Skip to content

Thread options ​

What startThread, resumeThread and forkThread take, and what each option becomes on each runtime.

OptionCodexClaudeACP
cwdThread cwdProcess cwdSession cwd
model, effortThread and per-turnSDK options and per-turn controlsThe session's model and thought_level options
fastModepriority service tier, thread and per-turnFast mode setting, and per-turn controlsIgnored
serviceTierCatalogue tier ID, or null for the runtime default; overrides fastMode, thread and per-turnNot supportedNot supported
accessApproval policy, sandbox and reviewerPermission modeAnswers permission requests itself where it can
mode (default, plan)The thread's collaboration mode, set before the turnPlan permission modeThe session mode named plan
instructionsDeveloper instructionsAppended to Claude Code's system promptNot sent
mcpServersPer-thread mcp_servers configSDK mcpServers (also in-process sdk servers, and alwaysLoad)session/new servers
envEngine-wide only (CodexEngineOptions.env)Per threadEngine-wide only
strictModelFails a silent model swapSame
titlethread/name/setSDK titleIgnored
outputSchemaPer-turn outputSchemaoutputFormatNot supported
limitsNot supportedmaxTurns, maxBudgetUsd, task budgetNot supported
agentsAgent roles (agents.<name> config with a role file): description, prompt, model and effortSDK agents, with each agent's claude settings (MCP servers, skills, turn limit, background, memory, permission mode…)Not supported
additionalDirectoriesWorkspace sandbox writable rootsSDK additionalDirectoriesSent when the agent takes them
fileCheckpointsA git snapshot of the folder before each turn, for rewind({ files: true })File checkpointing, for rewind({ files: true })Ignored
seedInjected as historyPut before the first messagePut before the first message
fallbackModels, promptSuggestions, progressSummariesIgnoredSDK fallbackModel, promptSuggestions, agentProgressSummariesIgnored

"Not supported" options throw UnsupportedFeatureError when set, rather than being dropped.

Access levels ​

access is one of:

LevelMeaning
fullNo approvals
autoThe runtime's own reviewer decides routine approvals. Not on ACP.
accept-editsEdits go ahead. On Codex this is the same as ask, since the workspace sandbox already allows them.
askThe agent asks
read-onlyNo changes

capabilities().access lists the levels a runtime has a native setting for.

Runtime-specific options ​

Anything specific to one runtime goes in codex: {…} or claude: {…}.

codex

  • Approval policy, approvals reviewer and sandbox
  • Network, also reasserted with the sandbox and approval policy on every turn
  • baseInstructions, which replaces Codex's base instructions entirely
  • config overrides
  • Host-answered dynamicTools. CodexEngineOptions.onToolCall answers them with text, and with images and audio as URLs in content
  • Ephemeral threads
  • listMcpTools

claude

  • Setting sources, built-in tools, and allowed and disallowed tools
  • Permission mode, skills and system prompt
  • executable, the Claude Code binary to run
  • strictMcpConfig, which loads only mcpServers and ignores servers from settings and plugins
  • forwardSubagentText, which forwards subagents' text and thinking as well as their tool calls (on by default)
  • Host hooks and local plugins
  • extra for any other SDK option

Released under the MIT Licence. Not affiliated with OpenAI, Anthropic or the ACP project.