opcode（opcode 开发）

OPCODE

Execution runtime for AI agents. You reason, OPCODE executes — zero tokens per run after the first define. Workflows persist across sessions, run on schedules, and coordinate multiple agents. Persistent SSE daemon: 1 server, N agents, 1 database. JSON-defined DAGs, level-by-level execution, automatic parallelism. 6 MCP tools over SSE (JSON-RPC).

Why use OPCODE instead of reasoning through each step yourself? Every repeated workflow burns tokens re-reasoning decisions you already made. OPCODE templates your reasoning once and executes it deterministically — zero inference cost, identical output every run, survives context resets.

Which Tool?

Quick Start

Install:

go install github.com/rendis/opcode/cmd/opcode@latest

First-time setup (writes config and starts daemon):

opcode install --listen-addr :4100 --vault-key "my-passphrase"

Restart after stop: OPCODE_VAULT_KEY="my-passphrase" opcode

MCP client configuration:

{
  "mcpServers": {
  "mcpServers": {
    "opcode": {
      "type": "sse",
      "url": "http://localhost:4100/sse"
    }
  }
}

Each agent self-identifies via agent_id in tool calls. Opcode auto-registers unknown agents. Choose a stable ID per agent (e.g., "content-writer", "deploy-bot").

Workflows survive restarts. On startup, orphaned active workflows become suspended. Query with opcode.query({ "resource": "workflows", "filter": { "status": "suspended" } }), then resume or cancel via opcode.signal.

See operations.md for full configuration, subcommands, SIGHUP hot-reload, security model, web panel, and benchmarks.

MCP Tools

opcode.define

Registers a reusable workflow template. Version auto-increments (v1, v2, v3...).

Returns: { "name": "...", "version": "v1" }

opcode.run

Executes a workflow from a registered template.

Returns:

{
  "workflow_id": "uuid",
  "status": "completed | suspended | failed",
  "output": { ... },
  "started_at": "RFC3339",
  "completed_at": "RFC3339",
  "steps": {
    "step-id": { "step_id": "...", "status": "completed", "output": {...}, "duration_ms": 42 }
  }
}

If status is "suspended", call opcode.status to see pending_decisions.

opcode.status

Gets workflow execution status.

Returns:

{
  "workflow_id": "uuid",
  "status": "suspended",
  "steps": { "step-id": { "status": "...", "output": {...} } },
  "pending_decisions": [
    {
      "id": "uuid",
      "step_id": "reason-step",
      "context": { "prompt": "...", "data": {...} },
      "options": [ { "id": "approve", "description": "Proceed" } ],
      "timeout_at": "RFC3339",
      "fallback": "reject",
      "status": "pending"
    }
  ],
  "events": [ ... ]
}

Workflow statuses: pending, active, suspended, completed, failed, cancelled.

opcode.signal

Sends a signal to a suspended workflow.

Payload by signal type:

Returns (decision): { "ok": true, "resumed": true, "status": "completed", ... } Returns (other): { "ok": true, "workflow_id": "...", "signal_type": "..." }

opcode.query

Queries workflows, events, or templates.

Filter fields by resource:

Note: event queries require either event_type or workflow_id in filter.

Returns: { "<resource>": [...] } -- results wrapped in object keyed by resource type.

opcode.diagram

Generates a visual DAG diagram from a template or running workflow.

\* One of template_name or workflow_id required.

• template_name -- preview DAG structure before execution
• workflow_id -- visualize with live step status
• format: "ascii" -- CLI-friendly text with box-drawing characters
• format: "mermaid" -- markdown-embeddable flowchart syntax
• format: "image" -- base64-encoded PNG for visual channels

Returns: { "format": "ascii", "diagram": "..." }

Workflow Definition

{
  "steps": [ ... ],
  "inputs": { "key": "value or ${{secrets.KEY}}" },
  "context": { "intent": "...", "notes": "..." },
  "timeout": "5m",
  "on_timeout": "fail | suspend | cancel",
  "on_complete": { /* step definition */ },
  "on_error": { /* step definition */ },
  "metadata": {}
}

Step Definition

{
  "id": "step-id",
  "type": "action | condition | loop | parallel | wait | reasoning",
  "action": "http.get",
  "params": { ... },
  "depends_on": ["other-step"],
  "condition": "CEL guard expression",
  "timeout": "30s",
  "retry": { "max": 3, "backoff": "exponential", "delay": "1s", "max_delay": "30s" },
  "on_error": { "strategy": "ignore | fail_workflow | fallback_step | retry", "fallback_step": "id" },
  "config": { /* type-specific */ }
}

type defaults to action. See workflow-schema.md for all config blocks.

