How to pause your agent for human input

Some tasks reach a decision point the agent cannot resolve on its own — a target environment, an approval threshold, a destructive step that needs sign-off. murmur/tool-request-input gives the model a ready-made tool to pause at that point, surface a question to the operator, and resume exactly where it left off once the operator replies.

The relevant manifest options are:

Option	Controls
artifacts[].runtime	Whether a WASM artifact is a model-visible tool, inference driver, or hook
lifecycle.input_timeout_secs	Maximum seconds to wait before failing the task if no reply arrives

---

Step 1 — write the manifest

Create a murmur.yaml file. Add murmur/tool-request-input to artifacts with runtime: tool, and write a system_prompt that tells the model when to pause:

AnthropicOpenAIDeepSeek

name: my-agent
version: "0.1.0"

network:
  internal_port: 52222

artifacts:
  - name: murmur/driver-anthropic
    version: "0.19.0"
    runtime: driver
    gateway:
      endpoint: https://api.anthropic.com
      api_key: ${ANTHROPIC_API_KEY}
  - name: murmur/tool-request-input
    version: "0.7.0"
    runtime: tool

inference:
  transport: http
  model: claude-sonnet-5
  driver:
    artifact: murmur/driver-anthropic
  system_prompt: |
    You are a deployment agent. When you reach a decision you cannot make on your own —
    such as which environment to target or whether to proceed with a destructive step —
    call tool-request-input with a clear, specific question for the operator.

name: my-agent
version: "0.1.0"

network:
  internal_port: 52222

artifacts:
  - name: murmur/driver-openai
    version: "0.19.0"
    runtime: driver
    gateway:
      endpoint: https://api.openai.com/v1
      api_key: ${OPENAI_API_KEY}
  - name: murmur/tool-request-input
    version: "0.7.0"
    runtime: tool

inference:
  transport: http
  model: o3-mini-high
  driver:
    artifact: murmur/driver-openai
  system_prompt: |
    You are a deployment agent. When you reach a decision you cannot make on your own —
    such as which environment to target or whether to proceed with a destructive step —
    call tool-request-input with a clear, specific question for the operator.

name: my-agent
version: "0.1.0"

network:
  internal_port: 52222

artifacts:
  - name: murmur/driver-deepseek
    version: "0.17.0"
    runtime: driver
    gateway:
      endpoint: https://api.deepseek.com
      api_key: ${DEEPSEEK_API_KEY}
  - name: murmur/tool-request-input
    version: "0.7.0"
    runtime: tool

inference:
  transport: http
  model: deepseek-r1
  driver:
    artifact: murmur/driver-deepseek
  system_prompt: |
    You are a deployment agent. When you reach a decision you cannot make on your own —
    such as which environment to target or whether to proceed with a destructive step —
    call tool-request-input with a clear, specific question for the operator.

network.internal_port pins the capsule to a fixed port on every run. Without it the runtime picks an OS-assigned port at startup, which changes between runs.

runtime: tool registers the WASM component with the capsule runtime. The model sees it as a callable tool named tool-request-input, the part of the artifact name after the /, with one parameter:

Parameter	Type	Required	Description
prompt	string	yes	The question to present to the operator

The system prompt controls when the model pauses. A vague instruction ("ask when unsure") produces over-cautious behavior; a concrete rule ("call the tool before any destructive step") produces predictable escalation.

---

Step 2 — install dependencies

With murmur.yaml in place, fetch all declared artifacts:

mur install

Different ways to install artifacts

mur install fetches an artifact from the source of its scope. murmur/ artifacts come from the built-in source, murmur-nexus/default-artifacts, with no configuration.

Option A — name the source of a scope in ~/.murmur/config.yaml:

registry:
  sources:
    - repo: <owner>/<repo>
      token: "${GITHUB_TOKEN}"

Then install by full name, with or without a version:

mur install <owner>/<artifact-name>@<version>

Option B — pass a full GitHub reference and skip configuration entirely:

mur install github:<owner>/<repo>@<tag>

See Installing artifacts to learn more.

murmur/tool-request-input is a WASM artifact — no platform tag is required. mur install resolves the correct variant for every artifact in the manifest automatically.

