Check inputs and requiredness
List every input field, its type, whether it is required, its allowed range or format, and which earlier step supplies it.
Guide 05 // Agent Preflight
CHECK THE CONTRACT BEFORE THE TOOL.
A tool contract says what a producer supplies, what a consumer expects and where the boundary stops. Review it before a compatibility check so a clean result has a precise meaning.
Why this matters
Names are not contracts.
Two tools can both call a value id while disagreeing about whether it is a customer identifier, a request identifier or an opaque handle. They can also disagree about units, time zones, list ordering, nulls, retry behaviour or who is allowed to use the result.
Declared contracts make those seams reviewable. They let a person spot a missing input or stranded output before an agent is asked to contact a live tool. They also keep compatibility separate from authority: a structurally compatible draft is not permission to access a secret, change a destination or execute work.
Use this checklist for a draft
Keep examples synthetic or redacted. Record the account or workspace scope. Do not paste credentials, bearer tokens, private payloads or production responses into a compatibility review.
Check outputs and consumers
List every output field, its type and requiredness, then name the later step that consumes it or record why it is intentionally terminal.
Check meaning and units
Write down what each value means, its unit, scale, time basis and allowed null or empty state; do not rely on matching field names alone.
Check cardinality and limits
Confirm whether each value is singular or a list, the ordering rule, maximum size, rate, timeout and payload limits, and what happens at the boundary.
Check identity and the auth boundary
Record the account or workspace scope and the identity expected by the receiving tool, while keeping credentials, tokens and payloads out of the review material.
Check failure handling
Define the response shape for validation errors, timeouts, partial results and retries, including whether a failure is safe to repeat or needs human review.
Check authority separately
State what the draft may describe and keep permission to access secrets, change a destination, send a request or execute work outside the compatibility check.
Record an evidence-safe next action
Save the contract version, the checked seam, the redacted result and one bounded correction or review action without claiming availability, delivery or system safety.
Synthetic example
Make the seam concrete.
Suppose a draft connects ticket.lookup to case.attach. The producer returns ticket_id as a string and opened_at as an ISO 8601 UTC timestamp. The consumer expects case_id, a string owned by the receiving workspace, plus opened_at_ms, an integer epoch timestamp in milliseconds.
A redacted compatibility note can say: incompatible at field mapping; ticket_id is not declared as a case_id, and the timestamp unit differs. Proposed correction: add a named lookup adapter that resolves a ticket to a case within the same declared account, then convert the timestamp explicitly. No token, real ticket body or live request is needed to explain the seam.
What the result can establish
It can establish that the declared fields, types, meanings and routing either fit or have a named mismatch at the time of review.
What it cannot establish
It cannot prove endpoint availability, delivery, vendor or framework compatibility, permission, confidentiality, operational reliability or overall system safety.
Turn the check into a safe next step.
Agent Preflight is an account-scoped, read-only review surface. It does not accept or reveal secrets, require sensitive payloads, execute tools, configure deployments or provide public self-service access. It has no promise of availability or SLA, and a named framework is not compatible merely because its fields look similar.
After the review, keep the redacted contract and result with the draft. If the seam fails, use the tool-chain mismatch guide to choose the smallest visible correction. If it fits, use the preflight guide to record what was checked and what still requires a separate human decision.