replay logs are the recovery source when the JSON snapshot is missing.
  • Session restore: When resumeSessionId is provided, loads the saved session record and restores AI context. The project session store first reads .robota/sessions/{id}.json; if it is missing, it replays .robota/logs/{id}.jsonl through agent-session replay readers and reconstructs messages/history from history_mutation events. For non-fork resumes with sandboxSnapshotId, the SDK restores the sandbox before constructing the underlying Session and before injecting messages. Messages are stored as pendingRestoreMessages and injected via session.injectMessage() after async initialization completes (deferred injection pattern). Memory event history and the last used memory references are restored for /memory used and debugging. This avoids injection failures caused by the Session not yet being fully initialized when the constructor runs.
  • forkSession option: forkSession?: boolean (default false). When false (resume), the original session ID is passed to the Session constructor so it reuses the same file. When true (fork), sessionId is omitted, generating a fresh UUID — the original session remains untouched.
  • getName()/setName(name): Get or set the session's user-facing name. Persists to the session record when a store is configured.
  • attachTransport(transport): attachTransport(transport: ITransportAdapter) — attaches a transport adapter to this session. Calls transport.attach(this). Used by consumers to compose transports consistently: session.attachTransport(transport); await transport.start();
  • Testing: Accepts an optional pre-built Session via options.session to enable unit testing without I/O setup

Command API Layer (SDK-Specific)

  • Package: agent-framework/command-api/
  • Purpose: Stable SDK-owned API layer consumed by built-in and third-party command modules. It is pure TypeScript, render-agnostic, provider-neutral, and has no CLI/TUI dependency.
  • Contracts:
    • ISystemCommand — command metadata, lifecycle, model/user visibility, and execute function.
    • ICommandModule — composition unit contributing command sources, executable commands, descriptors, and session requirements.
    • ICommandHostContext — narrow command-facing facade over session/context/runtime capabilities. Command modules must not require InteractiveSession, React state, CLI settings files, or TUI hooks directly.
    • ICommandResult — command output, structured diagnostics, typed host effects, and generic interactions.
    • TCommandEffect — typed host-applied effects such as model/language change, restart, exit, session picker, plugin UI, plugin registry reload, rename, and statusline patch.
    • ICommandInteraction / TCommandInteractionPrompt — generic command-owned follow-up prompts rendered by host UIs. Prompt descriptors may include a provider-neutral description string for host-rendered help text.
  • Provider common APIs: agent-framework/command-api/provider/ owns provider settings document types, provider profile merge/validation/delete helpers, environment reference helpers, setup-flow primitives including fixed-profile edit defaults, provider-owned setup help link projection, provider profile name suggestion helpers, provider command settings adapter contracts, and provider probe defaults. provider command behavior lives in @robota-sdk/agent-command-provider and consumes these APIs as an external command module.
  • Context/compact common APIs: agent-framework/command-api/context/ owns command-facing context-state reads, automatic compact policy reads, active-session policy updates, settings-adapter persistence helpers, and manual compact host-facade helpers. context and compact command behavior lives in @robota-sdk/agent-command-context and @robota-sdk/agent-command-compact; both consume these APIs as external command modules.
  • Model common APIs: agent-framework/command-api/model/ owns model-command metadata constants and subcommand projection helpers. model command behavior lives in @robota-sdk/agent-command-model and consumes these APIs as an external command module.
  • Language common APIs: agent-framework/command-api/language/ owns language-command metadata constants, recommended subcommands, argument parsing, and usage formatting. language command behavior lives in @robota-sdk/agent-command-language and consumes these APIs as an external command module.
  • Memory common APIs: agent-framework/command-api/memory/ owns memory-command metadata constants, subcommand projection helpers, project/pending memory store facades, sensitive-content checks, used-memory reference reads, and memory-event recording helpers. memory command behavior lives in @robota-sdk/agent-command-memory and consumes these APIs as an external command module.
  • Background common APIs: agent-framework/command-api/background/ owns background-command metadata constants, subcommand projection helpers, task-list/log formatting helpers, and list/read/cancel/close facades over ICommandHostContext. background command behavior lives in @robota-sdk/agent-command-background and consumes these APIs as an external command module.
  • Help common APIs: agent-framework/command-api/help/ owns help-command metadata constants and generic command-list formatting. help command behavior lives in @robota-sdk/agent-command-help and consumes this API as an external command module.
  • Permission common APIs: agent-framework/command-api/permissions/ owns permission-mode constants, descriptor subcommands, validation, permission-state reads, permission-state formatting, and command-facing adapter resolution. Canonical permission command behavior lives in @robota-sdk/agent-command-permissions, which owns /permissions [mode]. Legacy /mode behavior lives in @robota-sdk/agent-command-mode only for applications that explicitly compose that optional module. Both consume these APIs as external command modules.
  • Statusline common APIs: agent-framework/command-api/statusline/ owns statusline command metadata constants, subcommand projection helpers, default settings shape, typed settings patch contracts, and patch validation. statusline command behavior lives in @robota-sdk/agent-command-statusline and emits typed host-applied effects instead of importing CLI settings utilities.
  • Plugin common APIs: agent-framework/command-api/plugin/ owns plugin command metadata constants, subcommand projection helpers, ICommandPluginAdapter, reload result contracts, and plugin host effect factories. plugin and reload-plugins command behavior lives in @robota-sdk/agent-command-plugin and consumes these APIs as an external command module while hosts keep concrete plugin storage/UI wiring.
  • Session common APIs: agent-framework/command-api/session/ owns command-facing session-history helpers, session-name parsing, session-info reads, and effect factories for host-rendered history/name/picker/exit state. clear, rename, resume, and cost command behavior lives in @robota-sdk/agent-command-session; exit command behavior lives in @robota-sdk/agent-command-exit. Both consume these APIs as external command modules.
  • Settings/process effects: agent-framework/command-api/effects.ts owns the typed settings-reset-requested effect. reset command behavior lives in @robota-sdk/agent-command-reset and emits that effect without importing host settings file I/O.
  • Checkpoint common APIs: agent-framework/command-api/checkpoint/ owns command-facing checkpoint metadata constants, subcommand projection helpers, and inspect/list/restore/rollback facades over ICommandHostContext. rewind command behavior lives in @robota-sdk/agent-command-rewind and consumes these APIs as an external command module.
  • Boundary: command-api may define contracts and reusable command-facing helpers. It must not own product UI, concrete settings file I/O, process restart/exit, provider construction, or command-specific flows that can live in agent-command-* packages.

Transparent Workflow Contract (SDK-Specific)

The cross-cutting contract lives in ../../../.agents/specs/transparent-workflow.md. The SDK is the designated owner for reusable transparent workflow contracts and projections:

  • any new action provenance types and execution eligibility helpers;
  • mapping runtime task states into the shared user-facing state vocabulary;
  • execution workspace read models for main-thread, background task, and background group switching;
  • any new memory and preference inspection/removal API shapes;
  • command-facing facades that let agent-command-* expose status and memory controls without importing CLI code.

IExecutionOrigin is the current task/workspace origin projection. It is not command authorization provenance by itself. Future transparent workflow implementation PRs must add or extend typed action provenance before new host command/process execution surfaces depend on it.

User-local preferences, remembered values, and session state may influence display and navigation, but they must not execute commands. Shell/process/harness command execution must originate from direct user input or an assistant suggestion accepted through explicit UI approval or the current user-selected permission policy.

User-Local Storage Foundation (SDK-Specific)

The cross-cutting storage policy lives in ../../../.agents/specs/user-local-storage.md. The SDK is the designated owner for baseline workflow storage root resolution, repo-outside validation, category contracts, and item inspection/removal projections.

Existing projectPaths(cwd) helpers remain valid for explicit project-owned features such as project settings, session replay/debug logs, edit checkpoints, and current project memory. New baseline transparent workflow state must not use projectPaths(cwd) or ad hoc .robota/ paths. It must use SDK-owned user-local storage contracts.

Existing userPaths() helpers expose only current user settings and sessions paths. User-local workflow state uses the tested src/user-local/ APIs instead of CLI or command modules assembling category paths themselves.

User-Local Memory Transparency (SDK-Specific)

The baseline user-local memory contract lives in ../../../.agents/specs/user-local-memory.md. The SDK is the designated owner for memory item projection shapes, display/navigation disclosure rules, inspection APIs, delete/disable APIs, and disabled-item non-use.

User-local memory may influence display and navigation only. It must not execute shell/process commands, select repository harness commands, grant permissions, inject hidden prompt behavior, or become the execution cwd for a new command by itself.

Existing project memory under .robota/memory/ remains a separate explicit project-memory feature. New baseline local preferences, last-view state, and task associations must use the SDK user-local storage contract instead of project memory paths.

Transparent Process Execution (SDK-Specific)

The process execution contract lives in ../../../.agents/specs/process-execution.md. The SDK is the designated owner for process execution request/status projections that sit above runtime process tasks:

  • action provenance attached to user-directed process execution;
  • display-safe environment summaries;
  • working-directory projection;
  • foreground/background process status projection;
  • duration and terminal-result projection;
  • retention and transcript pointers consumed by command modules, transports, and CLIs.

Existing BackgroundProcess and execution workspace APIs are the current building blocks. Future user-facing process-run commands must use SDK/runtime contracts and must not let CLI components assemble process semantics from raw child-process state.

Repository Situational Awareness (SDK-Specific)

Passive repository context display is specified in ../../../.agents/specs/repository-situational-awareness.md. The SDK is the designated owner for context item projections, provenance fields, and bounded read contracts for cwd, repository root, branch, dirty summary, explicit context references, and active background workspace context.

Situational awareness projections must not infer commands, package managers, CI mappings, repository readiness, setup profiles, or harness contracts. Existing context loading may continue to serve prompt construction, but passive display surfaces must use explicit projection APIs instead of reusing broad context-loading internals for repository interpretation.