---

Step 3 — run the capsule and send a task

Start the capsule:

mur run
murmur: url localhost:52222
session: ses_019ed2af53da75c2aefee84ee10c34af

Note the URL. Set a shell variable for convenience, replacing 52222 with the port you see:

PORT=52222

Send a task that will require a human decision:

curl -s -X POST http://localhost:$PORT \
  -H "Content-Type: application/json" \
  -H "A2A-Version: 1.0" \
  -d '{
    "jsonrpc": "2.0",
    "id": 1,
    "method": "SendMessage",
    "params": {
      "message": {
        "messageId": "msg-001",
        "role": "ROLE_USER",
        "parts": [{"text": "Deploy the latest build. Confirm the target environment before proceeding."}]
      },
      "configuration": {"returnImmediately": true}
    }
  }'

Response:

{
  "jsonrpc": "2.0",
  "id": 1,
  "result": {
    "task": {
      "id": "tsk_01jw...",
      "contextId": "ctx_01jw...",
      "status": { "state": "TASK_STATE_SUBMITTED", "timestamp": "2026-10-13T09:00:00.123Z" }
    }
  }
}

Save the task id at result.task.id: you need it to detect the pause and to answer.

"returnImmediately": true has the door answer as soon as it accepts the task. Without it the door holds the request until the task pauses or ends, and answers with the task as it stands then: here, TASK_STATE_INPUT_REQUIRED with the agent's question in its prompt artifact, which saves the polling in step 4. See Agent Card: Answers.

---

Step 4 — detect when the agent is waiting

Poll GetTask with the task ID until the state changes to TASK_STATE_INPUT_REQUIRED:

curl -s -X POST http://localhost:$PORT \
  -H "Content-Type: application/json" \
  -H "A2A-Version: 1.0" \
  -d '{"jsonrpc":"2.0","id":2,"method":"GetTask","params":{"id":"<your_task_id>"}}'

While the agent is waiting, the response includes the question it formed:

{
  "jsonrpc": "2.0",
  "id": 2,
  "result": {
    "id": "tsk_01jw...",
    "contextId": "ctx_01jw...",
    "status": { "state": "TASK_STATE_INPUT_REQUIRED", "timestamp": "2026-10-13T09:00:02.841Z" },
    "artifacts": [
      {
        "artifactId": "prompt",
        "name": "prompt",
        "parts": [
          {
            "text": "Which environment should I deploy to? (staging, production)",
            "mediaType": "text/plain"
          }
        ]
      }
    ]
  }
}

The agent's question is at result.artifacts[0].parts[0].text. Read it and decide your answer.

---

Step 5 — send your answer

Send a message to the same capsule URL, naming the waiting task in taskId:

curl -s -X POST http://localhost:$PORT \
  -H "Content-Type: application/json" \
  -H "A2A-Version: 1.0" \
  -d '{
    "jsonrpc": "2.0",
    "id": 3,
    "method": "SendMessage",
    "params": {
      "message": {
        "messageId": "reply-001",
        "taskId": "<your_task_id>",
        "role": "ROLE_USER",
        "parts": [{"text": "staging"}]
      },
      "configuration": {"returnImmediately": true}
    }
  }'

The answer is delivered directly to the suspended tool call. The agent receives "staging" as the tool result and the loop resumes immediately:

{
  "jsonrpc": "2.0",
  "id": 3,
  "result": {
    "task": {
      "id": "tsk_01jw...",
      "contextId": "ctx_01jw...",
      "status": { "state": "TASK_STATE_WORKING", "timestamp": "2026-10-13T09:00:09.307Z" }
    }
  }
}

A message without taskId starts a new task instead, and never reaches the waiting one. A reply naming a task that is not waiting is refused: see Continuing a task.

Poll GetTask again until state reaches TASK_STATE_COMPLETED or TASK_STATE_FAILED.

---

Optional — set an input timeout

To fail the task automatically if no reply arrives within a deadline, add lifecycle.input_timeout_secs to the manifest:

lifecycle:
  task_acceptance: single
  input_timeout_secs: 300

When the deadline passes, the tool call fails and the attempt ends. The task then ends TASK_STATE_FAILED with status message "input-timeout", unless an on-task-end hook reopens it. Omit the field to wait indefinitely.

---

Summary

Feature / setting	How it works
murmur/tool-request-input	WASM tool artifact; runtime: tool; platform-independent
prompt parameter	The question the model asks the operator; string, required
Task state while waiting	"TASK_STATE_INPUT_REQUIRED"
Where to read the question	result.artifacts[0].parts[0].text from GetTask
How to resume the agent	Send a message to the same capsule URL whose taskId names the waiting task
State after reply	"TASK_STATE_WORKING" immediately; poll until "TASK_STATE_COMPLETED"
lifecycle.input_timeout_secs	Integer seconds to wait for a reply; absent = wait indefinitely