Weekly field notes: recover safely, test the send, curate memory
A dropped connection, an email test, and a memory write all ask the same operational question: what evidence proves the bot did the right thing? This issue links three live guides and gives a small check you can run for each boundary.
AI-assisted editorial: researched and drafted with AI using the primary sources linked below. Examples are illustrative; they are not measured product results. Editorial policy.
Key takeaways
- After an interrupted MCP tool call, check durable state or an application-level idempotency key before retrying a side effect.
- Use a provider test mode to check request handling, then a dedicated test inbox to verify rendering and actual delivery without contacting customers.
- Persist only useful facts with a clear owner and write rule; keep read-only reference material separate from facts the bot may change.
An interrupted tool call needs a result check
Our MCP retry guide follows the 2026-07-28 transport change: Streamable HTTP no longer provides Last-Event-ID replay for a broken response stream. A client must send a new request with a new request ID. That transport rule says nothing about whether a side-effecting tool already changed its destination before the connection failed.
Try a small test call that creates a record, then interrupt the response. Before retrying, read the destination by a stable operation key. If the record exists, reconcile that result; if it does not, retry with an application-level idempotency key that the server actually enforces. A new JSON-RPC request ID alone does not deduplicate the business action.
Sources: MCP 2026-07-28 key changes, MCP Streamable HTTP transport.
Separate request validation from a real test send
Our email-routine guide compares two provider test paths. SendGrid sandbox mode validates the shape of a Mail Send request without delivery or event records. Mailgun test mode accepts a message without sending it. Neither mode can show how the rendered message appears in a mailbox or prove arrival.
First assert that the bot selected the intended recipient, template, and approval state in a dry run. Then use the provider test mode appropriate to your transport. Finally send one end-to-end message to a dedicated mailbox you control and inspect the rendered content, links, reply-to, and delivery evidence. Keep real customer addresses out of the test.
Sources: SendGrid sandbox mode, Mailgun sending in test mode.
Make memory a deliberate write
Our memory guide asks which facts are useful after the next session starts. Anthropic’s memory tool is implemented by the application: Claude requests file operations, while the application decides where files live and enforces the allowed path. That makes the storage and write policy part of the product, not an automatic guarantee from the model.
For a small exercise, give a bot one read-only reference document and one writable task note. Ask it to save a durable preference, then feed it an untrusted tool result that asks for a conflicting memory update. Verify that only the approved note can change, that the stored fact records its source, and that a human can correct or delete it later.
Sources: Claude memory tool.
The public BotBento update
The three linked guides are live in the BotBento blog. Together they offer a practical sequence: verify an interrupted action, prove an email reached only a test destination, and review what the bot is allowed to remember. Each guide includes the primary documentation behind its factual claims and an illustrative exercise you can adapt.
BotBento remains a pre-release workspace. These articles describe patterns to test in your own systems; this issue makes no claim that BotBento runtime retries, email automation, or persistent bot memory are generally available. Newsletter email is sent only to confirmed subscribers after this archive is deployed and the private delivery checks pass.
Primary sources
Sources checked 2026-10-16. Standards and product documentation can change; follow the linked version when implementing.
- MCP 2026-07-28 key changes — Model Context Protocol
- MCP Streamable HTTP transport — Model Context Protocol
- SendGrid sandbox mode — Twilio SendGrid
- Mailgun sending in test mode — Mailgun
- Claude memory tool — Anthropic
BotBento is in development. Suggest a correction.