deeda Think — developer docs

Every runtime is available. Select one in workflow frontmatter with runtime_key + method. Use Choosing a method to decide between api, sdk, and cli routes, and the Workflow Schema for every field you can set.

Runtime key	Category	Methods	Recommended model	Needs API key
`anthropic-agent-sdk`	agent	`sdk`, `cli`	`claude-opus-4-7`	yes
`gemini-antigravity-cli`	agent	`cli`, `sdk`	`gemini-3.5-flash`	yes
`gemini-genai-sdk`	chat	`sdk`	`gemini-2.5-pro`	yes
`local-llm`	chat	`api`	`gemma-4-e2b`	yes
`local-voice`	voice	— (desktop/voice surface)	`gemma-4-e2b`	yes
`openai-agents-sdk`	agent	`sdk`	`gpt-5.5`	yes
`openai-codex-sdk`	agent	`cli`	`gpt-5.5`	yes
`openai-operator`	agent	`api`	`gpt-5.5`	yes
`openai-realtime-api`	voice	`api`	`gpt-realtime-2`	yes
`openai-responses-api`	chat	`api`	`gpt-5.5`	yes

anthropic-agent-sdk

Anthropic Agent SDK — Anthropic Claude Agent SDK / Claude Code When to use: Use when you want Claude Code style coding and shell-native agent behavior with Anthropic models. Capability flags: canParseVerdicts, canReview, supportsAdaptiveThinking, supportsBatchApi, supportsBuiltinCodeExec, supportsBuiltinSearch, supportsCacheControl, supportsCitations, supportsCloudTTS, supportsCodeInterpreter, supportsComputerUse, supportsContext1M, supportsDeveloperRole, supportsDocumentInput, supportsFileInput, supportsImageInput, supportsLocalRealtimeVoice, supportsLocalSTT, supportsLocalTTS, supportsMultiTurn, supportsReasoningEffort, supportsServiceTier, supportsSessionResume, supportsStreaming, supportsStructuredOutput, supportsThinkingBudget, supportsToolCalling, supportsWebSearch Knobs (set under harness_config.sdk_settings.anthropic):

Knob	Type	What it does
`thinking_type`	enum	Anthropic thinking mode.
`thinking_budget_tokens`	number	Extended thinking budget.
`max_turns`	number	Maximum Claude agent turns.
`permission_mode`	enum	Claude tool approval mode.

Starter workflow (copy, then extend with snippets):

---
id: starter-anthropic-agent-sdk
version: 1.0.0
agent:
  runtime_key: anthropic-agent-sdk
  provider: anthropic
  model: claude-opus-4-7
  role: worker
  task_template: "Run {{workflow_id}} for {{issue_ref}}."
  max_turns: 8
  routing_strategy: single_agent_routine
  single_agent:
    select: fixed
    candidates:
      - runtime_key: anthropic-agent-sdk
        method: sdk
budget:
  tokens: 60000
  wall_clock_minutes: 30
sandbox:
  workspace_write: true
tools:
  required:
    - fs.read
    - fs.write
---

# Starter: Anthropic Agent SDK

Plan, implement, verify; record evidence before finishing.

gemini-antigravity-cli

Google Antigravity 2.0 CLI — Google Antigravity CLI/SDK bridge When to use: Use for Google Antigravity 2.0 (I/O 2026-05-19) — local agentic coding loop via the agy CLI + google-antigravity Python SDK. 11 BuiltinTools (list_directory/search_directory/find_file/view_file/create_file/edit_file/run_command/ask_question/start_subagent/generate_image/finish), lifecycle Hook classes (runtime-enumerated; the SDK self-updates), 7-primitive policy DSL (allow/deny/ask_user/allow_all/deny_all/safe_defaults/workspace_only), MCP transports resolved from the installed SDK (stdio + streamable-http today; the vendor folded SSE reads into streamable-http), file-change + interval triggers, multimodal input (Image/Document/Audio/Video). Pairs with raw Gemini API/SDK in tandem. Capability flags: canParseVerdicts, canReview, supportsBuiltinCodeExec, supportsBuiltinSearch, supportsCloudTTS, supportsDocumentInput, supportsFileInput, supportsImageInput, supportsLocalRealtimeVoice, supportsLocalSTT, supportsLocalTTS, supportsMultiTurn, supportsStreaming, supportsStructuredOutput, supportsThinkingBudget, supportsThinkingLevel, supportsToolCalling, supportsVideoInput, supportsWebSearch Knobs (set under harness_config.sdk_settings.gemini):