System Command System (SDK-Specific)

  • Package: agent-framework/commands/
  • Purpose: SDK command infrastructure and command-facing common APIs — pure TypeScript, no React, no TUI dependency
  • Embedding: SystemCommandExecutor is embedded inside InteractiveSession. Consumers normally call session.executeCommand(name, args) directly. SystemCommandExecutor and createSystemCommands() are exported so independent command modules can compose and test against the same command contract.
  • Classes:
    • SystemCommandExecutor — registry + executor for ISystemCommand instances (internal to InteractiveSession)
    • createSystemCommands() — SDK core executable command factory; currently returns an empty list because user-visible built-ins live in agent-command-*
    • createBuiltinCommandModule() — SDK core compatibility module; currently empty
  • Design: Commands return ICommandResult with message, success, and optional SDK-owned effects and interaction contracts. data remains available for command-specific diagnostic payloads, but callers must not invent command-specific side-effect keys. User-facing follow-up prompts are represented by ICommandInteraction, and host actions such as restart, shutdown, plugin UI, plugin registry reload, session picker, model/language changes, session rename, and status-line updates are represented by typed TCommandEffect values.
  • Single owner rule: SDK-default built-in command metadata is derived from executable ISystemCommand records. A built-in command must not be added to autocomplete/help metadata without an executable owner module.
  • Lifecycle policy: ISystemCommand may declare command lifecycle metadata. Blocking foreground commands share the same InteractiveSession execution guard and thinking events as prompt execution. Inline commands execute immediately and must not call model-backed long-running operations.
  • Command identity: ICommand.name, ISystemCommand.name, ICapabilityDescriptor.name, and projected model-command reverse mappings use slash-free canonical command ids such as skills, agent, and memory. Slash syntax such as /skills or /agent belongs only to UI/transport input parsing and display.
  • SDK core built-ins: SDK core has no user-visible built-in commands. skills is owned by @robota-sdk/agent-command-skills, which consumes SDK skill discovery and activation APIs like any other command module.
  • Product-specific built-in commands: User-visible internal commands outside SDK-owned discovery are provided by product-composed command modules.
  • Product-composed built-in command modules: skills is provided by @robota-sdk/agent-command-skills. It is user- and model-invocable, lists registered skill metadata, and activates a skill through ICommandHostContext.executeSkillCommandByName(). Model-side activation uses the projected robota_command_skills tool with skill arguments in args.
  • Product-composed built-in command modules: help is provided by @robota-sdk/agent-command-help and renders the composed command list through SDK help common APIs.
  • Product-composed built-in command modules: model is provided by @robota-sdk/agent-command-model, reuses SDK model-command common APIs for subcommand metadata, and emits model-change-requested effects for host application.
  • Product-composed built-in command modules: permissions is provided by @robota-sdk/agent-command-permissions, reuses SDK permission common APIs for validation/subcommand metadata, state reads/formatting, and permission-mode updates through the command host adapter facade, and stays user-invocable only.
  • Optional legacy command modules: mode is provided by @robota-sdk/agent-command-mode only when an application explicitly composes that module. Product CLIs should prefer the canonical permissions command for permission-mode changes.
  • Product-composed built-in command modules: language is provided by @robota-sdk/agent-command-language, reuses SDK language command common APIs for usage/subcommand metadata, and emits language-change-requested effects for host application.
  • Product-composed built-in command modules: statusline is provided by @robota-sdk/agent-command-statusline, reuses SDK statusline common APIs for subcommand metadata and typed patch effects, and leaves status bar rendering/settings persistence to the host.
  • Product-composed built-in command modules: clear, rename, resume, and cost are provided by @robota-sdk/agent-command-session. clear reuses SDK session command common APIs to clear SDK session history and emits conversation-history-cleared so hosts clear rendered history through their own UI state. rename reuses SDK session command common APIs to normalize the requested name and emits session-renamed so hosts update title/status/persistence through their own adapters. resume emits session-picker-requested so hosts display saved-session picker UI through their own adapters. cost reads session id and message count through SDK session command common APIs.
  • Product-composed built-in command modules: reset is provided by @robota-sdk/agent-command-reset. It emits settings-reset-requested so hosts apply concrete settings deletion and shutdown at their own adapter/UI boundary.
  • Product-composed built-in command modules: rewind is provided by @robota-sdk/agent-command-rewind. It reuses SDK checkpoint command common APIs to list prompt-turn checkpoints, inspect captured files and restore plans, restore code to a selected checkpoint, or roll back through a selected checkpoint.
  • Product-composed built-in command modules: memory is provided by @robota-sdk/agent-command-memory. It reuses SDK memory command common APIs to inspect project memory, save durable entries, review pending candidates, record memory audit events, and report memory provenance.
  • Product-composed built-in command modules: background is provided by @robota-sdk/agent-command-background. It reuses SDK background command common APIs to list tasks, read logs, cancel queued/running work, and close terminal task records without SDK core embedding command registration.
  • Product-composed built-in command modules: context is provided by @robota-sdk/agent-command-context and reports context window usage plus auto-compact policy through the SDK command host facade. context auto ... uses the same common API layer to update the active session immediately and persist through host-provided settings adapters.
  • Product-composed built-in command modules: compact is provided by @robota-sdk/agent-command-compact, declares blocking lifecycle metadata through the same ISystemCommand contract, and is exposed as a model-invocable write capability. Auto-compaction remains a deterministic session policy and emits structured compaction events instead of relying on the model to decide routine compaction.
  • Product-composed built-in command modules: exit is provided by @robota-sdk/agent-command-exit. It reuses the SDK session-exit effect helper, stays user-invocable only, and leaves concrete shutdown/process exit to the host effect handler.
  • Product-composed built-in command modules: plugin and reload-plugins are provided by @robota-sdk/agent-command-plugin. They reuse SDK plugin command common APIs, send host UI opening through plugin-tui-requested, refresh host plugin command sources through plugin-registry-reload-requested, and perform install/uninstall/enable/disable/marketplace/reload operations through a host-provided ICommandPluginAdapter.
  • Model-invocable built-ins: Product-composed command modules such as skills, agent, memory, and compact expose descriptors so explicit user/model requests can execute through SDK-projected provider-safe command tools such as robota_command_skills. The descriptor owns usage metadata and autonomous-use guidance; the system prompt composer must not add separate behavior instructions.
  • rewind: User-invocable product-composed code checkpoint command. rewind list lists prompt-turn checkpoints; rewind inspect <checkpoint-id> shows captured files plus restore/rollback ranges; rewind restore <checkpoint-id> and rewind code <checkpoint-id> restore files to the selected checkpoint. It is not model-invocable by default.
  • Command modules: Optional ICommandModule instances may contribute ICommandSource palette metadata, ISystemCommand handlers, model-visible descriptors, and session requirements. The SDK does not know command names contributed by modules in advance. Product assemblies can inject host-owned built-ins such as plugin and product-composed command packages such as exit and statusline without adding CLI-specific code to SDK core.

Slash Command Registry (SDK-Specific)

  • Package: agent-framework/commands/ — SSOT owner; agent-cli re-exports from here
  • Classes:
    • CommandRegistry — aggregates multiple ICommandSource instances; filters by prefix; resolves plugin-qualified names
    • BuiltinCommandSource — SDK core compatibility command source; currently empty
    • SkillCommandSource — SDK common API that discovers SKILL.md files from project and user directories; command modules may use it for virtual skill palette metadata
    • PluginCommandSource — discovers commands exposed by installed bundle plugins (moved from agent-cli to agent-framework)
  • Migration note: These classes were previously in agent-cli/src/commands/. They were moved to agent-framework so any client can use slash command discovery without a TUI dependency. PluginCommandSource was also moved from agent-cli to agent-framework as part of the scope redesign.

Config Loading (SDK-Specific)

  • Package: agent-framework/config/
  • Rationale: .robota/settings.json file-based configuration is for local development environments only (servers use environment variables/DB)
  • Implementation: settings file merge, $ENV:VAR substitution for provider API keys, Zod validation, provider profile resolution
  • Provider profiles: settings may define currentProvider and providers. The active profile is resolved from providers[currentProvider], then normalized into IResolvedConfig.provider. Profile identity is the profile key, not the provider type or model pair. Setup helpers suggest readable model-derived keys and append numeric suffixes when the key already exists. Generic provider credentials use apiKey; provider-specific advanced authentication belongs in provider-owned options or injected clients, not in generic profile fields.
  • Legacy compatibility: legacy provider settings remain supported and are used when no active provider profile is configured.

Provider profile shape:

json
{
  "currentProvider": "supergemma4-26b-uncensored-v2",
  "providers": {
    "supergemma4-26b-uncensored-v2": {
      "type": "gemma",
      "model": "supergemma4-26b-uncensored-v2",
      "apiKey": "lm-studio",
      "baseURL": "http://localhost:1234/v1"
    },
    "gpt-4o": {
      "type": "openai",
      "model": "gpt-4o",
      "apiKey": "$ENV:OPENAI_API_KEY"
    },
    "qwen3-6-plus": {
      "type": "qwen",
      "model": "qwen3.6-plus",
      "apiKey": "$ENV:DASHSCOPE_API_KEY",
      "options": {
        "builtInWebTools": {
          "webSearch": true,
          "webFetch": true
        }
      }
    }
  }
}

Gemma-family local models should be configured through type: "gemma" so provider-specific stream projection is applied. DeepSeek API profiles should be configured through type: "deepseek" so provider-specific defaults, model catalog metadata, and thinking controls remain provider-owned. type: "openai" remains a model-family neutral OpenAI-compatible transport profile.

Provider profile options are preserved as provider-owned data. SDK config loading validates that the value is universal/JSON-like and passes it through; SDK code must not interpret provider-specific option keys. OpenAI-compatible local endpoints such as LM Studio should use local WebSearch/WebFetch function tools for web access unless their concrete provider package documents and enables provider-native hosted web capabilities.

Generated provider profile keys are normalized to lowercase ASCII slugs. The setup flow prefers the selected model id, falls back to provider type, and appends -2, -3, etc. for duplicates. Secrets, organizations, accounts, and API key fragments must not be included in generated keys.

Provider setup help links come from injected IProviderDefinition.setupHelpLinks records. The SDK formats those provider-owned links for generic prompts, but it does not choose provider URLs or branch on provider names. Link priority is API key issuance URL, then official console URL, then official provider documentation or homepage URL.

