openhands-api

This skill documents the OpenHands Cloud API (V1) and provides small, easy-to-copy clients.

It is intentionally focused on common OpenHands Cloud workflows:

• Defaults to OpenHands Cloud (https://app.all-hands.dev).
• Targets the V1 app server REST API under /api/v1/....
• Includes a few agent server endpoints (inside a sandbox) that use X-Session-API-Key.
• Covers the multi-conversation delegation pattern: start separate Cloud conversations when you want fresh context windows or background work.

When to use this skill

Use this skill when you need to:

• start or inspect OpenHands Cloud conversations from code
• monitor async startup via start-task polling
• monitor execution status for long-running jobs
• create separate Cloud conversations for parallel or background work
• access sandbox agent-server endpoints once a conversation is running

Auth

App server (Cloud)

Use Bearer auth:

• Header: Authorization: Bearer <OPENHANDS_CLOUD_API_KEY>
• Preferred env var: OPENHANDS_CLOUD_API_KEY
• Backward-compatible env var: OPENHANDS_API_KEY

Agent server (inside a sandbox)

Use session auth:

• Header: X-Session-API-Key: <session_api_key>

How to obtain agent_server_url and session_api_key:

1. Start or fetch an app conversation via the app server (Bearer auth), e.g.:
   • POST /api/v1/app-conversations
   • or GET /api/v1/app-conversations?ids=<conversation_id>
2. In the returned JSON, look for sandbox/runtime connection fields (names vary slightly by deployment/version). Common patterns:
   • a sandbox object containing agent_server_url (or similar)
   • a session key such as session_api_key (or similar)
3. Use those values to call the agent server directly:
   • Base: {agent_server_url}/api/...
   • Header: X-Session-API-Key: <session_api_key>

Example (common field names; adjust to your deployment):

# using the minimal Python client (`OpenHandsAPI`)
conv = api.app_conversation_get(app_conversation_id)

session_api_key = conv.get("session_api_key")
conversation_url = conv.get("conversation_url", "")

# `conversation_url` often looks like: https://<runtime-host>/api/conversations/<id>
agent_server_url = conversation_url.rsplit("/api/conversations", 1)[0]

If those fields are not present on the conversation record, list/search sandboxes (GET /api/v1/sandboxes/search) and use the sandbox referenced by the conversation to locate the agent server URL + session key.

Common V1 app server endpoints

The following are the main endpoints implemented in the minimal client:

• GET /api/v1/users/me — validate auth and inspect current account
• GET /api/v1/app-conversations/search?limit=... — list recent conversations
• GET /api/v1/app-conversations?ids=... — fetch conversation records by id (batch)
• GET /api/v1/app-conversations/count — count conversations
• POST /api/v1/app-conversations — start a new conversation (creates a sandbox)
• GET /api/v1/app-conversations/start-tasks?ids=... — check async start-task status
• GET /api/v1/conversation/{app_conversation_id}/events/search?limit=... — read conversation events
• GET /api/v1/conversation/{app_conversation_id}/events/count — count events
• GET /api/v1/sandboxes/search?limit=... — list sandboxes
• POST /api/v1/sandboxes/{sandbox_id}/pause / .../resume — manage sandbox lifecycle
• GET /api/v1/app-conversations/{app_conversation_id}/download — download trajectory zip

Delegating work with additional Cloud conversations

Use the Cloud API when you want a separate OpenHands conversation with its own fresh context window. This is useful for:

• background jobs that can run independently
• parallel investigations or implementation tasks
• long-running work where you want to keep the current conversation focused
• task-specific contexts, such as one conversation building a component while another runs tests

Delegation checklist

When you start a delegated Cloud conversation:

1. Write a self-contained task description. Do not assume the new conversation has any context from the current one.
2. Include the repository, branch, relevant file paths, constraints, and expected output.
3. Start the new conversation with POST /api/v1/app-conversations.
4. Poll the start-task until status is READY and you have an app_conversation_id.
5. Monitor the delegated conversation via GET /api/v1/app-conversations?ids=....
6. Share or store the Cloud URL: https://app.all-hands.dev/conversations/<app_conversation_id>.

