Agents

Claude Code and Codex

Choose the coding agent that works on your tasks, and tune its model, permissions, effort and timeouts.

SI Fabrik runs one of two agent runtimes for each project, set with runner.kind. Both get the same prompt, the same tracker tools and the same sandbox.

	Claude Code	Codex
`runner.kind`	`claude_code` (default)	`codex`
Protocol	`claude -p` with stream-json input and output	`codex app-server` (JSON-RPC)
Session	One long-lived process per attempt; continuation turns reuse it	One app-server session per attempt
Permissions	`claude.permission_mode` (default `bypassPermissions`)	Approvals accepted for the session; policies passed through
Credential in a sandbox	Anthropic API key, `claude setup-token` token, or a gateway key	OpenAI API key, or a gateway key
In the agent image	Yes	Yes

Claude Code#

SI Fabrik starts Claude Code in print mode with a stream-json protocol:

Shell

claude -p --input-format stream-json --output-format stream-json --verbose \
  --permission-mode bypassPermissions \
  --mcp-config '{"mcpServers":{"symphony":{...},"browser":{...}}}' \
  --append-system-prompt '…'

Each turn writes one user message, and the turn completes on the result event. Print mode never blocks on a question, and MCP elicitations are declined.

WORKFLOW.md

runner:
  kind: claude_code
claude:
  command: claude
  model: claude-opus-5-5          # --model
  permission_mode: bypassPermissions
  effort: high                    # --effort: low | medium | high | xhigh | max
  max_budget_usd: 5               # --max-budget-usd, per session
  allowed_tools: []               # --allowedTools
  disallowed_tools: [WebFetch]    # --disallowedTools
  append_system_prompt: |
    Prefer small, focused commits.
  extra_args: []

Permission modes#

Mode	Meaning
`bypassPermissions`	No prompts at all. High-trust: keep a sandbox on. The agent image runs as a non-root user, so bypass is allowed inside the container.
`acceptEdits`	File edits are accepted; other actions follow Claude Code's rules.
`default`	Claude Code's default rules.
`plan`	Plan mode.

Nobody is there to answer a permission prompt, so with the stricter modes the agent can't do anything those rules don't already allow.

Codex#

WORKFLOW.md

runner:
  kind: codex
codex:
  command: codex app-server
  model: gpt-5.5-codex
  approval_policy: never
  thread_sandbox: workspace-write
  # turn_sandbox_policy: …

Command and file-change approvals are accepted for the session (acceptForSession). approval_policy, thread_sandbox and turn_sandbox_policy are passed to Codex unchanged.
A request for user input (item/tool/requestUserInput) fails the turn with turn_input_required.
A dynamic tool call returns success: false and the session continues.
When Codex runs on the host with a gateway, it gets an isolated per-project CODEX_HOME, so a personal ~/.codex login can't take over from the gateway key.

Timeouts (both runners)#

Following the Symphony spec, the timeouts live under codex: but apply to both runners:

Key	Default	Effect
`codex.turn_timeout_ms`	1 hour	Silence timeout: the turn fails after this long without any agent output.
`codex.stall_timeout_ms`	5 minutes	A run with no activity for this long is stopped and retried (`stalled`). `0` turns it off.
`codex.read_timeout_ms`	5 seconds	Protocol read timeout.

The wizard writes turn_timeout_ms: 3600000 and stall_timeout_ms: 600000.

Models#

Leave claude.model (or codex.model) empty to use the runner's default, or the gateway's mapping. The wizard and the Workflow form offer Claude Opus 5.5, Sonnet 5.5 and Haiku 5.5 (claude-opus-5-5, claude-sonnet-5-5, claude-haiku-5-5), or the models your gateway lists.

Change the runner#

Switch Runner in Workflow → Settings → Agents, or edit runner.kind, and save. The project reloads the workflow right away.