Contract Ops
Schema + fixture + tool call → deterministic reliability report
One-shot validation for agent contracts without execution, replay, signup, or another model call. Compare schema drift, validate representative arguments, inspect fixture chains, sanitize regression packs, and keep the evidence copyable.
For 1–5 related contracts, shared fixtures, and cross-contract reconciliation, build one deterministic portfolio release packet below. The one-contract single-contract release gate and lightweight agent-contract-check remain separate and available.
Portfolio reliability
Reconcile 1–5 related contracts before release
Compare inline JSON Schema, tool-contract, or OpenAPI operation revisions against shared submitted fixtures. The run is deterministic and browser-local; it never resolves $ref, executes tools, reads a repository, or writes back.
Current valid run
Release packet
Cross-contract findings
Fixture results
Unknowns
Assertions
Exact receipt
Advanced: inert API and MCP requests
These are copyable text generated from the same canonical input. Copying does not execute either request.
POST /api/run
MCP run_tool
Exact contract
Bounded inputs, explicit results
agent-contract-check · 2026-07-16-contract-ops-v1Each run compares one saved/current JSON Schema pair, validates one representative argument object, and returns drift, diagnostics, retry safety, and one bounded hint.
Use a safe fixture id up to 64 characters, or let the core derive a deterministic id from the schema pair. Hints are capped at 512 characters.
The browser-local fixture runner accepts JSON-RPC, tool-call, HTTP, and typed-error fixtures in a packet capped at 256,000 bytes and 30 levels of nesting.
Fixture bindings may reference completed steps only. The runner validates structure and expectations; it does not execute tools or replay network requests.
Exports use utilito.agent-fixtures.v1 and utilito.agent-regression-pack.v1, strip dangerous keys, redact recognized secret keys and Bearer/sk--like values, and compare versions by stable fixture id. Redaction is a safety aid, not an exhaustive secret scanner.
AGENT_CONTRACT_CHECK_FAILED means schema drift, argument diagnostics, or retry safety needs review. It is a deterministic result, not an internal server error.
Browser-local
Keep sensitive fixtures on-device
Use the Tool Calling Validator for schema drift and argument checks, or the Agent Regression Pack Workbench for fixture chains, sanitization, versioned packs, and deterministic diffs.
HTTP API
Validate one contract
Call POST /api/run with tool_id: agent-contract-check. Free light use needs no key or account; load the schema first for the current bounded input contract.
MCP
Use the compact router
Load agent-contract-check with get_tool_schema, then call run_tool. Contract Ops does not add another top-level MCP tool.
Contract Ops CA3
Contract Review Packs
Move one deterministic agent-contract-check result into Context Workbench. The contract-review-pack prioritizes breaking drift, diagnostics, and unsafe retry state; the contract-review-report produces escaped Markdown or standalone HTML.
Raw argument values are omitted by design. Argument keys, exact omission counts, drift telemetry, and typed failures remain visible. The pack does not execute or replay the tool.
Boundaries
No recurring monitoring
Contract checks inspect submitted schemas, arguments, failures, and fixtures. They do not call providers, run the described tool, or replay HTTP requests.
Browser-local work stays on-device. API/MCP results are response-only; schemas, arguments, and fixtures are not stored or indefinitely retained by Contract Ops.
CO0 has no scheduled checks, webhook notifications, status history, or monitoring dashboard. Contract Monitor remains evidence-gated.
Contract Ops is a module in the Context Ops Builder Pilot, not a separate subscription. Free one-shot browser and light API/MCP validation remain useful.
Evidence before expansion
Request a named contract workflow
Need a versioned recipe, higher-volume composition, CI integration, or recurring drift check? Name the exact workflow and support boundary. A real request can select the next phase; it does not authorize generic monitoring or fixture storage.