Resolved provider fields:

FieldDescription
nameProvider type used by session model config (anthropic, openai, gemma)
modelActive model id
apiKeyAPI key or local placeholder token
baseURLOptional OpenAI-compatible endpoint override
timeoutOptional provider idle timeout in milliseconds. Also passed to provider construction when supported
optionsOptional provider-owned options bag preserved for CLI/provider composition

Context Loading (SDK-Specific)

  • Package: agent-framework/context/
  • Rationale: AGENTS.md/CLAUDE.md walk-up discovery is for local development environments only
  • Implementation: Directory traversal from cwd to root, project type/language detection, .robota/memory/MEMORY.md startup memory loading, active task context loading, system prompt assembly
  • Response Language: IResolvedConfig.language (from settings.json language field) is rendered as neutral metadata by buildSystemPrompt(). Persists across compaction because system message is preserved.
  • Compact Instructions: Extracts "Compact Instructions" section from CLAUDE.md and passes to Session for compaction
  • Skill Discovery Paths: Skills are discovered from .agents/skills/*/SKILL.md (project), .claude/skills/*/SKILL.md, .claude/commands/*.md, and ~/.robota/skills/*/SKILL.md. Used by conditional SDK skill metadata injection when /skills is model-invocable, and by @robota-sdk/agent-command-skills for virtual skill command palette metadata.

Active Task Context (SDK-Specific)

  • Package: agent-framework/context/task-context.ts
  • Purpose: Treat active .agents/tasks/*.md files as bounded working-memory metadata for the current session.
  • Discovery: Only direct Markdown files under .agents/tasks/ are eligible. README.md and files under .agents/tasks/completed/ are excluded.
  • Selection: Task selection is bounded. Matching - **Branch**: metadata for the current git branch takes precedence, followed by in-progress, todo, then unknown status. Completed tasks are excluded.
  • Formatting: formatTaskContext() renders selected task metadata as neutral Markdown under Active Task Context. It includes path, title, status, branch, scope, objective, and unchecked completion items. It must not add behavior instructions.
  • Prompt integration: loadContext() stores formatted task context in ILoadedContext.taskContext; buildSystemPrompt() renders it after project memory and before runtime metadata. Compaction preserves it because the system message is preserved.
  • Status synchronization: updateTaskFileStatus() updates or inserts the task status metadata and appends a dated progress entry when a progress message is supplied. The function accepts an injected clock for deterministic tests.

Project Memory (SDK-Specific)

  • Package: agent-framework/memory/
  • Storage: .robota/memory/MEMORY.md is the project memory index; .robota/memory/topics/*.md stores topic details.
  • Startup injection: loadContext() reads the memory index into ILoadedContext.memoryMd; buildSystemPrompt() renders it under the neutral Project Memory section. Topic files are not injected at startup.
  • Caps: Startup memory is capped to the first 200 lines and at most 25KB.
  • Command-driven access: memory is the model-visible project memory interface when the product composes @robota-sdk/agent-command-memory. It is exposed through the SDK-projected robota_command_memory tool using the injected command descriptor. The descriptor guides the model to inspect memory when stored context may help, add only durable reusable facts, review pending candidates, report provenance, and avoid storing secrets.
  • Sensitive data policy: Candidate policy must skip obvious secret, token, password, private-key, payment-card, and national-ID style content instead of silently saving it. Additional extractors may be composed later, but they must feed the same policy/store contracts.
  • No hidden turn side effects: InteractiveSession must not automatically prepend topic memory to user prompts and must not create pending memory candidates after a completed turn. Topic retrieval and memory writes happen through explicit /memory command execution, whether user-invoked or model-invoked.
  • Reusable retrieval/capture internals: MemoryRetrievalService, MemoryCandidateExtractor, MemoryPolicyEvaluator, and PendingMemoryStore remain reusable building blocks for explicit commands or future command modules. They are not wired as implicit session lifecycle side effects.
  • Deduplication: ProjectMemoryStore.append() returns deduplicated and must avoid repeating the same normalized topic entry.
  • Command: memory list | show [topic] | add <user|feedback|project|reference> <topic> <text> | pending | approve <id> | reject <id> | used.
  • Audit trail: /memory approve, /memory reject, and future explicit memory workflows append memory events to the session record as memoryEvents for resume/debugging. High-frequency streaming data is not part of the memory event stream.
  • Ownership: SDK owns memory stores, memory policy primitives, and command-facing memory APIs. @robota-sdk/agent-command-memory owns command behavior. CLI only composes the module and renders command results/autocomplete metadata.
  • Prompt composition boundary: The system prompt may include the neutral Project Memory startup index and the /memory descriptor under Built-in Commands; it must not include extra hardcoded memory behavior instructions outside descriptor data.
  • User-local memory boundary: This project memory feature is not baseline user-local memory. User-local display/navigation preferences are governed by ../../../.agents/specs/user-local-memory.md and must not be stored in .robota/memory/.

User-Local Storage

  • Package: agent-framework/user-local/
  • Purpose: Resolve and inspect baseline workflow storage under user-local storage outside the active repository.
  • Default root: ~/.robota.
  • Validation: SDK APIs reject empty roots, relative roots, roots equal to the active repository, and roots inside the active repository, including symlink-resolved paths when possible.
  • Categories: preferences, view-state, memory-projections, task-associations, workflow-metadata, and inspection-index.
  • Inspection projection: SDK returns root, active repository root, category summaries, item summaries, storage locations, enabled/delete/disable metadata, and timestamps when available.
  • Command boundary: @robota-sdk/agent-command-user-local formats provider-free user-local storage list --format json output from SDK projections. agent-cli only routes the direct product command before provider setup and prints the command-owned output.
  • Repository independence: SDK user-local APIs must not create repository .robota/ baseline workflow state.

User-Local Memory

  • Package: agent-framework/user-local/
  • Purpose: Persist explicit display/navigation memory items under the user-local storage root.
  • Storage category: memory-projections.
  • Allowed categories: view-preference, last-visible-cwd, background-selection, task-association, display-preference, and inspection-choice.
  • Projection fields: category, key, summary, value summary, source, scope, storage location, timestamps, enabled state, display/navigation rule, delete/disable availability, and commandExecutionEffect: "none".
  • Mutation APIs: SDK owns set, list, inspect, disable, delete, and enabled-item read behavior.
  • Disabled-item rule: disabled items remain inspectable but readEnabledUserLocalMemoryItem returns null, so they cannot affect display/navigation defaults.
  • Command boundary: @robota-sdk/agent-command-user-local formats provider-free user-local memory ... output from SDK projections. agent-cli only routes the product command and passes terminal options such as --summary, --source, and --format.
  • Repository independence: user-local memory APIs must not write baseline memory inside the active repository or project .robota/.

Context Window Management

  • Token tracking: agent-session Session tracks cumulative input tokens from provider response metadata
  • Usage state: session.getContextState() returns IContextWindowState (usedTokens, maxTokens, usedPercentage)
  • Auto-compaction: Triggers at the configured context-window threshold, defaults to ~83.5%, and can be disabled per session
  • Manual compaction: session.compact(instructions?) generates LLM summary, replaces history
  • Model sizes: Lookup table per model (200K for Sonnet/Haiku, 1M for Opus)
  • Compact Instructions: Extracted from CLAUDE.md "Compact Instructions" section, passed to summary prompt
  • Hooks: PreCompact/PostCompact events in agent-core, fired before/after compaction
  • Callbacks: onCompact in createQuery() options for notification when compaction occurs

Public API

InteractiveSession — Central Client-Facing API

Wraps Session (composition) to provide event-driven interaction for any client (CLI, web, API server, worker). Manages streaming text accumulation, tool execution state tracking, prompt queuing, abort orchestration, and message history. Logic previously embedded in CLI React hooks.

The SDK is pure TypeScript with no React dependency. The CLI is a thin TUI-only layer that subscribes to InteractiveSession events and maps them to React/Ink state.

typescript
import { InteractiveSession } from '@robota-sdk/agent-framework';
import { AnthropicProvider } from '@robota-sdk/agent-provider/anthropic';

// Consumer creates provider and passes it to InteractiveSession.
// Config and context are loaded internally from cwd.
const provider = new AnthropicProvider({ apiKey: process.env.ANTHROPIC_API_KEY });
const session = new InteractiveSession({ cwd: process.cwd(), provider });

// Event-driven — subscribe to state changes
session.on('text_delta', (delta: string) => { /* streaming text chunk */ });
session.on('tool_start', (state: IToolState) => { /* tool execution began */ });
session.on('tool_end', (state: IToolState) => { /* tool execution finished */ });
session.on('thinking', (isThinking: boolean) => { /* execution state changed */ });
session.on('complete', (result: IExecutionResult) => { /* prompt completed */ });
session.on('error', (error: Error) => { /* execution error */ });
session.on('context_update', (state: IContextWindowState) => { /* token usage updated */ });
session.on('interrupted', (result: IExecutionResult) => { /* abort completed */ });
session.on('skill_activation', (event: ISkillActivationEvent) => { /* skill activation state */ });

// Submit prompt. Queues if already executing (max 1 queued).
// displayInput: shown in UI (e.g., "/audit") instead of full built prompt
// rawInput: passed to Session.run() for hook matching
await session.submit(input, displayInput?, rawInput?);

// Execute a named system command. Virtual `/skill-name` entries are normalized by the SDK
// command registry into the composed `skills` command with `<skill-name> [args]`.
const result = await session.executeCommand('context', '');
// result.message — human-readable string
// result.success — boolean
// result.data   — command-specific structured data

// List all registered system commands (for transport adapters)
const commands = session.listCommands(); // Array<{ name, description }>

// Abort current execution and clear queue
session.abort();

// Cancel queued prompt without aborting current execution
session.cancelQueue();

// Graceful shutdown: reject new prompts, abort foreground work, cancel managed background tasks,
// persist final session state, and fire SessionEnd through agent-session.
await session.shutdown({ reason: 'prompt_input_exit', message: 'User requested exit' });

// State queries
session.isExecuting();       // boolean
session.getPendingPrompt();  // string | null
session.getMessages();       // TUniversalMessage[] — backward-compatible; returns chat entries only
session.getFullHistory();    // IHistoryEntry[] — full history including event entries (tool summaries, skill invocations)
session.getContextState();   // IContextWindowState
session.getStreamingText();  // string (accumulated so far)
session.getActiveTools();    // IToolState[]

Self-Hosting Verification Planner

The SDK exports pure planning/state helpers for clients that need to drive a safe edit/build/verify loop without coupling to CLI or TUI rendering.

typescript
import {
  planSelfHostingVerification,
  transitionSelfHostingLoop,
} from '@robota-sdk/agent-framework';

const plan = planSelfHostingVerification({
  changedFiles: ['packages/agent-framework/src/index.ts'],
  packageScopes: ['@robota-sdk/agent-framework'],
  baseRef: 'origin/develop',
});

let state = transitionSelfHostingLoop('idle', 'checkpoint_created');
state = transitionSelfHostingLoop(state, 'edits_started');
state = transitionSelfHostingLoop(state, 'edits_applied');
state = transitionSelfHostingLoop(state, 'verify_passed');

plan.steps is an ordered, provider-neutral command plan. Consumers execute commands in child processes and keep the current SDK process alive as the old runtime. The planner does not write files, restore checkpoints, or render UI.

