Skip to content

opencode.md

Latest commit

 

History

History
89 lines (65 loc) · 4.52 KB

opencode.md

File metadata and controls

89 lines (65 loc) · 4.52 KB

connector-opencode

Agentic coding via a locally-running opencode serve daemon. A bot hands off coding tasks in plain language — "fix this bug", "add tests", "refactor X" — and gets the agent's reply back. The daemon does the file edits and command runs on the host; from Springtale's sandbox the connector only makes HTTP calls to the loopback daemon.

  Bot / rule engine        connector-opencode          opencode serve
        │                          │                  (127.0.0.1:4096)
        │ execute(run_task,        │                          │
        │   {prompt})              │                          │
        ├─────────────────────────►│ POST /session            │
        │                          ├─────────────────────────►│
        │                          │ POST /session/{id}/message
        │                          ├─────────────────────────►│
        │                          │      (agent edits files, │
        │                          │       runs commands)     │
        │                          │◄─────────────────────────┤
        │ {session_id, reply}      │  info + parts            │
        │◄─────────────────────────┤                          │

Fig. 1. OpenCode data flow. Action-only — no triggers. Coding actions are read_only: false, so the chat-approval gate (W2) fronts them.

1. Configuration

TABLE I. CONFIG FIELDS

Field Type Default Description
base_url String "http://127.0.0.1:4096" Base URL of the running opencode serve daemon
password Option<Secret<String>> None Daemon HTTP basic-auth password (its OPENCODE_SERVER_PASSWORD). Omit when the daemon runs without auth
model Option<String> None Model id passed through on each prompt (e.g. anthropic/claude-sonnet-4). None uses the daemon's configured default
agent Option<String> None Route prompts to a specific opencode agent

2. Authentication

HTTP basic auth against the daemon. The username is fixed (opencode); only the password varies. When password is unset, requests are sent unauthenticated — fine for a loopback-only daemon.

3. Triggers

None. The connector is action-only; all work flows to the local daemon on demand.

4. Actions

TABLE II. ACTIONS

Name Input fields Output fields
run_task prompt (the coding task, in plain language), title (optional session title for later reference) session_id, reply (the agent's text reply), response (raw message response: info + parts)
continue_session session_id (from a previous run_task), prompt (the follow-up instruction, e.g. "now add tests") session_id, reply, response

Both actions are read_only: false — agentic coding edits files and runs commands on the host, so they are never auto-granted under a Reconnoiter intent and the approval gate fronts them.

5. Capabilities Required

Capability Parameter
NetworkOutbound host:port parsed from base_url (default 127.0.0.1:4096)

6. Example Rule

[rule]
name = "fix-on-demand"

[trigger]
type = "ConnectorEvent"
connector = "connector-telegram"
event = "message_received"

[[conditions]]
type = "FieldContains"
field = "trigger.text"
value = "/fix"

[[actions]]
type = "RunConnector"
connector = "connector-opencode"
action = "run_task"

[actions.params]
prompt = "${trigger.text}"
title = "telegram-requested fix"

7. Security Notes

  • The daemon is the trust boundary. opencode serve edits files and executes commands with the privileges of whoever started it. Springtale's sandbox constrains the connector (loopback HTTP only), not the daemon. Run the daemon in the workspace you intend it to touch, nothing wider.
  • Approval-gated. Because both actions are mutating, formation intent decomposition never selects them for monitor-only intents, and interactive surfaces require explicit approval before dispatch.
  • Keep it loopback. Pointing base_url at a non-loopback host sends your prompts (and the agent's file contents) over the network. Don't, unless you control the transport end-to-end.