Knob	Type	What it does
`antigravity_model`	enum	SDK default “gemini-3.5-flash” (types.py:79). Other documented: “gemini-3.1-pro-preview” (types.py:125), image model “gemini-3.1-flash-image-preview” (types.py:80).
`antigravity_thinking_level`	enum	ThinkingLevel: MINIMAL \| LOW \| MEDIUM \| HIGH (types.py:83).
`antigravity_enable_subagents`	boolean	CapabilitiesConfig.enable_subagents — allow start_subagent tool.
`antigravity_enabled_tools`	string[]	BuiltinTools allowlist (mutually exclusive with disabled_tools).
`antigravity_disabled_tools`	string[]	BuiltinTools denylist (mutually exclusive with enabled_tools).
`antigravity_compaction_threshold`	number	Token count after which context window may be compacted.
`antigravity_image_model`	string	CapabilitiesConfig.image_model (default gemini-3.1-flash-image-preview).
`antigravity_finish_schema_json`	string	JSON schema for the finish tool — drives structured_output().
`antigravity_workspaces`	string[]	Workspace directories — file tools auto-restricted via policy.workspace_only().
`antigravity_mcp_servers`	json	McpServerConfig union — transport discriminators resolved from the installed SDK (stdio + streamable-http today; the historic McpSseServer was removed by the vendor).
`antigravity_vertex`	boolean	GeminiConfig(vertex=True) — Vertex AI backend via GCP ADC.
`antigravity_sandbox`	boolean	`agy --sandbox` flag — terminal-restricted sandbox.
`antigravity_add_dir`	string[]	`agy --add-dir` (repeatable) workspace directories.
`antigravity_print_timeout`	string	`agy --print-timeout` (default 5m0s).
`antigravity_log_file`	string	`agy --log-file` override path.
`antigravity_continue`	boolean	`agy --continue`/`-c` resume the most recent conversation.
`antigravity_conversation_id`	string	`agy --conversation <id>` resume specific past conversation.

Starter workflow (copy, then extend with snippets):

---
id: starter-gemini-antigravity-cli
version: 1.0.0
agent:
  runtime_key: gemini-antigravity-cli
  provider: gemini
  model: gemini-3.5-flash
  role: worker
  task_template: "Run {{workflow_id}} for {{issue_ref}}."
  max_turns: 8
  routing_strategy: single_agent_routine
  single_agent:
    select: fixed
    candidates:
      - runtime_key: gemini-antigravity-cli
        method: cli
budget:
  tokens: 60000
  wall_clock_minutes: 30
sandbox:
  workspace_write: true
tools:
  required:
    - fs.read
    - fs.write
---

# Starter: Google Antigravity 2.0 CLI

Plan, implement, verify; record evidence before finishing.

gemini-genai-sdk

Gemini GenAI SDK — Google GenAI SDK When to use: Use when you want Gemini native streaming, Google Search, code execution, or Vertex AI-backed deployment. Capability flags: canParseVerdicts, canReview, supportsAudioInput, supportsBuiltinCodeExec, supportsBuiltinSearch, supportsCloudTTS, supportsDocumentInput, supportsFileInput, supportsImageInput, supportsLocalRealtimeVoice, supportsLocalSTT, supportsLocalTTS, supportsMultiTurn, supportsStreaming, supportsStructuredOutput, supportsThinkingBudget, supportsThinkingLevel, supportsToolCalling, supportsVideoInput, supportsWebSearch Knobs (set under harness_config.sdk_settings.gemini):

Knob	Type	What it does
`gemini_google_search`	boolean	Enable built-in Google Search grounding.
`gemini_code_execution`	boolean	Enable Gemini code execution.
`gemini_use_vertex_ai`	boolean	Switch auth/transport to Vertex AI.
`gemini_gcp_project`	string	Vertex AI GCP project.
`gemini_gcp_location`	string	Vertex AI region.