Task Context Helpers

The SDK exports pure helpers for discovering, selecting, formatting, and updating active task files.

typescript
import { loadTaskContext, updateTaskFileStatus } from '@robota-sdk/agent-framework';

const taskContext = loadTaskContext(process.cwd(), {
  currentBranch: 'feat/context-injection-task-files',
  maxTasks: 3,
});

updateTaskFileStatus('.agents/tasks/CLI-BL-017-context-injection-from-task-files.md', 'completed', {
  progressMessage: 'Verified task context injection.',
});

These helpers operate on Markdown files under .agents/tasks/. They do not render UI and do not inject behavior instructions into the prompt; the formatted task context is neutral metadata.

IToolState:

typescript
interface IToolState {
  toolName: string;
  firstArg: string;
  isRunning: boolean;
  result?: 'success' | 'error' | 'denied';
  diffLines?: IDiffLine[];
  diffFile?: string;
  toolResultData?: string;
}

diffLines is structured Edit tool display metadata. For completed Edit tools, InteractiveSession derives it from the Edit arguments, tool result startLine, and the modified file contents when readable. Diff lines may include type: 'hunk', context, remove, and add. toolResultData is the already-truncated tool result payload emitted by the permission/session layer; transports may derive bounded previews from it, but SDK/session records remain the source for full transcript recovery. The SDK persists this metadata so all transports can replay the same tool summary; CLI owns visual rendering only.

IExecutionResult:

typescript
interface IExecutionResult {
  response: string;
  history: IHistoryEntry[]; // Full history including chat + event entries
  toolSummaries: IToolSummary[];
  contextState: IContextWindowState;
  usage?: IUsageSnapshot;
}

IUsageSnapshot is the SDK-owned provider-neutral execution usage record:

typescript
interface IUsageSnapshot {
  kind: 'exact' | 'estimated';
  scope: 'turn';
  totalTokens: number;
  promptTokens?: number;
  completionTokens?: number;
  contextUsedTokens: number;
  contextMaxTokens: number;
  contextUsedPercentage: number;
  costStatus: 'unknown' | 'estimated' | 'exact';
}

InteractiveSession appends a usage-summary event entry after the assistant response when exact provider usage is available. The entry is persisted in IHistoryEntry[] so /resume, headless transports, and debugging can display usage without reparsing assistant prose.

IInteractiveSessionEvents:

typescript
interface IInteractiveSessionEvents {
  text_delta: (delta: string) => void;
  tool_start: (state: IToolState) => void;
  tool_end: (state: IToolState) => void;
  thinking: (isThinking: boolean) => void;
  complete: (result: IExecutionResult) => void;
  error: (error: Error) => void;
  context_update: (state: IContextWindowState) => void;
  interrupted: (result: IExecutionResult) => void;
  skill_activation: (event: ISkillActivationEvent) => void;
  background_task_event: (event: TBackgroundTaskEvent) => void;
}

ITransportAdapter:

typescript
interface ITransportAdapter {
  /** Human-readable transport name (e.g., 'http', 'ws', 'mcp', 'headless') */
  readonly name: string;

  /** Attach an InteractiveSession to this transport. */
  attach(session: InteractiveSession): void;

  /** Start serving. What this means depends on the transport. */
  start(): Promise<void>;

  /** Stop serving and clean up resources. */
  stop(): Promise<void>;
}

Common interface for all transport adapters. Defined in src/interactive/types.ts and exported from @robota-sdk/agent-framework. Each agent-transport-* package provides a factory that returns an ITransportAdapter implementation.

Background and Subagent Runtime Exports

BackgroundTaskManager is re-exported from agent-executor as the generic runtime registry for long-running work. It owns task IDs, queueing, bounded concurrency, lifecycle events, targeted cancellation, shutdown, terminal close/dismiss, optional send/log controls, watchdogs, and immutable state snapshots.

Runner adapters receive IBackgroundTaskStart.emit(event) for progress reporting. The manager stamps task IDs onto runner events, updates currentAction for tool start/end events, and forwards the resulting TBackgroundTaskEvent to subscribers.

Background task runtime exports:

ExportKindDescription
BackgroundTaskManagerclassGeneric in-memory background task registry and scheduler
BackgroundTaskErrorclassTyped background task error with category and recoverability
createLimitedOutputCapturefunctionRuntime-owned UTF-8-safe bounded output capture helper
appendPrefixedLogLinesfunctionRuntime-owned source-prefixed log line projection helper
createBackgroundTaskLogPagefunctionRuntime-owned cursor-based log pagination helper
IBackgroundTaskManagerinterfaceGeneric manager API for spawn/wait/list/get/cancel/close/shutdown/send
IBackgroundTaskRunnerinterfacePort implemented by agent/process runner adapters
ILimitedOutputCaptureinterfaceRuntime-owned bounded output capture contract
TBackgroundTaskIdFactorytypeRequest-aware task ID factory used by composed managers
IBackgroundTaskStateinterfaceRuntime lifecycle state for one background task
IBackgroundTaskRequesttypeDiscriminated union of agent/process background task requests
IBackgroundTaskResultinterfaceCompleted background task output
TBackgroundTaskEventtypeRuntime-owned lifecycle/progress event union
TBackgroundTaskRunnerEventtypeRunner-owned progress event union without task IDs
TBackgroundTaskModetypeforeground or background
TBackgroundTaskStatustypeShared task lifecycle status union
TBackgroundTaskTimeoutReasontypeWatchdog reason union projected onto failed task state
transitionBackgroundTaskStatusfunctionPure lifecycle transition function
BackgroundJobOrchestratorclassSDK-owned grouping/wait layer above BackgroundTaskManager
IBackgroundJobGroupStateinterfaceParent-session-scoped background task group snapshot
IBackgroundJobGroupSummaryinterfacePresentation-neutral group completion counts and result lines
TBackgroundJobWaitPolicytypedetached, wait_all, wait_any, or manual group completion policy
createExecutionWorkspaceSnapshotfunctionSDK-owned main-thread/task/group workspace projection
createExecutionWorkspaceTaskSpawnerfunctionOrigin-bound task spawning port for commands, skills, and transports
IExecutionWorkspaceEntryinterfacePresentation-neutral selectable execution entry
IExecutionWorkspaceSnapshotinterfaceSession-scoped execution workspace read model
IExecutionWorkspaceTaskSpawnerinterfaceSDK task creation port for agent/process tasks and groups
IExecutionOrigininterfaceSDK-owned task provenance projected from opaque runtime metadata

Background agent watchdog configuration is provider-neutral. Agent requests may set idleTimeoutMs, maxRuntimeMs, outputLimitBytes, maxTextDeltas, repetitionWindow, and repetitionThreshold; the runtime refreshes lastActivityAt from runner progress events and fails runaway jobs with timeoutReason.

InteractiveSession subscribes to background task events, persists every event including streaming text deltas into the session record for local debugging/resume, and emits background_task_event for transports and TUI state projection. It also maps background agent lifecycle events into Claude Code-compatible SubagentStart and SubagentStop hooks.

BackgroundJobOrchestrator is the SDK-owned layer above BackgroundTaskManager for parent-request orchestration. It groups related task IDs, applies a wait policy, emits group lifecycle events, and produces result envelopes with task IDs, labels, terminal status, concise output summaries, output references, and errors. It also exposes presentation-neutral summary helpers for command/transport/UI adapters. The orchestrator does not run processes, own provider calls, mutate TUI state, or inject hardcoded prompt instructions.

InteractiveSession exposes background job group controls:

APIBehavior
createBackgroundJobGroup(request)Create a parent-session-scoped group over existing task IDs
listBackgroundJobGroups()Return cloned group snapshots
getBackgroundJobGroup(groupId)Return one cloned group snapshot
waitBackgroundJobGroup(groupId)Resolve when the group's wait policy reaches a terminal result
summarizeBackgroundJobGroup(group)Return counts and concise result lines for an existing group

InteractiveSession emits background_job_group_event with TBackgroundJobGroupEvent. When session persistence is enabled, group snapshots and group events are stored alongside background task snapshots/events so resume/debugging can reconstruct group provenance.

