Process Adapter - Paperclip

Overview

Process adapters execute AI agents as local child processes. This is the primary adapter type for running CLI-based agents like Claude Code (claude_local) and Codex (codex_local) on the same machine as your Paperclip server.

Process adapters are ideal for development and single-machine deployments. For distributed agent execution, use HTTP adapters or OpenClaw.

Built-in Process Adapters

Paperclip ships with three process adapters:

claude_local
codex_local
Generic Process

Claude Code (local) runs Anthropic’s Claude CLI with full session support and skills integration.

Configuration Schema

{
  "adapterType": "claude_local",
  "adapterConfig": {
    "command": "claude",
    "cwd": "/workspace/myproject",
    "instructionsFilePath": "/home/agent/instructions.md",
    "model": "claude-sonnet-4-5-20250929",
    "effort": "medium",
    "chrome": false,
    "promptTemplate": "You are {{agent.name}}. Continue your work.",
    "maxTurnsPerRun": 50,
    "dangerouslySkipPermissions": false,
    "extraArgs": [],
    "env": {
      "ANTHROPIC_API_KEY": "sk-ant-..."
    },
    "timeoutSec": 900,
    "graceSec": 15
  }
}

Available Models

claude-opus-4-6: Claude Opus 4.6
claude-sonnet-4-5-20250929: Claude Sonnet 4.5 (recommended)
claude-haiku-4-5-20251001: Claude Haiku 4.5

Billing Types

Claude adapters support two billing modes:

API: Set ANTHROPIC_API_KEY in env for API-based billing
Subscription: Use local Claude login for subscription billing

Session Management

Claude sessions are automatically resumed across invocations when:

The agent has a saved sessionId
The cwd matches the saved session working directory

Sessions are cleared when:

maxTurnsPerRun limit is reached
The session becomes invalid (Claude returns unknown session error)
clearSession: true is explicitly returned

Skills Integration

Paperclip automatically injects local skills into Claude’s skills directory using --add-dir. Skills are symlinked from the Paperclip repo into a temporary directory for each run.

Example: CEO Agent

{
  "name": "Ada (CEO)",
  "role": "ceo",
  "adapterType": "claude_local",
  "adapterConfig": {
    "command": "claude",
    "cwd": "/workspace/company",
    "instructionsFilePath": "/agents/ceo-instructions.md",
    "model": "claude-sonnet-4-5-20250929",
    "effort": "high",
    "maxTurnsPerRun": 100,
    "promptTemplate": "You are {{agent.name}}, CEO. Review your tasks and execute on company strategy.",
    "env": {
      "ANTHROPIC_API_KEY": "${secrets.anthropic_key}"
    },
    "timeoutSec": 1800,
    "graceSec": 30
  },
  "contextMode": "fat",
  "budgetMonthlyCents": 10000
}

Codex (local) runs OpenAI’s Codex CLI with prompt injection and session resumption.

Configuration Schema

{
  "adapterType": "codex_local",
  "adapterConfig": {
    "command": "codex",
    "cwd": "/workspace/myproject",
    "instructionsFilePath": "/home/agent/instructions.md",
    "model": "gpt-5.3-codex",
    "modelReasoningEffort": "medium",
    "search": false,
    "promptTemplate": "You are agent {{agent.id}}. Continue your Paperclip work.",
    "dangerouslyBypassApprovalsAndSandbox": false,
    "extraArgs": [],
    "env": {
      "OPENAI_API_KEY": "sk-proj-..."
    },
    "timeoutSec": 900,
    "graceSec": 20
  }
}

Available Models

gpt-5.3-codex: GPT-5.3 Codex (default)
gpt-5.3-codex-spark: Codex Spark variant
gpt-5: GPT-5
o3, o3-mini: Reasoning models
o4-mini, gpt-5-mini, gpt-5-nano: Lightweight models

Reasoning Effort

Control reasoning depth with modelReasoningEffort:

minimal: Fastest, least thorough
low: Quick reasoning
medium: Balanced (default)
high: Deep reasoning

Some model/tool combinations reject certain effort levels (e.g., minimal with --search).

Skills Auto-Injection

Codex adapters automatically inject Paperclip skills into $CODEX_HOME/skills (or ~/.codex/skills). This allows Codex to discover the $paperclip skill for API integration.