Starter workflow (copy, then extend with snippets):

---
id: starter-gemini-genai-sdk
version: 1.0.0
agent:
  runtime_key: gemini-genai-sdk
  provider: gemini
  model: gemini-2.5-pro
  role: worker
  task_template: "Run {{workflow_id}} for {{issue_ref}}."
  max_turns: 8
  routing_strategy: single_agent_routine
  single_agent:
    select: fixed
    candidates:
      - runtime_key: gemini-genai-sdk
        method: sdk
budget:
  tokens: 60000
  wall_clock_minutes: 30
sandbox:
  workspace_write: true
tools:
  required:
    - fs.read
    - fs.write
---

# Starter: Gemini GenAI SDK

Plan, implement, verify; record evidence before finishing.

local-llm

Local LLM — Local OpenAI-compatible LLM runtime When to use: Use for managed llama.cpp or a catalog-backed local endpoint covering the V1 local universe: llama.cpp, Ollama, MLX, and LocalAI. Sprint 07 signoff is llama.cpp only. Capability flags: canParseVerdicts, canReview, supportsCloudTTS, supportsImageInput, supportsLocalRealtimeVoice, supportsLocalSTT, supportsLocalTTS, supportsMultiTurn, supportsStreaming, supportsStructuredOutput, supportsToolCalling Knobs (set under harness_config.sdk_settings.local):

Knob	Type	What it does
`temperature`	number	Sampling temperature.
`max_tokens`	number	Maximum generation length.
`LOCAL_LLM_BASE_URL`	string	Catalog-backed local loopback base URL.

Starter workflow (copy, then extend with snippets):

---
id: starter-local-llm
version: 1.0.0
agent:
  runtime_key: local-llm
  provider: local
  model: gemma-4-e2b
  role: worker
  task_template: "Run {{workflow_id}} for {{issue_ref}}."
  max_turns: 8
  routing_strategy: single_agent_routine
  single_agent:
    select: fixed
    candidates:
      - runtime_key: local-llm
        method: api
budget:
  tokens: 60000
  wall_clock_minutes: 30
sandbox:
  workspace_write: true
tools:
  required:
    - fs.read
    - fs.write
---

# Starter: Local LLM

Plan, implement, verify; record evidence before finishing.

local-voice

Local Voice — Desktop voice shell When to use: Use when you want local STT plus a provider-agnostic voice shell that can drive any text runtime. Capability flags: canParseVerdicts, canReview, supportsAudioInput, supportsAudioOutput, supportsCloudTTS, supportsImageInput, supportsLocalRealtimeVoice, supportsLocalSTT, supportsLocalTTS, supportsMultiTurn, supportsRealtimeVoice, supportsStreaming, supportsStructuredOutput, supportsToolCalling Knobs (set under harness_config.sdk_settings.local):

Knob	Type	What it does
`voice_stt_backend`	enum	Speech-to-text backend.
`voice_tts_backend`	enum	Text-to-speech backend.
`voice_tts_voice_id`	string	Preferred voice identifier.