SubagentManager and its associated types are exported for clients that need to compose managed subagent execution. It is now a compatibility facade over BackgroundTaskManager for kind: 'agent' tasks, preserving the existing subagent API while moving lifecycle semantics to the shared background layer.

typescript
import { SubagentManager } from '@robota-sdk/agent-framework';
import type { ISubagentRunner } from '@robota-sdk/agent-framework';

const runner: ISubagentRunner = createRunner();
const manager = new SubagentManager({ runner, maxConcurrent: 2 });

const job = await manager.spawn({
  type: 'general-purpose',
  label: 'General purpose',
  parentSessionId: 'session_parent',
  mode: 'foreground',
  depth: 1,
  cwd: process.cwd(),
  prompt: 'Review the codebase',
});

const result = await manager.wait(job.id);

Agent subagent requests may set isolation: 'worktree'. The SDK treats this as a contract flag and propagates it through agent command arguments, ISubagentSpawnRequest, and background task metadata. Worktree isolation is explicit unless a host assembly provides and documents a capability-aware default policy; SDK core must not silently infer or fallback between isolated and non-isolated execution. agent-executor owns WorktreeSubagentRunner, which decorates any ISubagentRunner with worktree lifecycle, metadata, cleanup, and hook behavior. Runtime shells provide an ISubagentWorktreeAdapter implementation for concrete local Git/filesystem operations. If a preserved worktree is returned by a runner, IBackgroundTaskResult.metadata.worktreePath, branchName, worktreeStatus, worktreeNextAction, worktreeBaseRevision, and parentWorktreeStatus are projected onto matching IBackgroundTaskState fields.

createBackgroundProcessTool(deps) is exported for SDK composition. The tool is registered only when a runtime shell injects a process background runner through createSession({ backgroundTaskRunners }); default Bash foreground behavior remains unchanged.

createSession() also accepts subagentRunnerFactory?: TSubagentRunnerFactory. When omitted, SDK composition uses createInProcessSubagentRunner. Runtime shells such as agent-cli may inject a factory that receives the same assembled dependency bundle and returns a process-backed ISubagentRunner.

Exported subagent runtime types:

ExportKindDescription
SubagentManagerclassRe-export from agent-executor; in-memory subagent job facade
createInProcessSubagentRunnerfunctionRunner adapter that executes subagent jobs with createSubagentSession()
WorktreeSubagentRunnerclassRe-export from agent-executor; worktree isolation runner decorator
createWorktreeSubagentRunnerfunctionFactory for WorktreeSubagentRunner
createDefaultToolsfunctionDefault tool assembly helper exported for CLI fork-worker composition
ISubagentManagerinterfaceRe-export from agent-executor; manager API
ISubagentRunnerinterfaceRe-export from agent-executor; single-job runner port
ISubagentWorktreeAdapterinterfaceRe-export from agent-executor; concrete worktree I/O port
IPreparedSubagentWorktreeinterfaceRe-export from agent-executor; prepared worktree handoff
IInProcessSubagentRunnerDepsinterfaceDependencies captured by the in-process runner adapter
TSubagentRunnerFactorytypeFactory seam for runtime shells to replace the default subagent runner
ISubagentJobHandleinterfaceRe-export from agent-executor; targeted job handle
ISubagentJobStateinterfaceRe-export from agent-executor; subagent job projection
ISubagentSpawnRequestinterfaceRe-export from agent-executor; spawn request
ISubagentJobResultinterfaceRe-export from agent-executor; completion output and metadata
TSubagentJobModetypeRe-export from agent-executor; foreground or background
TSubagentJobStatustypeRe-export from agent-executor; lifecycle status union

History Entry Types

InteractiveSession manages history as IHistoryEntry[]. Each entry has a category field:

CategoryDescription
'chat'A standard conversation message (TUniversalMessage). Returned by getMessages() as-is.
'event'A structured non-message event (tool summary, skill invocation, system notification, etc.).

Tool summary entry (appended by InteractiveSession after each execution round):

typescript
// category: 'event', type: 'tool-summary'
{
  id: string;
  timestamp: Date;
  category: 'event';
  type: 'tool-summary';
  data: {
    summary: string;
    tools: Array<{
      toolName: string;
      firstArg: string;
      isRunning: boolean;
      result?: 'success' | 'error' | 'denied';
      diffLines?: IDiffLine[];
      diffFile?: string;
      toolResultData?: string;
    }>;
  }
}

Usage summary entry (appended by InteractiveSession after each completed turn when usage exists):

typescript
{
  category: 'event',
  type: 'usage-summary',
  data: {
    kind: 'exact',
    scope: 'turn',
    promptTokens: 1000,
    completionTokens: 200,
    totalTokens: 1200,
    contextUsedTokens: 1200,
    contextMaxTokens: 200000,
    contextUsedPercentage: 0.6,
    costStatus: 'unknown',
  }
}

Skill activation entry (appended by InteractiveSession when a real skill activation starts):

typescript
// category: 'event', type: 'skill-activation'
{
  id: string;
  timestamp: Date;
  category: 'event';
  type: 'skill-activation';
  data: {
    skillName: string;
    source: 'skill' | 'plugin';
    invocation: 'user-slash' | 'model-tool';
    mode: 'inject' | 'fork';
    status: 'started' | 'completed' | 'failed';
    message: string;
    qualifiedName?: string;
    error?: string;
  }
}

Legacy skill-invocation entries may still be rendered when resuming older sessions, but new SDK execution records use skill-activation.

Consumers that need only AI messages call getMessages() (returns TUniversalMessage[] — backward-compatible). Consumers that need the full picture (e.g., rendering a rich message list) call getFullHistory() (returns IHistoryEntry[]).

System Commands — Embedded in InteractiveSession

SystemCommandExecutor is embedded inside InteractiveSession. Consumers access system commands via session.executeCommand(name, args). Command module packages may import ISystemCommand, ICommandModule, SystemCommandExecutor, and createSystemCommands() for composition tests.

The command types and result interface are exported for consumers that need to inspect results:

typescript
import type { ICommandResult, ISystemCommand } from '@robota-sdk/agent-framework';

// Execute a named command on the session (returns null if command not found)
const result: ICommandResult | null = await session.executeCommand('context', '');
// result.message — human-readable string
// result.success — boolean
// result.data   — command-specific structured data

Product-composed command modules:

CommandDescription
helpCommand module for rendering registered commands
clearOptional command module for clearing conversation and rendered host history
compactCompress context window (optional focus instructions)
languageRequest response language update through language-change-requested effect
costOptional session command module for session ID and message count
contextToken usage: used / max / percentage
permissions [mode]Current mode, session-approved tools, and permission mode changes
statuslineOptional command module for statusline visibility and git branch patch flows
memoryList/show/add/review project memory and report used memory references
rewindList edit checkpoints, restore later edits, or rollback through a checkpoint
resetRequests settings reset through settings-reset-requested effect
resumeOptional command module for requesting session picker through effect
renameOptional command module for requesting session rename through effect
providerOptional command module for provider current/list/use/add/test flows

ISystemCommand:

typescript
interface ISystemCommand {
  name: string;
  description: string;
  modelInvocable?: boolean;
  userInvocable?: boolean;
  argumentHint?: string;
  safety?: TCapabilitySafety;
  subcommands?: readonly ICommand[];
  lifecycle?: 'inline' | 'blocking' | 'background';
  execute(context: ICommandHostContext, args: string): Promise<ICommandResult> | ICommandResult;
}

ICommandHostContext is the command-facing facade supplied by the SDK executor. Command implementations must depend on the specific context methods or typed host adapters they need rather than accepting InteractiveSession, CLI state, or UI hooks.

ICommandHostAdapters is the host-provided adapter bag exposed through ICommandHostContext.getCommandHostAdapters(). It currently includes settings persistence, process lifecycle, permission-mode access, and plugin-management adapters. Command modules may request only the adapter they need; they must not import a concrete CLI/TUI implementation.

ICommandModule:

typescript
interface ICommandModule {
  name: string;
  commandSources?: readonly ICommandSource[];
  systemCommands?: readonly ISystemCommand[];
  commandDescriptors?: readonly ICapabilityDescriptor[];
  sessionRequirements?: readonly TCommandModuleSessionRequirement[];
}

sessionRequirements is how command modules request optional SDK wiring. The current requirement is agent-executor, which enables agent definitions and the shared background/subagent managers for command-owned agent execution.

ICommandResult:

typescript
interface ICommandResult {
  message: string;
  success: boolean;
  data?: Record<string, TCommandResultDataValue>;
  effects?: readonly TCommandEffect[];
  interaction?: ICommandInteraction;
}

type TCommandEffect =
  | { type: 'model-change-requested'; modelId: string }
  | { type: 'language-change-requested'; language: string }
  | { type: 'settings-reset-requested' }
  | { type: 'session-exit-requested'; reason?: TSessionEndReason; message?: string }
  | { type: 'session-restart-requested'; reason: TSessionEndReason; message: string }
  | { type: 'plugin-tui-requested' }
  | { type: 'plugin-registry-reload-requested' }
  | { type: 'session-picker-requested' }
  | { type: 'session-renamed'; name: string }
  | { type: 'conversation-history-cleared' }
  | { type: 'statusline-settings-patch'; patch: TStatusLineCommandSettingsPatch };

interface ICommandInteraction {
  prompt: ICommandInteractionPrompt;
  submit(value: string): Promise<ICommandResult> | ICommandResult;
  cancel?(): Promise<ICommandResult> | ICommandResult;
}

ICommandInteractionPrompt is the generic prompt descriptor used by UI hosts. It supports choice and text prompts with optional description text, masked text, and validation metadata. Hosts render the prompt and pass submitted values back to the interaction; they do not inspect command-specific state.

CommandRegistry, BuiltinCommandSource, SkillCommandSource, PluginCommandSource

Command discovery and aggregation for clients that expose a slash command palette or autocomplete UI. Owned by agent-framework; agent-cli re-exports CommandRegistry from here. PluginCommandSource was moved from agent-cli to agent-framework so all clients benefit from plugin command discovery. Command modules can be added through registry.addModule(module) without the registry knowing their command names. Hosts can call registry.replaceSource(name, source) to refresh dynamic sources such as plugin-provided commands after a successful reload effect.

typescript
import {
  CommandRegistry,
  SkillCommandSource,
  PluginCommandSource,
} from '@robota-sdk/agent-framework';

const registry = new CommandRegistry();
registry.addModule(commandModule);
registry.addSource(new SkillCommandSource(process.cwd()));

registry.getCommands(); // ICommand[] — all composed commands and virtual entries
registry.getCommands('mod'); // filtered by prefix (for autocomplete)
registry.resolveQualifiedName('audit'); // "my-plugin:audit" or null
registry.getSubcommands('mode'); // ICommand[] — subcommands

BuiltinCommandSource remains exported as an empty SDK-core compatibility source. Product command entries come from composed ICommandModule values such as @robota-sdk/agent-command-skills.