Minimal cURL flow

curl -X POST "https://app.all-hands.dev/api/v1/app-conversations" \
  -H "Authorization: Bearer ${OPENHANDS_CLOUD_API_KEY}" \
  -H "Content-Type: application/json" \
  -d '{
    "initial_message": {
      "content": [{"type": "text", "text": "Investigate flaky tests in tests/test_api.py. Report the root cause and propose a fix."}]
    },
    "selected_repository": "owner/repo"
  }'

If the response does not already include app_conversation_id, poll the start-task:

curl -s "https://app.all-hands.dev/api/v1/app-conversations/start-tasks?ids=${START_TASK_ID}" \
  -H "Authorization: Bearer ${OPENHANDS_CLOUD_API_KEY}"

Then check execution status:

curl -s "https://app.all-hands.dev/api/v1/app-conversations?ids=${APP_CONVERSATION_ID}" \
  -H "Authorization: Bearer ${OPENHANDS_CLOUD_API_KEY}"

Minimal Python flow

from openhands_api import OpenHandsAPI

api = OpenHandsAPI()  # prefers OPENHANDS_CLOUD_API_KEY

start = api.app_conversation_start(
    initial_message=(
        "Implement the requested dashboard component in src/dashboard.tsx. "
        "Update any related tests and summarize the changes."
    ),
    selected_repository="owner/repo",
    selected_branch="main",
    title="Dashboard component task",
)

ready = start
if not ready.get("app_conversation_id"):
    ready = api.poll_start_task_until_ready(start["id"])

conversation_id = ready["app_conversation_id"]
print(f"Delegated conversation: {api.base_url}/conversations/{conversation_id}")

status = api.app_conversation_get(conversation_id)
print(status.get("sandbox_status"), status.get("execution_status"))

api.close()

Parallelism guidance

• Prefer 5 or fewer concurrently running delegated conversations.
• Before starting more, check recent conversations and count how many are still execution_status == "running".
• Batch specific conversation lookups with GET /api/v1/app-conversations?ids=... when you already know their ids.

Example:

items = api.app_conversations_search(limit=50).get("items", [])
running = [item for item in items if item.get("execution_status") == "running"]
if len(running) >= 5:
    print("Wait for some delegated conversations to finish before starting more.")

Start-task vs app_conversation_id (common pitfall)

In many deployments, POST /api/v1/app-conversations is asynchronous and returns a start-task object:

• id is the start_task_id
• app_conversation_id is the id you should use for conversation operations like:
  • GET /api/v1/app-conversations/{app_conversation_id}/download
  • GET /api/v1/conversation/{app_conversation_id}/events/...

If app_conversation_id is not present in the initial response, fetch it via:

• GET /api/v1/app-conversations/start-tasks?ids=<start_task_id>

If you pass a start_task_id to /download, you will get 404 Not Found.

Common agent server endpoints

These run against agent_server_url (not the app server):

• POST {agent_server_url}/api/bash/execute_bash_command
• GET {agent_server_url}/api/file/download/<absolute_path>
• POST {agent_server_url}/api/file/upload/<absolute_path> (multipart)
• GET {agent_server_url}/api/conversations/{conversation_id}/events/search
• GET {agent_server_url}/api/conversations/{conversation_id}/events/count

Counting events (recommended approach)

If you need to know how many events a conversation has, you can:

1. App server count (fastest when working)
   • GET /api/v1/conversation/{app_conversation_id}/events/count
2. Agent server count (reliable fallback)
   • GET {agent_server_url}/api/conversations/{app_conversation_id}/events/count
3. Trajectory zip fallback (heavier, but still one call + gives full payloads)
   • GET /api/v1/app-conversations/{app_conversation_id}/download
   • Unzip and count event_*.json files

Do not rely on the last event id to infer the total number of events. In the agent-server API, event IDs are UUIDs (not monotonically increasing integers).

Troubleshooting

For common issues and solutions, see TROUBLESHOOTING.md.

Event structure (for debugging)

Events returned by:

• app server: GET /api/v1/conversation/{id}/events/search
• agent server: GET {agent_server_url}/api/conversations/{id}/events/search

…share the same high-level shape.