No turn-dispatch method: this runtime is used through desktop/voice surfaces, not workflow method selection.

openai-agents-sdk

OpenAI Agents SDK — OpenAI Agents SDK When to use: Use when you need declarative agents, hosted tools, typed handoffs, and guardrails in the OpenAI Agents SDK. Capability flags: canParseVerdicts, canReview, supportsAudioInput, supportsAudioOutput, supportsBatchApi, supportsBuiltinCodeExec, supportsBuiltinSearch, supportsCloudTTS, supportsCodeInterpreter, supportsCodexSandboxMode, supportsCodexThreadResume, supportsComputerUse, supportsCustomTools, supportsDeveloperRole, supportsDocumentInput, supportsFileInput, supportsHostedConnectors, supportsHostedContainer, supportsHostedConversations, supportsHostedFileSearch, supportsHostedMcp, supportsHostedSkills, supportsImageGeneration, supportsImageInput, supportsLocalRealtimeVoice, supportsLocalSTT, supportsLocalTTS, supportsMultiTurn, supportsPromptCacheKey, supportsPromptTemplate, supportsReasoningEffort, supportsServerAgentLoop, supportsServerCompaction, supportsServiceTier, supportsStreaming, supportsStructuredOutput, supportsToolCalling, supportsWebSearch Knobs (set under harness_config.sdk_settings.openai):

Knob	Type	What it does
`max_tokens`	number	Maximum response tokens.
`reasoning_effort`	enum	Reasoning effort for supported models.
`openai_tool_profile`	enum	Hosted tool set selection.

Starter workflow (copy, then extend with snippets):

---
id: starter-openai-agents-sdk
version: 1.0.0
agent:
  runtime_key: openai-agents-sdk
  provider: openai
  model: gpt-5.5
  role: worker
  task_template: "Run {{workflow_id}} for {{issue_ref}}."
  max_turns: 8
  routing_strategy: single_agent_routine
  single_agent:
    select: fixed
    candidates:
      - runtime_key: openai-agents-sdk
        method: sdk
budget:
  tokens: 60000
  wall_clock_minutes: 30
sandbox:
  workspace_write: true
tools:
  required:
    - fs.read
    - fs.write
---

# Starter: OpenAI Agents SDK

Plan, implement, verify; record evidence before finishing.

openai-codex-sdk

OpenAI Codex SDK — OpenAI Codex CLI/SDK bridge When to use: Use when you want Codex CLI autonomy, workspace-aware edits, and approval-driven coding loops. Capability flags: canParseVerdicts, canReview, supportsAudioInput, supportsAudioOutput, supportsBatchApi, supportsBuiltinCodeExec, supportsBuiltinSearch, supportsCloudTTS, supportsCodeInterpreter, supportsCodexSandboxMode, supportsCodexThreadResume, supportsComputerUse, supportsCustomTools, supportsDeveloperRole, supportsDocumentInput, supportsFileInput, supportsHostedContainer, supportsHostedSkills, supportsImageGeneration, supportsImageInput, supportsLocalRealtimeVoice, supportsLocalSTT, supportsLocalTTS, supportsMultiTurn, supportsPromptCacheKey, supportsPromptTemplate, supportsReasoningEffort, supportsServerAgentLoop, supportsServerCompaction, supportsServiceTier, supportsStreaming, supportsStructuredOutput, supportsToolCalling, supportsWebSearch Knobs (set under harness_config.sdk_settings.openai):

Knob	Type	What it does
`reasoning_effort`	enum	Reasoning effort passed through to Codex.
`codex_sandbox_mode`	enum	Filesystem sandbox mode.
`codex_approval_policy`	enum	CLI approval policy.
`codex_config_json`	json	Raw Codex TOML/CLI config overrides.

Starter workflow (copy, then extend with snippets):

---
id: starter-openai-codex-sdk
version: 1.0.0
agent:
  runtime_key: openai-codex-sdk
  provider: openai
  model: gpt-5.5
  role: worker
  task_template: "Run {{workflow_id}} for {{issue_ref}}."
  max_turns: 8
  routing_strategy: single_agent_routine
  single_agent:
    select: fixed
    candidates:
      - runtime_key: openai-codex-sdk
        method: cli
budget:
  tokens: 60000
  wall_clock_minutes: 30
sandbox:
  workspace_write: true
tools:
  required:
    - fs.read
    - fs.write
---

# Starter: OpenAI Codex SDK

Plan, implement, verify; record evidence before finishing.