SkillCommandSource scans (highest priority first):

  1. <cwd>/.claude/skills/*/SKILL.md
  2. <cwd>/.claude/commands/*.md (Claude Code compatible)
  3. ~/.robota/skills/*/SKILL.md
  4. <cwd>/.agents/skills/*/SKILL.md

createQuery() — Convenience Factory

createQuery({ provider }) is a factory that returns a prompt-only function. The caller creates the provider; the factory captures it and returns a simple async function that accepts a prompt string.

typescript
import { createQuery } from '@robota-sdk/agent-framework';
import { AnthropicProvider } from '@robota-sdk/agent-provider/anthropic';

const provider = new AnthropicProvider({ apiKey: process.env.ANTHROPIC_API_KEY });
const query = createQuery({ provider });

const response = await query('Show me the file list');

const queryWithOptions = createQuery({
  provider,
  cwd: '/path/to/project',
  permissionMode: 'acceptEdits',
  maxTurns: 10,
  onTextDelta: (delta) => process.stdout.write(delta),
});

const detailedResponse = await queryWithOptions('Analyze the code');

createSession() is an internal assembly factory — it is not exported from @robota-sdk/agent-framework. Config and context loading, tool assembly, and provider wiring happen inside InteractiveSession and createQuery().

Session — Direct Usage (Generic)

typescript
import { Session } from '@robota-sdk/agent-session';

// Session requires pre-constructed tools, provider, and systemMessage
const session = new Session({ tools, provider, systemMessage, terminal });
const response = await session.run('Hello');

Public Surface Ownership

The top-level @robota-sdk/agent-framework entrypoint exposes SDK-owned APIs and explicit SDK facades. It must not pass through general-purpose agent-core, agent-session, or agent-tools exports only for convenience. See PUBLIC-SURFACE.md for the export classification.

Allowed public classes:

  • SDK-owned APIs: InteractiveSession, createQuery, command contracts/common APIs, project memory, checkpoints, reversible execution, plugin management, and task context helpers.
  • SDK facades: project session store helpers, subagent assembly helpers, agent/background process tools, and command host/common APIs that narrow lower-level behavior through SDK contracts.
  • Explicit runtime facades: background-task and subagent lifecycle contracts re-exported through src/background-tasks/index.ts and src/subagents/index.ts.

Owner-direct APIs:

  • agent-core owns history helpers, provider interfaces, permissions, hooks, context window types, and generic message utilities.
  • agent-tools owns direct built-in tool exports and tool result types.
  • agent-session owns generic session APIs and terminal output primitives.

pnpm harness:scan:sdk-public-surface prevents broad export * barrels, top-level lower-owner pass-through exports, and runtime re-exports outside the documented SDK facade barrels.

History Types — Owner Package

typescript
import {
  IHistoryEntry,
  isChatEntry,
  chatEntryToMessage,
  messageToHistoryEntry,
  getMessagesForAPI,
} from '@robota-sdk/agent-core';
ExportKindDescription
IHistoryEntryinterfaceRich history entry: id, timestamp, category ('chat' | 'event'), type, data
isChatEntryfunctionType guard that narrows IHistoryEntry to chat entries
chatEntryToMessagefunctionConverts a chat IHistoryEntry to TUniversalMessage
messageToHistoryEntryfunctionConverts a TUniversalMessage to a chat IHistoryEntry
getMessagesForAPIfunctionExtracts TUniversalMessage[] from IHistoryEntry[] (filters to chat entries only)

Built-in Tools — Direct Usage

@robota-sdk/agent-framework assembles built-in tools internally for SDK sessions. Direct tool usage imports from @robota-sdk/agent-tools:

typescript
import {
  bashTool,
  editTool,
  globTool,
  grepTool,
  readTool,
  webFetchTool,
  webSearchTool,
  writeTool,
} from '@robota-sdk/agent-tools';

Permissions — Direct Usage

typescript
import { evaluatePermission } from '@robota-sdk/agent-core';

promptForApproval is exported from agent-framework for CLI and transport adapters that implement a non-TUI permission flow:

ExportKindDescription
promptForApprovalfunctionPrompts the user for allow/deny approval before a tool runs using ITerminalOutput.select()

Skill Prompt Utilities

substituteVariables and preprocessShellCommands are pure helpers for skill prompt processing:

ExportKindDescription
substituteVariablesfunctionSubstitutes $VAR / ${VAR} placeholders in a skill prompt string from a context
preprocessShellCommandsfunctionExtracts shell commands embedded in skill prompt text for pre-execution
SkillPromptContexttypeVariable substitution context shape for substituteVariables

Path Helpers

projectPaths and userPaths are SDK-owned path helpers for project-local and user-local file resolution:

ExportKindDescription
projectPathsfunctionReturns structured project-local paths under .robota/ for a given cwd
userPathsfunctionReturns structured user-local paths under ~/.robota/ (settings, sessions, memory)

These helpers are used by SDK assembly and command modules. Transparent workflow baseline storage must not use projectPaths(cwd) or ad hoc .robota/ paths — use the user-local storage root resolver instead.

Import Rules

These rules define which packages each layer is allowed to import from. Violations break the layered architecture.

CLI (agent-cli)

SourceAllowedNotes
agent-frameworkAll SDK-owned public APIsInteractiveSession, createQuery, runtime contracts re-exported by SDK
agent-executor❌ Direct import discouragedCLI should receive runtime ports through SDK composition/re-exports
agent-corePublic types + utilities onlyTUniversalMessage, TPermissionMode, createSystemMessage, getModelName
agent-core❌ Internal engine classesRobota, ExecutionService, ConversationStore are forbidden
agent-session❌ ForbiddenSDK provides its own session types; CLI must not import sessions directly
agent-tools❌ ForbiddenSDK assembles tools internally
agent-provider-*Provider creation onlyAnthropicProvider, GeminiProvider (CLI picks which to use)

SDK (agent-framework)

SourceAllowedNotes
agent-coreFull access
agent-executorFull accessBackground task/subagent lifecycle primitives
agent-sessionFull access
agent-toolsFull access
agent-provider-*❌ ForbiddenSDK is provider-neutral; provider comes from consumer

Transport packages (agent-transport-*)

SourceAllowedNotes
agent-frameworkInteractiveSession and related types
agent-corePublic types only (TUniversalMessage etc.)

Design Decision Records

Claude Code vs Claude Agent SDK Relationship (Research)

  • Claude Agent SDK extracts the Claude Code runtime (running the CLI as a subprocess)
  • Robota adopts a direct code sharing approach rather than subprocess
  • Layer hierarchy: agent-cli → agent-framework → agent-session → agent-core (upper layers import lower layers)
  • Research document: docs/superpowers/research/2026-03-19-claude-code-vs-agent-sdk.md

General/Specialized Separation Criteria

Each module's placement is determined by "Is this used only in the SDK, or is it general-purpose?":

ModuleVerdictRationale
PermissionsGeneral → agent-coreTool permission checks are needed on servers too
HooksGeneral → agent-coreAudit/validation is needed on servers too
Built-in toolsGeneral → agent-toolsFile system tools are needed in playground/server environments too
SessionGeneral → agent-sessionSession management is needed in any environment
Config loadingSDK-specific → agent-framework.robota/settings.json is for local environments only
Context loadingSDK-specific → agent-frameworkAGENTS.md walk-up is for local environments only
Agent runtime depsSDK-specific → agent-frameworkSub-session creation dependencies are assembled by SDK and consumed through command/runtime APIs
InteractiveSessionSDK-specific → agent-frameworkClient-facing event wrapper; no CLI/React dependency; reusable by all clients
SystemCommandExecutorSDK-specific → agent-frameworkEmbedded in InteractiveSession; accessed via session.executeCommand(); exported for command module composition tests
CommandRegistry et al.SDK-specific → agent-frameworkSlash command discovery is useful for any client; moved from CLI to SDK
ITerminalOutputGeneral → agent-sessionTerminal I/O abstraction (SSOT in permission-enforcer.ts; agent-cli has a duplicate)

Existing Package Refactoring History

  • agent-session: Removed existing SessionManager/ChatInstance (zero consumers, no-op persistence), replaced with Session/SessionStore from agent-framework
  • agent-tools: Added 8 built-in tools in builtins/ directory (Bash, Read, Write, Edit, Glob, Grep, WebFetch, WebSearch), added TToolResult type
  • agent-core: Added permissions/ and hooks/ directories
  • agent-provider-anthropic: Multi-block content handling (text + tool_use), streaming chatWithStreaming, onTextDelta support

Hook Type Executors (SDK-Specific)

agent-framework provides two additional IHookTypeExecutor implementations that extend the hook system beyond agent-core's built-in command and http executors:

ExecutorHook TypeDescription
PromptExecutorpromptInjects the hook's prompt text into the session context as a system-level instruction
AgentExecutoragentCreates a nested agent session (via createSession) to process the hook input and return a result

These executors are registered with runHooks via the executors map during session creation in createSession().

Settings Configuration

Settings are loaded with a 6-file precedence model (lowest priority first). .robota/ is the primary configuration convention; .claude/ paths are supported for Claude Code compatibility.

LayerPathScope
1~/.robota/settings.jsonUser global
2~/.claude/settings.jsonUser global (Claude Code compatible)
3.robota/settings.jsonProject
4.robota/settings.local.jsonProject (local)
5.claude/settings.jsonProject (Claude Code compatible)
6.claude/settings.local.jsonProject (local, Claude Code compatible)

The .claude/settings.json layers provide Claude Code compatibility — settings written by Claude Code are automatically picked up by Robota. Higher layers override lower layers via deep merge. $ENV:VAR substitution is applied after merge for provider API keys.

Provider resolution order:

  1. currentProvider plus providers[currentProvider]
  2. Legacy provider
  3. Existing defaults

Provider profile schema:

FieldDescription
typeProvider implementation type such as anthropic or openai
modelDefault model ID for the profile
apiKeyLiteral key or $ENV:<name> reference
baseURLOptional OpenAI-compatible or provider-specific endpoint
timeoutOptional provider idle timeout and provider construction timeout when supported

currentProvider must point to an existing profile key. Missing profiles and profiles without type are configuration errors. Profile keys are stable user-facing identifiers; two profiles may have the same type and model when they represent different credentials, accounts, endpoints, or operational defaults. Legacy provider remains accepted for backward compatibility, but it must not override an explicit active provider profile.

