Skip to content

cloud-agent-operations.md

Latest commit

 

History

History
99 lines (82 loc) · 4.79 KB

cloud-agent-operations.md

File metadata and controls

99 lines (82 loc) · 4.79 KB
Raw

Hosted Cloud Agent Operations

Agent Bounties can turn an unstructured digital-work objective into measurable draft bounty terms without running a model on a contributor's computer.

Runtime Boundary

  • POST /v1/cloud-agent/bounty-drafts runs in the hosted API service.
  • draft_bounty_with_cloud_agent in MCP proxies that API; the MCP service does not receive the model credential.
  • GET /v1/cloud-agent/readiness reports the provider, model, public access, limits, missing configuration, and the fact that there is no local fallback.
  • The hosted endpoints remain available for explicit API, evaluation, and objective-compilation workflows. They are not the default website composer.
  • The default website flow hands the request to the person's own ChatGPT, Claude, Gemini, or other MCP-capable AI account and validates the returned draft locally. It never receives the person's provider credential.

Cloud output is untrusted draft data. It cannot publish terms, hold a key, request a wallet signature, fund a contract, verify a submission, settle a bounty, or prove payment. The creator must review the draft, select an executable verifier, and publish and fund through the canonical protocol.

Provider Configuration

The adapter supports OpenAI Responses, OpenAI-compatible chat completions, and Anthropic Messages. Production uses GPT-5.6 through the Responses API with a strict JSON Schema for every generation path: objective graphs, bounty drafts, demo solutions, and published-terms analysis. Flexible nested benchmark and evidence objects cross the model boundary as bounded JSON strings and are decoded and validated before they enter a public response. Only agent-bounties-api receives CLOUD_AGENT_API_KEY. The Blueprint declares it directly on that service with sync: false; Render does not support sync: false inside environment groups. The exact-SHA deployment controller reconciles all nonsecret settings and, when the repository Actions secret is present, copies the model credential directly to the API service without including its value in deployment evidence.

Required production settings:

CLOUD_AGENT_ENABLED=true
CLOUD_AGENT_PUBLIC_DRAFTS=true
CLOUD_AGENT_PROVIDER=openai
CLOUD_AGENT_PROTOCOL=openai_responses
CLOUD_AGENT_ENDPOINT=https://api.openai.com/v1/responses
CLOUD_AGENT_MODEL=gpt-5.6-luna
CLOUD_AGENT_REASONING_EFFORT=low
CLOUD_AGENT_API_KEY=<Render secret>
CLOUD_AGENT_MAX_OUTPUT_TOKENS=12000
CLOUD_AGENT_MAX_DAILY_DRAFTS=50
CLOUD_AGENT_TIMEOUT_SECONDS=90

Store the same credential as the repository Actions secret CLOUD_AGENT_API_KEY, then dispatch Render Deploy Recovery. Existing Render Blueprints ignore newly added sync: false placeholders, so the deployment controller is the repeatable configuration path. Rotating the Actions secret and dispatching again rotates the API-service value and forces a redeploy. The controller then fails closed unless readiness reports hosted execution, draft-only authority, no local fallback, and available: true.

The default public quota is 50 fresh drafts or objective plans per UTC day per API process. GPT-5.6 Luna runs at low reasoning effort with a 12,000-token combined reasoning/output ceiling and a 90-second provider timeout. This leaves enough headroom for strict multi-task JSON while bounding spend and interactive wait. Idempotent retries return the cached draft and do not consume another model call. Inputs, outputs, timeout, arrays, URLs, and idempotency keys are bounded.

Failure Policy

  • Missing credentials or model configuration: readiness is unavailable and drafting returns 503; manual exact-term entry remains available.
  • Invalid objective: structured 400; no terms are published.
  • Model output that fails deterministic validation: structured 502; objective plans receive one bounded repair attempt first; no terms are published.
  • Daily quota exhausted: 429; no local process or local model is invoked.
  • Provider outage: 503; no wallet or protocol transition occurs.

No cloud-model response is allowed to authorize settlement. Deterministic verifier modules or a precommitted verifier quorum remain the only verification authorities, and only a confirmed canonical BountySettled event proves payment.

Cloud-Only Availability

Render runs the API, MCP service, Postgres, and Base indexer. GitHub Actions runs the scheduled inventory, verifier, relay, and deployment-control loops. GitHub Pages serves agentbounties.app. Turning off a maintainer workstation must not affect any of those paths.

Verify after each production deployment:

curl https://api.agentbounties.app/health
curl https://api.agentbounties.app/v1/cloud-agent/readiness
curl "https://api.agentbounties.app/v1/base/autonomous-bounties/inventory-summary?network=base-mainnet&claimable_only=true"
curl https://mcp.agentbounties.app/tools