openai-operator

OpenAI Operator — OpenAI computer-use/operator path When to use: Use when you need computer-use style browser automation, screenshots, and hosted connectors through OpenAI tool primitives. Capability flags: canParseVerdicts, canReview, supportsAudioInput, supportsAudioOutput, supportsAudioStreaming, supportsBatchApi, supportsBuiltinCodeExec, supportsBuiltinSearch, supportsCloudTTS, supportsCodeInterpreter, supportsCodexSandboxMode, supportsCodexThreadResume, supportsComputerUse, supportsCustomTools, supportsDeveloperRole, supportsDocumentInput, supportsFileInput, supportsHostedConnectors, supportsHostedContainer, supportsHostedConversations, supportsHostedFileSearch, supportsHostedMcp, supportsHostedSkills, supportsImageGeneration, supportsImageInput, supportsLocalRealtimeVoice, supportsLocalSTT, supportsLocalTTS, supportsMultiTurn, supportsPromptCacheKey, supportsPromptTemplate, supportsRealtimeSIP, supportsRealtimeVoice, supportsRealtimeWebRTC, supportsReasoningEffort, supportsServerAgentLoop, supportsServerCompaction, supportsServerVAD, supportsServiceTier, supportsStreaming, supportsStructuredOutput, supportsToolCalling, supportsWebSearch, supportsWebhookDelivery Knobs (set under harness_config.sdk_settings.openai):

Knob	Type	What it does
`openai_connectors`	string[]	Hosted connector IDs (Drive, Gmail, etc.); emitted as {type:mcp, connector_id} tools.
`openai_connector_authorization`	string	Per-connector OAuth access token map (connector id → token); required by the API for every connector tool.
`computer_use_display`	string	Display config {width,height,environment}; selects the older computer_use_preview tool (the current computer tool takes no display parameters).
`tool_approval_mode`	enum	Approval mode for computer-use actions.
`openai_include_computer_call_image`	boolean	Include screenshot/image payloads in runtime events.

Starter workflow (copy, then extend with snippets):

---
id: starter-openai-operator
version: 1.0.0
agent:
  runtime_key: openai-operator
  provider: openai
  model: gpt-5.5
  role: worker
  task_template: "Run {{workflow_id}} for {{issue_ref}}."
  max_turns: 8
  routing_strategy: single_agent_routine
  single_agent:
    select: fixed
    candidates:
      - runtime_key: openai-operator
        method: api
budget:
  tokens: 60000
  wall_clock_minutes: 30
sandbox:
  workspace_write: true
tools:
  required:
    - fs.read
    - fs.write
---

# Starter: OpenAI Operator

Plan, implement, verify; record evidence before finishing.

openai-realtime-api

OpenAI Realtime API — OpenAI Realtime API When to use: Use when you need low-latency audio or multimodal Realtime sessions with server VAD and voice output. Capability flags: canParseVerdicts, canReview, supportsAudioInput, supportsAudioOutput, supportsAudioStreaming, supportsBatchApi, supportsBuiltinCodeExec, supportsBuiltinSearch, supportsCloudTTS, supportsCodeInterpreter, supportsCodexSandboxMode, supportsCodexThreadResume, supportsComputerUse, supportsCustomTools, supportsDeveloperRole, supportsDocumentInput, supportsFileInput, supportsHostedContainer, supportsHostedFileSearch, supportsHostedMcp, supportsHostedSkills, supportsImageGeneration, supportsImageInput, supportsLocalRealtimeVoice, supportsLocalSTT, supportsLocalTTS, supportsMultiTurn, supportsPromptCacheKey, supportsRealtimeSIP, supportsRealtimeVoice, supportsRealtimeWebRTC, supportsReasoningEffort, supportsServerAgentLoop, supportsServerCompaction, supportsServerVAD, supportsServiceTier, supportsStreaming, supportsStructuredOutput, supportsToolCalling, supportsWebSearch Knobs (set under harness_config.sdk_settings.openai):

Knob	Type	What it does
`voice`	string	Realtime voice name.
`modalities`	string[]	Realtime modalities to request.
`turn_detection`	enum	Server VAD / turn detection strategy.