Each event typically includes:

• id (UUID)
• timestamp
• kind
• source

Common kind values:

Linking tool calls:

• ActionEvent.tool_call_id == ObservationEvent.tool_call_id
• ObservationEvent.action_id == ActionEvent.id

Example (simplified):

{
  "id": "<action-event-uuid>",
  "kind": "ActionEvent",
  "source": "agent",
  "tool_name": "terminal",
  "tool_call_id": "toolu_...",
  "action": {"command": "ls"}
}

{
  "id": "<observation-event-uuid>",
  "kind": "ObservationEvent",
  "source": "environment",
  "tool_name": "terminal",
  "tool_call_id": "toolu_...",
  "action_id": "<action-event-uuid>",
  "observation": {"exit_code": 0, "stdout": "..."}
}

Debugging one-liners (events)

These assume you're querying the app server endpoint. For agent-server queries, swap the URL base + use X-Session-API-Key.

Print a quick timeline

curl -s "${BASE_URL:-https://app.all-hands.dev}/api/v1/conversation/${APP_CONVERSATION_ID}/events/search?limit=100" \
  -H "Authorization: Bearer ${OPENHANDS_CLOUD_API_KEY:-$OPENHANDS_API_KEY}" \
  -H "Accept: application/json" | \
python3 - <<'PY'
import json, sys
items = (json.load(sys.stdin) or {}).get("items", [])
for i, e in enumerate(items):
    print(f"{i:04d}  {e.get('timestamp','')}  {e.get('source','')}  {e.get('kind','')}")
PY

Find error-like events

curl -s "${BASE_URL:-https://app.all-hands.dev}/api/v1/conversation/${APP_CONVERSATION_ID}/events/search?limit=200" \
  -H "Authorization: Bearer ${OPENHANDS_CLOUD_API_KEY:-$OPENHANDS_API_KEY}" \
  -H "Accept: application/json" | \
python3 - <<'PY'
import json, sys
items = (json.load(sys.stdin) or {}).get("items", [])
for i, e in enumerate(items):
    if e.get("kind") == "ErrorEvent" or ("code" in e and "detail" in e):
        print(i, e.get("kind"), e.get("code"), str(e.get("detail", ""))[:400])
PY

Check tool-call matching (unmatched actions / duplicate observations)

curl -s "${BASE_URL:-https://app.all-hands.dev}/api/v1/conversation/${APP_CONVERSATION_ID}/events/search?limit=200" \
  -H "Authorization: Bearer ${OPENHANDS_CLOUD_API_KEY:-$OPENHANDS_API_KEY}" \
  -H "Accept: application/json" | \
python3 - <<'PY'
import json, sys
from collections import Counter
items = (json.load(sys.stdin) or {}).get("items", [])
action_ids = {e.get("id") for e in items if e.get("kind") == "ActionEvent"}
obs_action_ids = [e.get("action_id") for e in items if e.get("kind") == "ObservationEvent" and e.get("action_id")]
observed = set(obs_action_ids)
print("actions:", len(action_ids))
print("observations:", len(observed))
unmatched = action_ids - observed
print("unmatched actions:", list(unmatched)[:20] if unmatched else "none")
dups = [aid for aid, c in Counter(obs_action_ids).items() if c > 1]
print("duplicate observation action_ids:", list(dups)[:20] if dups else "none")
PY

Quick start (Python)

# Copy `skills/openhands-api/scripts/openhands_api.py` into your project (e.g. as `openhands_api.py`),
# then import it normally:
from openhands_api import OpenHandsAPI

api = OpenHandsAPI()  # prefers OPENHANDS_CLOUD_API_KEY

me = api.users_me()
print(me)

recent = api.app_conversations_search(limit=5)
print(recent)

api.close()

CLI examples

Search conversations:

export OPENHANDS_CLOUD_API_KEY="..."
python skills/openhands-api/scripts/openhands_api.py search-conversations --limit 5

Start a conversation from a prompt file:

python skills/openhands-api/scripts/openhands_api.py start-conversation \
  --prompt-file skills/openhands-api/references/example_prompt.md \
  --repo owner/repo \
  --branch main

Notes for AI agents extending this client

• Prefer .../search endpoints with a small limit.
• Avoid loops that could generate many API calls.
• Start conversations only when asked: it may create sandboxes and cost money.
• For sandbox file operations and command execution, use the agent server endpoints with X-Session-API-Key.

See also:

• skills/openhands-api/scripts/openhands_api.py
• The original inspiration client: enyst/llm-playground → openhands-api-client-v1/scripts/cloud_api_v1.py
• Troubleshooting content and real-world usage feedback → https://github.com/jpshackelford/.openhands/tree/main/skills/openhands-cloud-api