Agent Runtimes

CORAL works with any coding agent that can run as a subprocess. Pick the runtime in task.yaml via agents.runtime. Claude Code is the default and needs no special config beyond an Anthropic API key. The other runtimes are documented below.

DeepSeek Harness

Install the official DeepSeek Harness CLI and make sure dsh is on PATH:

npm install --global @deepseek-ai/dsh
export DEEPSEEK_API_KEY=your-key

Select its canonical runtime name in the task config:

agents:
  runtime: dsh
  model: deepseek-v4-flash

CORAL invokes dsh --profile headless <prompt>, points DSH_HOME at the agent's shared .dsh/ directory, and writes the task instructions to AGENTS.md at the worktree root. It emits a final Cordis overlay so agents.model overrides the profile default. The aliases deepseek, deepseek-harness, and deepseek_harness are accepted for compatibility, but new configurations should use dsh.

When the CORAL LiteLLM gateway is enabled, gateway_url and gateway_api_key are exposed to DSH as DEEPSEEK_BASE_URL and DEEPSEEK_API_KEY. Direct runs inherit the same environment variables from the shell.

Optional runtime_options:

agents:
  runtime: dsh
  runtime_options:
    command: /usr/local/bin/dsh       # override binary path
    profile: headless                 # override the DSH profile
    provider: deepseek-official       # provider for agents.model
    patch: [./dsh.cordis.yml]         # repeatable --patch overlays
    permission_mode: workspace-write  # DSH_PERMISSION_MODE
    tools_mode: native                # DSH_TOOLS_MODE

Note: The current DSH headless profile creates a fresh persisted session but does not expose its session ID or a resume argument. After a CORAL-managed restart, DSH starts a new session with a continuation prompt and recovers context from the worktree and shared .dsh/ state.

OpenCode

OpenCode reads permissions and provider config from an opencode.json file. Place it in your task's seed/ directory so it gets copied into each agent's worktree.

Example (from examples/circle_packing/seed/opencode.json):

{
  "$schema": "https://opencode.ai/config.json",
  "permission": {
    "external_directory": "allow",
    "question": "deny",
    "doom_loop": "allow",
    "bash": "allow",
    "edit": "allow",
    "read": "allow",
    "write": "allow",
    "webfetch": "deny",
    "websearch": "deny",
    "codesearch": "allow",
    "lsp": "allow",
    "skill": "allow"
  },
  "provider": {
    "claude": {
      "npm": "@ai-sdk/anthropic",
      "name": "claude",
      "options": {
        "baseURL": "http://localhost:4000/v1",
        "apiKey": "xxx"
      },
      "models": {
        "claude-opus-4-6": {
          "name": "claude-opus-4-6"
        }
      }
    }
  }
}

Key points:

• Set all permissions to "allow" except question, webfetch, websearch (set those to "deny") so the agent runs autonomously without interactive prompts.
• The provider section configures which model to use. When using the gateway, point baseURL at http://localhost:<gateway_port>/v1 and set apiKey to any placeholder value — the gateway handles authentication.
• Place opencode.json in your seed directory so it gets copied into each agent's worktree.

Task config:

agents:
  runtime: opencode
  model: claude/claude-opus-4-6  # must match a model defined in opencode.json

Cursor Agent

Install Cursor Agent and authenticate once:

curl -fsSL https://cursor.com/install | bash
cursor-agent login

Point your task at it:

agents:
  runtime: cursor          # alias for cursor_agent; "cursor-agent" also works
  model: auto              # or any model id supported by your Cursor plan

CORAL spawns each agent as cursor-agent --print --output-format stream-json --force --workspace <wt> [--model] [--mode] [--resume] <prompt>. The --force flag is always passed (cursor-agent requires it for write tools in --print mode). The full task brief is written to AGENTS.md at the worktree root, and CORAL also drops a .cursor/rules/coral.mdc always-apply rule with short guardrails (use coral eval, don't touch .coral/private/, share via .cursor/notes/ and .cursor/skills/) so they survive context pressure.

Optional runtime_options:

agents:
  runtime: cursor
  runtime_options:
    command: /usr/local/bin/cursor-agent  # override binary path
    mode: plan                            # --mode plan|ask
    stream_partial_output: true           # --stream-partial-output

Note: Cursor Agent uses its own auth and does not route through the LiteLLM gateway. Login state lives in your Cursor account, not in CORAL config.

Kiro

Install Kiro (kiro-cli) and authenticate via Kiro's setup. Then:

agents:
  runtime: kiro
  model: auto              # or any Kiro-supported model

CORAL spawns each agent as kiro-cli chat <prompt> --no-interactive -a (the -a flag trusts all tools). Instructions are read from KIRO.md at the worktree root. Kiro does not currently expose a session id we can resume from, so restarted agents start fresh — they still see all prior attempts and notes via the shared .kiro/ directory.

Like Cursor, Kiro uses its own auth and does not route through the LiteLLM gateway.