Starter workflow (copy, then extend with snippets):

---
id: starter-openai-realtime-api
version: 1.0.0
agent:
  runtime_key: openai-realtime-api
  provider: openai
  model: gpt-realtime-2
  role: worker
  task_template: "Run {{workflow_id}} for {{issue_ref}}."
  max_turns: 8
  routing_strategy: single_agent_routine
  single_agent:
    select: fixed
    candidates:
      - runtime_key: openai-realtime-api
        method: api
budget:
  tokens: 60000
  wall_clock_minutes: 30
sandbox:
  workspace_write: true
tools:
  required:
    - fs.read
    - fs.write
---

# Starter: OpenAI Realtime API

Plan, implement, verify; record evidence before finishing.

openai-responses-api

OpenAI Responses API — OpenAI Responses API When to use: Use when you need the broadest OpenAI knob surface, structured output, hosted tools, background runs, and webhook delivery. Capability flags: canParseVerdicts, canReview, supportsAudioInput, supportsAudioOutput, supportsBatchApi, supportsBuiltinCodeExec, supportsBuiltinSearch, supportsCloudTTS, supportsCodeInterpreter, supportsCodexSandboxMode, supportsCodexThreadResume, supportsComputerUse, supportsCustomTools, supportsDeveloperRole, supportsDocumentInput, supportsFileInput, supportsHostedConnectors, supportsHostedContainer, supportsHostedConversations, supportsHostedFileSearch, supportsHostedMcp, supportsHostedSkills, supportsImageGeneration, supportsImageInput, supportsLocalRealtimeVoice, supportsLocalSTT, supportsLocalTTS, supportsMultiTurn, supportsPromptCacheKey, supportsPromptTemplate, supportsReasoningEffort, supportsServerAgentLoop, supportsServerCompaction, supportsServiceTier, supportsStreaming, supportsStructuredOutput, supportsToolCalling, supportsWebSearch, supportsWebhookDelivery Knobs (set under harness_config.sdk_settings.openai):

Knob	Type	What it does
`reasoning_effort`	enum	Reasoning effort for GPT-5/o-series models.
`openai_background`	boolean	Background delivery via webhook.
`openai_connectors`	string[]	Hosted connector IDs.
`openai_webhook_url`	string	Webhook receiver for async delivery.

Starter workflow (copy, then extend with snippets):

---
id: starter-openai-responses-api
version: 1.0.0
agent:
  runtime_key: openai-responses-api
  provider: openai
  model: gpt-5.5
  role: worker
  task_template: "Run {{workflow_id}} for {{issue_ref}}."
  max_turns: 8
  routing_strategy: single_agent_routine
  single_agent:
    select: fixed
    candidates:
      - runtime_key: openai-responses-api
        method: api
budget:
  tokens: 60000
  wall_clock_minutes: 30
sandbox:
  workspace_write: true
tools:
  required:
    - fs.read
    - fs.write
---

# Starter: OpenAI Responses API

Plan, implement, verify; record evidence before finishing.

Getting Started

Concepts

Agent Harness

Provider Capabilities

Runtimes

anthropic-agent-sdk

gemini-antigravity-cli

gemini-genai-sdk

local-llm

local-voice

openai-agents-sdk

openai-codex-sdk

openai-operator

openai-realtime-api

openai-responses-api

​anthropic-agent-sdk

​gemini-antigravity-cli

​gemini-genai-sdk

​local-llm

​local-voice

​openai-agents-sdk

​openai-codex-sdk

​openai-operator

​openai-realtime-api

​openai-responses-api

anthropic-agent-sdk

gemini-antigravity-cli

gemini-genai-sdk

local-llm

local-voice

openai-agents-sdk

openai-codex-sdk

openai-operator

openai-realtime-api

openai-responses-api