The SDK remains provider-neutral: it resolves provider metadata for session assembly, but consumers such as agent-cli still construct concrete provider instances. During session assembly, config.provider.timeout is forwarded to Session.providerTimeout; when omitted, SDK assembly uses a 120-second provider idle timeout so headless/TUI sessions cannot wait forever for a stalled provider call.

Bundle Plugin System

Bundle plugins package reusable extensions (tools, hooks, permissions, system prompt additions) into installable units.

Types

TypeDescription
IBundlePluginManifestPlugin metadata: name, version, description, author, keywords
ILoadedBundlePluginFull bundle: manifest + tools, hooks, permissions, systemPrompt

Loader

BundlePluginLoader loads a bundle plugin from a directory path. It reads the manifest, resolves tool/hook definitions, and validates the bundle structure.

Installer

BundlePluginInstaller manages plugin installation and uninstallation:

  • Installs bundles to ~/.robota/plugins/ (user) or .robota/plugins/ (project)
  • Tracks installed plugins in a registry file
  • Handles enable/disable state per plugin

Marketplace Client

MarketplaceClient provides plugin discovery and installation from remote sources.

  • Source management: Add, remove, and list marketplace sources
  • Default marketplace: Built-in default source URL for the Robota plugin marketplace
  • Search: Query available plugins by name, keyword, or category
  • Install: Download and install plugins via BundlePluginInstaller

System Prompt Skill and Agent Injection

Skills discovered from skill directories are exposed to the system prompt by metadata only when the session has a composed model-invocable skills command descriptor. The metadata includes name and description only. The ## Skills section owns model-visible skill selection metadata and must not include extra hardcoded behavior instructions. skills is owned by @robota-sdk/agent-command-skills as a normal built-in command module. Full SKILL.md content is loaded only when the composed skills command calls SDK skill activation through ICommandHostContext.executeSkillCommandByName(). Skills with disable-model-invocation: true are omitted from model-visible metadata and rejected for model-sourced skills activation.

When at least one model-invocable command exists, createSession() projects each descriptor into a provider-safe tool named robota_command_<command>. The projection layer keeps a reverse map from provider-visible tool name to slash-free command id, validates collisions before session assembly, and routes execution through the same ISystemCommand handler used by user-entered slash commands. skills uses the projected robota_command_skills route with args: "<skill-name> [args]". createSession() must not register ExecuteSkill or any parallel direct skill model tool. A model mentioning or recommending a skill in ordinary prose is not a skill activation.

For user prompts, InteractiveSession.submit() does not parse natural language for skill names or activation phrases. Natural-language skill selection belongs to the model-facing skills descriptor and the projected robota_command_skills tool route. Explicit slash input such as /audit src/index.ts is a virtual command alias normalized by executeCommand() into the composed skills command with args audit src/index.ts.

Projected command tool names must match provider naming constraints (^[A-Za-z0-9_-]{1,64}$) and use the robota_command_ namespace. Their provider-visible descriptions come from registered command descriptors so command owners, not the system prompt composer, own autonomous-use guidance. createSession() must not register projected command tools when no registered command descriptor is model-invocable. The legacy createCommandExecutionTool() helper remains exported for compatibility, but createSession() does not expose both routes for the same command behavior.

Selection must not be implemented with local keyword matching, alias tables, or natural-language pre-routing inside InteractiveSession.

Agent definitions are exposed to the system prompt by metadata only when an injected command module requests agent-executor. Without that session requirement, agent runtime dependencies, agent definitions, and model-visible agent metadata are omitted.

Agent execution is routed through command/runtime APIs such as agent and through context: fork skill execution. createSession() stores reusable agent runtime dependencies for those paths but does not register a separate model-visible Agent tool.

Skill Execution Semantics

InteractiveSession.executeCommand(name, args) is the only transport-facing slash execution path. When name is a virtual skill name and a skills command module is composed, the SDK normalizes the request to command skills with args <skill-name> [args]. TUI and headless transports must not call skill-specific execution methods.

InteractiveSession.executeSkillCommandByName(name, args, request) is the SDK host API consumed by the skills command module. It resolves the named skill from SDK-owned skill sources, validates the invocation source, loads the full SKILL.md, emits skill_activation, and returns structured command results/effects. Model-sourced calls return processed skill instructions as command result data; user-sourced calls submit the rendered prompt or fork execution into the active session and emit session-execution-started.

Skill metadataBehavior
no contextRender skill content and submit it into the current session
context: forkRun rendered skill content in an isolated subagent session using skill.agent or general-purpose
allowed-toolsRestrict fork-session tools to the listed names, after the selected agent definition denylist applies
disable-model-invocationHide from model-visible skill metadata; user slash invocation still works
user-invocable: falseHide from user slash menus; model metadata remains available unless model invocation is disabled

Fork skill execution must not rely on prompting the parent model to call the Agent tool. It must call createSubagentSession() directly through the per-session agent tool dependencies so the behavior is deterministic and unit-testable.

Every activation records an ISkillActivationEvent:

typescript
interface ISkillActivationEvent {
  readonly type: 'skill-activation';
  readonly skillName: string;
  readonly source: 'skill' | 'plugin';
  readonly invocation: 'user-slash' | 'model-tool';
  readonly mode: 'inject' | 'fork';
  readonly status: 'started' | 'completed' | 'failed';
  readonly timestamp: string;
  readonly qualifiedName?: string;
  readonly error?: string;
}

InteractiveSession stores skill activation events in skillActivationEvents when session persistence is enabled. The event list is restored with the session record and the started event is also represented in IHistoryEntry[] for UI rendering. Consumers must not report a skill as active unless this event exists.

Hook Wiring into Session Lifecycle

During createSession(), hooks from the merged settings configuration are wired into the session lifecycle:

  1. Hook configuration is extracted from the resolved config
  2. SDK-specific executors (PromptExecutor, AgentExecutor) are registered alongside core executors
  3. SessionStart hooks fire during session initialization
  4. PreToolUse/PostToolUse hooks are invoked by PermissionEnforcer around tool execution
  5. UserPromptSubmit hooks fire before each user message is processed
  6. Stop hooks fire on session termination

Background Task Execution

BackgroundTaskManager is owned by agent-executor and re-exported by agent-framework through the explicit runtime facade. It is the generic lifecycle layer for foreground/background agent and process jobs. It is provider-neutral and depends only on injected runner ports.

Responsibilities:

  • create addressable background task records
  • enforce bounded concurrency across registered task kinds
  • track lifecycle state: queued, running, waiting_permission, completed, failed, cancelled
  • expose spawn, wait, list, get, cancel, close, send, readLog, and subscribe
  • emit a single TBackgroundTaskEvent union for lifecycle/progress projection
  • keep runner implementation details out of TUI, transports, and tool code

The manager does not create providers, sessions, child processes, worktrees, or TUI state directly. Those concerns belong to runner adapters and outer composition layers. SDK code composes the manager with SDK-owned tools and InteractiveSession; it does not own the lifecycle state machine.

SDK runtime facade barrels also re-export runtime-owned helper primitives for bounded output capture and cursor-based log pagination so runtime shells can implement process adapters through the documented SDK facade instead of importing agent-executor directly.

InteractiveSession exposes background task controls:

MethodBehavior
listBackgroundTasks(filter?)Return cloned background task state snapshots
getBackgroundTask(taskId)Return one cloned task snapshot
cancelBackgroundTask(...)Targeted task cancellation
closeBackgroundTask(taskId)Remove a terminal task from the registry
sendBackgroundTask(...)Forward optional input to a supporting runner
readBackgroundTaskLog(...)Read optional runner logs

InteractiveSession emits background_task_event with TBackgroundTaskEvent.

InteractiveSession also exposes an SDK-owned execution workspace read model for clients that need to switch between the main conversation, background tasks, and background groups without owning lifecycle state:

MethodBehavior
getExecutionWorkspaceSnapshot()Return a presentation-neutral snapshot with the main-thread entry first
listExecutionWorkspaceEntries(filter?)Return selectable main-thread/task/group entries
getExecutionWorkspaceEntry(entryId)Return one execution workspace entry
readExecutionWorkspaceDetail(...)Return a normalized detail page for main transcript, task log, or group
createExecutionWorkspaceTaskSpawner(...)Return an origin-bound SDK task spawning port for commands/skills/hosts

The read model is the only shared contract for task-switching surfaces. agent-cli and transports may render entries, keep ephemeral selection state, and invoke explicit controls, but they must not infer lifecycle, retention, origin, unread/attention semantics, or control availability from raw events when this projection is available.

The cross-client background work state contract is defined in ../../../.agents/specs/background-work-state.md. The current IExecutionWorkspaceEntry shape covers stable ids, entry kind, origin, status, labels, preview, current action, attention, visibility, updated time, and advisory controls. Future fields such as started time, elapsed time, input-needed reason, terminal result, retention state, archive, and clear controls must be added to the SDK projection before CLI or transport surfaces render them.

Execution workspace entries use a common IExecutionWorkspaceEntry shape:

  • main_thread is an SDK projection backed by InteractiveSession history and current foreground execution state. It is not a BackgroundTaskManager record.
  • background_task entries are projections of IBackgroundTaskState.
  • background_group entries are projections of BackgroundJobOrchestrator groups.
  • origin is SDK-owned provenance. Runtime stores only opaque primitive metadata; the SDK maps it into IExecutionOrigin for commands, model commands, tool calls, skills, transports, and system work. This is presentation provenance; command execution eligibility for transparent workflow features must follow the action provenance contract.
  • controls is presentation-neutral and advisory. Selecting an entry is never a lifecycle mutation; cancellation, close, send, read, wait, and group summary remain explicit APIs.

Default visibility keeps active, permission-blocked, failed, cancelled, and unread-completed tasks in the workspace list. Clean completed tasks remain queryable through runtime state until close() or session cleanup, but clients may choose a collapsed recent/history presentation from the SDK entry metadata instead of deleting records.

