// knotbind notes
The API changed. The assistant did not know.

The short answer
An assistant does not notice that an API changed. It keeps sending the request it learned and reading the answer it expects. Knotbind runs scheduled Checks against the exact calls your integration makes, opens a Finding when two runs fail or a schema change breaks the contract, diagnoses the cause, drafts a repair, verifies it in your Sandbox, and applies it only after the Owner approves.
Ask anyone who shipped an AI assistant last year what went wrong and you will rarely hear "the model was bad". You will hear something quieter: the integration worked on Tuesday and was quietly wrong on Wednesday. Nobody deployed. Nobody announced anything. The API behind the assistant changed.
Knotbind exists for that Wednesday. This is the first post in a series on the failure modes we watch, why they are hard to see, and what the product actually does about each one.
What a silent break looks like
An assistant calls a tool. The tool answers with HTTP 200 and a body. Nothing errored, so nothing alerted. But the field the model was taught to read is gone, renamed, or now means something different — so the model reads empty space and keeps talking. The customer gets a confident, plausible answer that is quietly wrong. The logs show success.
The changes that do this are ordinary. Four come up over and over:
- A renamed field.
customer_emailbecomesemail. The response still parses; the value the assistant reads is now undefined. - A new required parameter. Calls that used to succeed now return 400. The assistant retries what it was told to send.
- A removed or renamed tool. In MCP, the agent discovers tools from the server's schema. Remove one from the schema and every client that had it cached is calling something that no longer exists.
- An auth change. A token endpoint moves, a client registration breaks, and onboarding that worked in demo does not work in production.
Why a status code is not the signal
The instinct is to alert on 5xx. But most integration breaks do not look like an outage. A contract change can keep every status code green while changing what the numbers mean. A Check that only asks "did it answer 200" will pass forever on an endpoint that stopped answering correctly.
That is why Knotbind Checks are written as expected behaviour, not as liveness probes: the exact call the integration makes, plus what a passing answer looks like. And when a Check passes for the first time, Knotbind records the shape of the answer as its Baseline. A later answer with a different shape fails the Check even if it is a 200, so an unannounced change still opens a Finding.
Why one failed run is not enough to act on
A single failure is usually weather: a timeout, a hiccup, a cold start. Reacting to one creates noise, and teams that get noise stop reading alerts. So Knotbind marks a connection suspect after one failed run and opens a Finding only when a second consecutive run fails, or when a schema change is breaking on its own terms. Breaking Drift skips the second run: if a tool schema loses or renames something a Check depends on, that does not need to happen twice to matter.
What happens after a Finding opens
Once a Finding is open, the loop does the boring part so a human can do the decision part:
- Diagnose. Knotbind compares the failing answer against the Baseline and the schema history, and names what changed and where your integration touches it.
- Propose. A Repair Proposal is drafted: the change to the request, the Check, or the mapping that would make the call correct again.
- Verify. The proposal runs against the customer's Sandbox, not production. The next passing run in the Sandbox marks it verified.
- Approve. A person decides. Knotbind applies a change only after Owner Approval, and every step — the failure, the diagnosis, the proposal, the decision — is written to an append-only Audit Record.

A fix nobody approved never ships. That is not a disclaimer; it is the design. Agents can ask for a change. Only a person can release one.
The same idea, on the code side
Some breaks are not in a running integration but in the repo behind it. Knotbind Mend watches the GitHub REST API, the OpenAI API, the Stripe API and the core Twilio API for breaking changes, finds the affected lines in connected TypeScript and Python GitHub repos, and opens a draft pull request with the fix, its severity, its confidence and the repo's own test result. When the tests fail or the confidence is low it opens an issue instead. It never writes to a default branch. A person merges what a machine drafted.
Check yours
If you run an API or an MCP server that assistants call, the failure mode above is not hypothetical: it is a matter of when. The Agent Readiness Check is a free, read-only scan of whether an assistant can actually discover your service, read its OpenAPI description, register, and connect to its MCP server without a human. It is a good place to find out whether your integration is built to survive a Wednesday.
Questions people ask about this
Does Knotbind change production by itself? No. It diagnoses a Finding, drafts a Repair Proposal, verifies it in your Sandbox, and applies it only after the Owner approves. Every step is written to an append-only Audit Record.
What is a Baseline? When a Check passes for the first time, Knotbind records the shape of the answer. A later answer with a different shape fails the Check, so a change nobody announced still opens a Finding.
Why two failures and not one? A single failed run is usually a network hiccup and the connection is marked suspect. Two consecutive failures, or a schema change that is breaking on its own, open a Finding.
What can Knotbind watch? Scheduled Checks against HTTP APIs described with OpenAPI and MCP servers over Streamable HTTP. Knotbind Mend watches the GitHub REST API, OpenAI API, Stripe API and core Twilio API for changes that break connected code.
Questions this raises
Does Knotbind change production by itself?
No. Knotbind diagnoses a Finding, drafts a Repair Proposal, verifies it in the customer's Sandbox, and applies it only after Owner Approval. Every step is written to an append-only Audit Record.
What is a Baseline?
When a Check passes for the first time, Knotbind records the shape of the answer as its Baseline. A later answer with a different shape fails the Check, so a change nobody announced still opens a Finding.
Why does a Finding open after two failures and not one?
A single failed run is usually a network hiccup, and the connection is marked suspect. Two consecutive failures, or a schema change that is breaking on its own terms, open a Finding.
What can Knotbind watch?
Scheduled Checks against HTTP APIs described with OpenAPI and MCP servers over Streamable HTTP. Knotbind Mend separately watches the GitHub REST API, OpenAI API, Stripe API and core Twilio API for changes that break connected TypeScript and Python code.
Does Knotbind merge the pull requests Mend opens?
No. Mend opens draft pull requests, or an issue when the repo's own tests fail or confidence is low. Merging is the customer's call, and Mend never writes to a default branch.