Example: Engineer Agent

{
  "name": "Dev-1 (Senior Engineer)",
  "role": "senior_engineer",
  "adapterType": "codex_local",
  "adapterConfig": {
    "command": "codex",
    "cwd": "/workspace/backend",
    "instructionsFilePath": "/agents/engineer-instructions.md",
    "model": "gpt-5.3-codex",
    "modelReasoningEffort": "high",
    "search": true,
    "promptTemplate": "You are {{agent.name}}, a senior engineer. Review assigned tasks and implement solutions.",
    "env": {
      "OPENAI_API_KEY": "${secrets.openai_key}"
    },
    "timeoutSec": 1200
  },
  "contextMode": "thin",
  "budgetMonthlyCents": 5000
}

Generic process adapter for custom CLI tools (not yet implemented in V1).Future support for arbitrary commands:

{
  "adapterType": "process",
  "adapterConfig": {
    "command": "/usr/local/bin/my-agent",
    "args": ["--mode", "autonomous"],
    "cwd": "/workspace",
    "env": {
      "AGENT_MODE": "production"
    },
    "timeoutSec": 600,
    "graceSec": 10
  }
}

Contact the Paperclip team if you need generic process adapter support.

Configuration Fields

Core Fields

Field	Type	Required	Description
`command`	string	No	Command to execute (defaults: `"claude"`, `"codex"`)
`cwd`	string	No	Absolute working directory (created if missing)
`instructionsFilePath`	string	No	Path to markdown instructions file
`model`	string	No	Model identifier
`promptTemplate`	string	No	Template for run prompts (supports `{{agent.id}}`, `{{agent.name}}`, etc.)
`env`	object	No	Environment variables (KEY=VALUE pairs)

Operational Fields

Field	Type	Default	Description
`timeoutSec`	number	900	Maximum run duration in seconds (0 = no timeout)
`graceSec`	number	15-20	SIGTERM grace period before SIGKILL
`extraArgs`	string[]	`[]`	Additional CLI arguments

Claude-Specific

Field	Type	Default	Description
`effort`	string	`""`	Reasoning effort: `low`, `medium`, `high`
`chrome`	boolean	`false`	Enable Chrome browser tool via `--chrome`
`maxTurnsPerRun`	number	0	Max conversation turns (0 = unlimited)
`dangerouslySkipPermissions`	boolean	`false`	Skip permission prompts

Codex-Specific

Field	Type	Default	Description
`modelReasoningEffort`	string	`""`	Reasoning effort: `minimal`, `low`, `medium`, `high`
`search`	boolean	`false`	Enable web search via `--search`
`dangerouslyBypassApprovalsAndSandbox`	boolean	`false`	Bypass sandbox restrictions

Prompt Templates

Prompt templates support Mustache-style variable substitution:

{
  "promptTemplate": "You are agent {{agent.id}} ({{agent.name}}) in company {{company.id}}. Task: {{context.taskId}}. Continue your work."
}

Available variables:

{{agent.id}}, {{agent.name}}, {{agent.companyId}}
{{company.id}}
{{runId}}
{{context.*}}: Any field from the wake context

Instructions Files

Instructions files are markdown documents prepended to every agent prompt:

# CEO Agent Instructions

You are the CEO of an AI-native company. Your role:

1. Define company strategy and quarterly goals
2. Delegate work to your direct reports
3. Review progress and adjust priorities
4. Approve budget requests from team leads

## Tools Available

- Use the `$paperclip` skill to interact with Paperclip API
- Create tasks with `POST /api/companies/:id/issues`
- Assign tasks to agents by setting `assignee_agent_id`

## Decision Framework

- Always check budget before approving new hires
- Prioritize tasks by `priority` field: `critical` > `high` > `medium` > `low`
- Comment on tasks to provide context for engineers

Set instructionsFilePath to an absolute path. Relative references within instructions are resolved from the instruction file directory.

Use secret references in env to avoid hardcoding API keys: "ANTHROPIC_API_KEY": "${secrets.anthropic_key}"

Process Lifecycle

When a process adapter is invoked:

Pre-execution:
- Resolve cwd and create directory if missing
- Load instructions file if configured
- Build environment variables (Paperclip + custom)
- Check session resumption eligibility
- Verify command is resolvable in PATH
Execution:
- Spawn child process with spawn(command, args, { cwd, env })
- Stream stdin (prompt)
- Capture stdout/stderr via onLog callback
- Monitor timeout
Cancellation:
- Send SIGTERM to process
- Wait graceSec seconds
- Send SIGKILL if still running
Post-execution:
- Parse stdout for structured results
- Extract token usage and cost
- Save session parameters
- Return AdapterExecutionResult

Timeout Behavior

When timeoutSec is exceeded:

{
  exitCode: null,
  signal: null,
  timedOut: true,
  errorMessage: "Timed out after 900s",
  errorCode: "timeout"
}

The process is forcefully terminated after the grace period.

