# EffectGuard: coordination for cooperating agents

Check new_paid_purchases_enabled in /.well-known/x402.json before any new purchase.
If false, new paid purchases are paused. Do not connect or sign through a wallet
warning. Existing purchases can still be recovered without a new payment.

Base URL: https://effectguard-x402-experiment.onrender.com
Stage: external pilot. The integrator executes its own tools.

Use cases:
- operation.once: coordinate duplicate writes using one stable business event ID.
- json.read: reuse matching JSON results across agents within 30 seconds.

Integration: /agent
Python SDK: /agent/sdk.zip
HubSpot pilot kit: /agent/hubspot.zip
HubSpot setup: /agent/hubspot.txt
Tester brief: /agent/tester.txt
JSON read guide: /agent/reads
OpenAPI: /openapi.json
Payment metadata: /.well-known/x402.json
Enrollment: POST /v1/enroll with {}. Store the returned API key privately.
Authenticate API calls with Authorization: Bearer <your API key>.

POST /v1/preflight is paid via x402 v2 on Base (eip155:8453), official USDC.
Read the current offer before signing. Current experiment price: 0.01 USDC
per new logical request, including REUSE, WAIT, REVIEW or BLOCK_DUPLICATE.
Recovery of the same purchased request, /v1/start and /v1/commit are free.
Persist request IDs and payment receipts. Never automatically repay an
uncertain settlement; recover the original request first. Enforce a wallet budget.

For operation.once, arguments contain namespace, operation, business_event_id
and parameters. Keep business_event_id stable for retries of one business event.
Use a separate persisted request_id for each logical preflight.
Agents must share the same enrolled tenant/account and operation contract.
Include every input affecting the operation or permissions in parameters.

On EXECUTE, call /v1/start with effect_id and execution_token, then execute
the external tool and call /v1/commit with outcome and result. Reuse confirmed
results for REUSE. Do not execute on BLOCK_DUPLICATE, WAIT or REVIEW.
If the external outcome is uncertain, commit UNKNOWN and reconcile; do not
claim success or blindly repeat the write. Use provider idempotency where available.
EffectGuard does not guarantee exactly-once execution of external services.

HubSpot note-once: the public Python pilot kit includes the tested adapter,
local demo and bounded hosted x402 test. The starter is limited to a dedicated
test contact named EffectGuard Pilot in the buyer's own portal. The owner
completed two paid checks with one HubSpot note POST, then recovered both
results with no new signature or note POST. External customer demand and
savings remain unverified. Quickstart: python -m scripts.crm_agent demo after
installing the kit. Two deterministic workers and a persistent simulated CRM;
no LLM, credentials or payments. Reruns recover the same source event.
python -m scripts.crm_agent hubspot --portal-id ID --contact-id ID prepares and
runs a local test on the buyer's dedicated contact. Explicit STANDARD opt-in
is required when applicable. Optional paid mode uses --accept-paid-test and a
local dedicated test-wallet signer with a durable maximum 0.02 USDC, or MetaMask
via scripts.hubspot_paid_pilot. It is not an unlimited production payment policy.
The feedback command prints a sanitized local summary; no automatic upload.
The kit is not a native MCP/n8n connector. Only the
portal ID and note payload hash reach this API, not the HubSpot token or body.
Use the same enrolled account for cooperating writers. A discovered preflight
is not a complete HubSpot action: install the adapter and enforce start/commit
in your runtime. Never execute HubSpot writes on BLOCK_DUPLICATE, WAIT or REVIEW.

Payment accounting: facilitator-confirmed settlements retain historical amount,
asset, network, receiver and transaction ID. Same-purchase recovery adds no entry.
CSV export is private to the operator. Fees and PLN valuation are not recorded.
Old payments without captured terms are explicitly incomplete.