The workspace state vocabulary follows the transparent workflow contract. Current runtime waiting_permission snapshots must be projected for clients as user-facing waiting-for-input state when the surface is not exposing raw runtime types for debugging.

When session persistence is enabled, InteractiveSession must persist background task state as part of the project-local session record. Lifecycle, tool start/end, permission, completion, failure, cancellation, and close events update the session JSON with the latest task snapshots and durable event summaries. High-frequency background_task_text_delta events must not rewrite the main session JSON per chunk; they are written to append-only JSONL session logs and task/subagent transcript files so debugging data is available while streaming is still in progress without risking partial JSON writes.

createSession() accepts backgroundTaskRunners?: IBackgroundTaskRunner[]. When a runner with kind: 'process' is present, SDK composition registers the model-callable BackgroundProcess tool:

  • BackgroundProcess starts a command as kind: 'process', mode: 'background'
  • it returns { success, background: true, output: '', taskId, status, command } immediately
  • stdout/stderr inspection and cancellation are routed through the shared manager APIs
  • existing Bash tool behavior is not changed

createSession() accepts subagentRunnerFactory?: TSubagentRunnerFactory. The SDK default remains createInProcessSubagentRunner(agentToolDeps). A runtime shell may supply a factory to run agent command jobs through a process-backed runner while reusing the same config/context/tool dependency bundle assembled by the SDK.

Runner progress semantics:

  • background_task_text_delta forwards partial output for preview surfaces
  • background_task_tool_start sets IBackgroundTaskState.currentAction
  • background_task_tool_end clears currentAction on success or stores the error/action on failure
  • progress events do not complete, fail, cancel, or close tasks; lifecycle remains manager-owned
  • progress and lifecycle events are diagnostic data, not just UI state; SDK composition must route them to session logging/persistence when those facilities are configured

The product-composed /background command module maps to these APIs:

CommandBehavior
/background or /background listList current background tasks
/background read <task-id> [offset]Read a task log page
/background cancel <task-id>Cancel one running/queued task
/background close <task-id>Dismiss one terminal task

Subagent Execution

SubagentManager

SubagentManager is owned by agent-executor and re-exported by agent-framework through the explicit runtime facade. It is the managed subagent facade. It depends on an injected ISubagentRunner port or an injected IBackgroundTaskManager and maps subagent jobs to BackgroundTaskManager agent tasks.

Responsibilities:

  • create addressable subagent job records
  • enforce bounded concurrency
  • track lifecycle state: queued, running, waiting_permission, completed, failed, cancelled
  • expose spawn, wait, list, get, cancel, close, and send operations
  • keep runner implementation details out of TUI and command-module code

SubagentManager does not create providers, sessions, child processes, worktrees, or TUI state directly. Those concerns belong to runner adapters and outer composition layers. It exposes getBackgroundTaskManager() so SDK InteractiveSession can forward generic background task events and controls without depending on subagent-specific types.

SubagentRunner Port

ISubagentRunner is owned by agent-executor and is the execution boundary for one subagent job. Implementations can run jobs in-process for tests or in a child process for CLI runtime.

typescript
interface ISubagentRunner {
  start(job: ISubagentJobStart): ISubagentJobHandle;
}

interface ISubagentJobStart {
  jobId: string;
  request: ISubagentSpawnRequest;
  emit?: (event: TBackgroundTaskRunnerEvent) => void;
}

interface ISubagentJobHandle {
  readonly jobId: string;
  readonly pid?: number;
  readonly logPath?: string;
  readonly transcriptPath?: string;
  result: Promise<ISubagentJobResult>;
  cancel(reason?: string): Promise<void>;
  send?(prompt: string): Promise<void>;
  readLog?(cursor?: IBackgroundTaskLogCursor): Promise<IBackgroundTaskLogPage>;
}

The runner reports completion through its result promise and supports targeted cancellation through cancel(). Follow-up routing via send() is optional until a runner supports it. Log reading via readLog() is optional, but process-backed subagent runners should implement it so /agent read AGENT_ID can inspect append-only transcripts while a job is still running.

createInProcessSubagentRunner(deps) is the default SDK adapter for foreground compatibility. It resolves the requested agent definition, creates an isolated child Session with createSubagentSession(), runs the prompt, and maps the response to ISubagentJobResult.

WorktreeSubagentRunner

WorktreeSubagentRunner is owned by agent-executor. It keeps worktree isolation behavior reusable across CLI, headless, or future runtime shells while keeping concrete Git commands outside the reusable runtime layer.

The decorator depends on:

  • an inner ISubagentRunner that performs the actual agent execution
  • an ISubagentWorktreeAdapter port that can prepare, inspect, and remove worktrees
  • optional THooksConfig and hook executors for worktree lifecycle notifications

When job.request.isolation !== 'worktree', the decorator delegates to the inner runner without changing the request.

When job.request.isolation === 'worktree', the decorator must:

  • call ISubagentWorktreeAdapter.prepare({ jobId, cwd })
  • invoke the inner runner with cwd, worktreePath, and branchName set to the prepared worktree
  • emit WorktreeCreate hook notification after preparation
  • remove clean worktrees exactly once on success, delegated failure, synchronous delegated start failure, or successful cancellation
  • preserve dirty worktrees and return worktreePath, branchName, worktreeStatus, and worktreeNextAction in ISubagentJobResult.metadata
  • include adapter-provided baseRevision and dirty parent checkout status in handoff metadata when available
  • preserve existing result metadata while adding worktree metadata
  • emit WorktreeRemove hook notification when a clean worktree is removed

createSubagentSession(options)

Assembles an isolated child Session for subagent execution. Unlike createSession, this factory does not load config files or context from disk — it receives pre-resolved config and context from the parent session.

Tool filtering order:

  1. Remove disallowed tools (denylist from agent definition)
  2. Keep only allowed tools (allowlist from agent definition, if specified)
  3. Always remove agent-spawning tools such as Agent and robota_command_agent (subagents cannot spawn subagents)

Model resolution: Agent definition model override (with shortcut expansion: sonnet, haiku, opus) takes priority; falls back to parent config model.

Agent Definitions

IAgentDefinition interface defines the shape for both built-in and custom agents:

FieldTypeRequiredDescription
namestringYesUnique agent identifier
descriptionstringYesHuman-readable purpose description
systemPromptstringYesMarkdown body used as the agent's system prompt
modelstringNoModel override (inherits parent when omitted)
maxTurnsnumberNoMaximum agentic turns
toolsstring[]NoAllowlist of tool names
disallowedToolsstring[]NoDenylist of tool names

Built-in agents:

NameModel OverrideTool RestrictionsPurpose
general-purpose(parent)None (inherits all)Full-capability task agent
Explore(parent)Denies Write, EditRead-only code exploration
Plan(parent)Denies Write, EditRead-only planning/research

Model-Requested Agent Invocation

Model-requested agent invocation is owned by @robota-sdk/agent-command-agent. The command module contributes agent as a model-invocable built-in command and requests the SDK agent-executor session requirement. The model route is the same projected command-tool path used by other built-ins: robota_command_agent({ args: "..." }).

The SDK stores agent runtime dependencies for the command module and for context: fork skills. It does not register a separate model-visible Agent function tool. Parallel, batch, detached, and worktree agent behavior belongs to agent command arguments and the shared runtime job APIs.

Structured command/background-task results are the only evidence that agent work started or completed. Assistant prose is not execution evidence.

When isolation: 'worktree' is requested, a runtime shell that supports worktree isolation must compose WorktreeSubagentRunner with a concrete ISubagentWorktreeAdapter. The runtime runner handles lifecycle, cleanup, handoff metadata, and WorktreeCreate / WorktreeRemove hook notifications; the shell adapter handles Git/filesystem I/O. Unsupported non-Git or shell states must fail with actionable messages unless the user explicitly requested non-isolated execution.

AgentDefinitionLoader (Internal)

AgentDefinitionLoader is an internal class — it is not exported from src/index.ts. It scans directories for custom .md agent definitions with YAML frontmatter, merged with built-in agents. Custom agents override built-in agents on name collision.

Scan directories (highest priority first):

  1. <cwd>/.robota/agents/ — project-level (Robota native)
  2. <cwd>/.agents/agents/ — project-level (Robota repository convention)
  3. <cwd>/.claude/agents/ — project-level (Claude Code compatible)
  4. <home>/.robota/agents/ — user-level (Robota native)
  5. <home>/.claude/agents/ — user-level (Claude Code compatible)

Framework System Prompt Suffixes

Two suffix modes appended to subagent system prompts:

  • Subagent suffix (default): Instructs the agent to report concisely to the caller
  • Fork worker suffix (isForkWorker: true): Instructs the agent to respond within 500 words, suitable for skill fork execution

assembleSubagentPrompt(options)

Assembles the full system prompt for a subagent session:

  1. Agent body (from agent definition systemPrompt)
  2. CLAUDE.md content (from parent context)
  3. AGENTS.md content (from parent context)
  4. Framework suffix (subagent or fork worker)

Subagent Transcript Logger

createSubagentLogger(parentSessionId, agentId, baseLogsDir) creates a FileSessionLogger for append-only subagent transcripts. Subagent sessions must run with sessionId = agentId, so the transcript is written to {baseLogsDir}/{parentSessionId}/subagents/{agentId}.jsonl.

Subagent transcript logs must include session initialization, prompts, tool calls/results, streaming text_delta chunks, final assistant output, context state, and errors. Parent sessions may store only transcript paths and task snapshots in .robota/sessions/*.json; the transcript JSONL remains the source of truth for high-frequency streaming data.

Unconnected Packages (Future Integration Targets)

PackageCurrent StateIntegration Direction
agent-tool-mcpUnconnectedConnect when MCP server is configured in InteractiveSession options
agent-teamUnconnectedReplace agent-tool.ts with agent-team delegation pattern
agent-event-serviceUnconnectedPublish Session lifecycle events
agent-plugin-*UnconnectedInject plugins during Session/Robota creation
agent-provider-openai/google/bytedanceUnconnectedConsumer passes provider to InteractiveSession({ cwd, provider })

Released under the MIT License.