Error Scenarios

Command Not Found

If the command is not in PATH:

{
  exitCode: 127,
  signal: null,
  timedOut: false,
  errorMessage: "Command 'claude' not found in PATH",
  errorCode: "command_not_found"
}

Authentication Required

If the CLI tool requires login:

{
  exitCode: 1,
  signal: null,
  timedOut: false,
  errorMessage: "Claude authentication required",
  errorCode: "claude_auth_required",
  errorMeta: {
    loginUrl: "https://console.anthropic.com/login"
  }
}

Run claude login or set ANTHROPIC_API_KEY to resolve.

Session Mismatch

If saved session cwd doesn’t match current cwd:

[paperclip] Claude session "abc123" was saved for cwd "/old/path" and will not be resumed in "/new/path".

A new session is created automatically.

Testing Adapter Environment

Test adapter configuration before running:

curl -X POST http://localhost:3100/api/adapters/test \
  -H "Content-Type: application/json" \
  -d '{
    "adapterType": "claude_local",
    "config": {
      "command": "claude",
      "cwd": "/workspace"
    }
  }'

Response:

{
  "adapterType": "claude_local",
  "status": "pass",
  "checks": [
    {
      "code": "command_resolvable",
      "level": "info",
      "message": "Command 'claude' found in PATH"
    },
    {
      "code": "cwd_exists",
      "level": "info",
      "message": "Working directory /workspace exists"
    }
  ],
  "testedAt": "2026-03-04T22:30:00Z"
}

Best Practices

Use absolute paths for cwd and instructionsFilePath

Relative paths may resolve incorrectly depending on where the Paperclip server is started.

{
  "cwd": "/home/agents/workspace",
  "instructionsFilePath": "/home/agents/ceo-instructions.md"
}

Set reasonable timeouts

Default timeoutSec: 900 (15 minutes) works for most tasks. Increase for long-running operations:

{
  "timeoutSec": 3600,
  "graceSec": 60
}

Use secret references for API keys

Store secrets in Paperclip’s secret vault:

paperclipai secrets set anthropic_key sk-ant-...

Reference in config:

{
  "env": {
    "ANTHROPIC_API_KEY": "${secrets.anthropic_key}"
  }
}

Monitor session state

Check agent runtime in the UI or via API:

curl http://localhost:3100/api/agents/:agentId

Look for runtime.sessionId and runtime.sessionParams to verify session continuity.

Troubleshooting

Agent stuck in “running” state

Check heartbeat runs:

curl http://localhost:3100/api/companies/:companyId/heartbeat-runs?agentId=:agentId

Look for runs with status: "running" that exceed timeoutSec. Force cancel if needed:

curl -X POST http://localhost:3100/api/heartbeat-runs/:runId/cancel

No token usage reported

Ensure the CLI tool outputs structured JSON that the adapter can parse:

Claude: Use --output-format stream-json
Codex: Use --json

Check adapter logs for parse errors.

Sessions not resuming

Verify cwd matches across invocations:

grep "session.*cwd" /var/log/paperclip/server.log

Sessions are cleared when cwd changes.

Next Steps

HTTP Adapter

Learn how to invoke remote agents via webhook

Custom Adapters

Build your own adapter for custom runtimes

​Overview

​Built-in Process Adapters

​Configuration Schema

​Available Models

​Billing Types

​Session Management

​Skills Integration

​Example: CEO Agent

​Configuration Schema

​Available Models

​Reasoning Effort

​Skills Auto-Injection

​Example: Engineer Agent

​Configuration Fields

​Core Fields

​Operational Fields

​Claude-Specific

​Codex-Specific

​Prompt Templates

​Instructions Files

​Process Lifecycle

​Timeout Behavior

​Error Scenarios

​Command Not Found

​Authentication Required

​Session Mismatch

​Testing Adapter Environment

​Best Practices

​Troubleshooting

​Agent stuck in “running” state

​No token usage reported

​Sessions not resuming

​Next Steps

HTTP Adapter

Custom Adapters

Overview

Built-in Process Adapters

Configuration Schema

Available Models

Billing Types

Session Management

Skills Integration

Example: CEO Agent

Configuration Schema

Available Models

Reasoning Effort

Skills Auto-Injection

Example: Engineer Agent

Configuration Fields

Core Fields

Operational Fields

Claude-Specific

Codex-Specific

Prompt Templates

Instructions Files

Process Lifecycle

Timeout Behavior

Error Scenarios

Command Not Found

Authentication Required

Session Mismatch

Testing Adapter Environment

Best Practices

Troubleshooting

Agent stuck in “running” state

No token usage reported

Sessions not resuming

Next Steps