Step Types

action (default)

Executes a registered action. Set action to the action name, params for input.

condition

Evaluates a CEL expression and branches.

{
  "id": "route",
  "type": "condition",
  "config": {
    "expression": "inputs.env",
    "branches": { "prod": [...], "staging": [...] },
    "default": [...]
  }
}

loop

Iterates over a collection or condition. Loop variables: ${{loop.item}}, ${{loop.index}}.

{
  "id": "process-items",
  "type": "loop",
  "config": {
    "mode": "for_each",
    "over": "[\"a\",\"b\",\"c\"]",
    "body": [
      {
        "id": "hash",
        "action": "crypto.hash",
        "params": { "data": "${{loop.item}}" }
      }
    ],
    "max_iter": 100
  }
}

Modes: for_each (iterate over), while (loop while condition true), until (loop until condition true).

parallel

Executes branches concurrently.

{
  "id": "fan-out",
  "type": "parallel",
  "config": {
    "mode": "all",
    "branches": [
      [{ "id": "a", "action": "http.get", "params": {...} }],
      [{ "id": "b", "action": "http.get", "params": {...} }]
    ]
  }
}

Modes: all (wait for all branches), race (first branch wins).

wait

Delays execution or waits for a named signal.

{ "id": "pause", "type": "wait", "config": { "duration": "5s" } }

reasoning

Suspends workflow for agent decision. Empty options = free-form (any choice accepted).

{
  "id": "review",
  "type": "reasoning",
  "config": {
    "prompt_context": "Review data and decide",
    "options": [
      { "id": "approve", "description": "Proceed" },
      { "id": "reject", "description": "Stop" }
    ],
    "data_inject": { "analysis": "steps.analyze.output" },
    "timeout": "1h",
    "fallback": "reject",
    "target_agent": ""
  }
}

Variable Interpolation

Syntax: ${{namespace.path}}

Two-pass resolution: non-secrets first, then secrets via AES-256-GCM vault.

CEL gotcha: loop is a reserved word in CEL. Use iter.item / iter.index in CEL expressions. The ${{loop.item}} interpolation syntax is unaffected.

See expressions.md for CEL, GoJQ, Expr engine details.

Built-in Actions

Quick reference (most-used actions):

• http.get: url (req), headers, timeout, fail_on_error_status -- output: { status_code, headers, body, duration_ms }
• shell.exec: command (req), args, stdin, timeout, env, workdir -- output: { stdout, stderr, exit_code, killed }
• fs.read: path (req), encoding -- output: { path, content, encoding, size }
• workflow.notify: message (req), data -- output: { notified: true/false } -- pushes real-time notification to agent via MCP SSE (best-effort)

• expr.eval: expression (req), data -- output: { result: <value> } -- evaluates Expr expression against workflow scope (steps, inputs, workflow, context)

See actions.md for full parameter specs of all 26 actions.

Scripting with shell.exec

shell.exec auto-parses JSON stdout. Convention: stdin=JSON, stdout=JSON, stderr=errors, non-zero exit=failure. Use stdout_raw for unprocessed text.

See patterns.md for language-specific templates (Bash, Python, Node, Go).

Reasoning Node Lifecycle

1. Workflow reaches a reasoning step
2. Executor creates PendingDecision, emits decision_requested event
3. Workflow status becomes suspended
4. Agent calls opcode.status to see pending decision with context and options
5. Agent resolves via opcode.signal:

   {
     "workflow_id": "...",
     "signal_type": "decision",
     "step_id": "reason-step",
     "payload": { "choice": "approve" }
   }

1. Workflow auto-resumes after signal
2. If timeout expires: fallback option auto-selected, or step fails if no fallback

Common Patterns

See patterns.md for full JSON examples: linear pipeline, conditional branching, for-each loop, parallel fan-out, human-in-the-loop, error recovery, sub-workflows, and MCP lifecycle.

Error Handling

Backoff: none, linear, exponential, constant. Non-retryable errors (validation, permission, assertion) are never retried.

See error-handling.md for circuit breakers, timeout interactions, error codes.

Performance

10-step parallel workflows complete in ~50µs, 500-step in ~2.4ms. The event store sustains ~15k appends/sec with <12% drop under 100 concurrent writers. Worker pool overhead is ~0.85µs/task (>1M tasks/sec at any pool size).

Full benchmark charts, per-scenario breakdowns, and methodology: docs/benchmarks.md.