ModelConnector/crates/model-connector/src/contract.rs
Meta Vens 5a4d8bd477 connector-contract-v3: delete the Responses fields and Refusal::Policy; type the request, outcome and refusals
ProviderRequest loses reasoning_summary, service_tier, output_format_name,
previous_response_id and prompt_cache_key; reasoning_effort is a ReasoningEffort,
max_output_tokens is required and cache_retention is added. Completed carries a
StopReason and Usage, Failed a FailureKind, and the receipt names its provider.
Refusal::Policy becomes ModelNotAllowed, OutputBound and PayloadBound, each its
own branch in begin. cultcache-rs moves to the one CultLib pin a4eae90.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
2026-10-28 11:11:53 +01:00

658 lines
20 KiB
Rust

use std::collections::HashSet;
use serde::{Deserialize, Serialize};
use sha2::{Digest, Sha256};
use thiserror::Error;
use zeroize::Zeroize;
use crate::envelope::ServiceError;
pub const PROVIDER_REQUEST_SCHEMA_ID: &str = "gamecult.codex.provider_request.v2";
pub const INVOCATION_SCHEMA_ID: &str = "gamecult.codex.transport_invocation.v2";
pub const RESULT_SCHEMA_ID: &str = "gamecult.codex.transport_result.v2";
pub const RECEIPT_SCHEMA_ID: &str = "gamecult.codex.transport_receipt.v2";
pub const ENVELOPE_SCHEMA_ID: &str = "gamecult.codex.transport_envelope.v2";
#[derive(Clone, PartialEq, Eq)]
pub struct TransportKey(pub(crate) [u8; 32]);
impl TransportKey {
pub fn from_connection_secret(secret: &str) -> Result<Self, ServiceError> {
if secret.trim().is_empty() || secret.trim() != secret {
return Err(ServiceError::InvalidAdmission);
}
Ok(Self(Sha256::digest(secret.as_bytes()).into()))
}
}
impl Drop for TransportKey {
fn drop(&mut self) {
self.0.zeroize();
}
}
#[derive(Debug, Clone, PartialEq, Eq, Serialize, Deserialize)]
pub struct ProviderRequest {
pub schema_id: String,
pub request_id: String,
pub conversation_id: String,
pub model: String,
pub instructions: String,
pub input: Vec<InputItem>,
pub reasoning_effort: Option<ReasoningEffort>,
pub cache_retention: Option<CacheRetention>,
pub tools: Vec<ToolDefinition>,
pub tool_choice: ToolChoice,
pub parallel_tool_calls: bool,
pub output_schema_json: Option<String>,
pub max_output_tokens: u32,
}
impl ProviderRequest {
pub fn new(
request_id: impl Into<String>,
conversation_id: impl Into<String>,
model: impl Into<String>,
instructions: impl Into<String>,
max_output_tokens: u32,
) -> Self {
Self {
schema_id: PROVIDER_REQUEST_SCHEMA_ID.to_string(),
request_id: request_id.into(),
conversation_id: conversation_id.into(),
model: model.into(),
instructions: instructions.into(),
input: Vec::new(),
reasoning_effort: None,
cache_retention: None,
tools: Vec::new(),
tool_choice: ToolChoice::Auto,
parallel_tool_calls: false,
output_schema_json: None,
max_output_tokens,
}
}
pub fn validate(&self) -> Result<(), ContractError> {
if self.schema_id != PROVIDER_REQUEST_SCHEMA_ID {
return Err(ContractError::Schema);
}
require_id(&self.request_id, "request_id")?;
require_id(&self.conversation_id, "conversation_id")?;
require_id(&self.model, "model")?;
require_content(&self.instructions, "instructions")?;
if self.max_output_tokens == 0 {
return Err(ContractError::Invalid("max_output_tokens"));
}
if let Some(schema) = &self.output_schema_json {
require_content(schema, "output_schema_json")?;
require_json_object(schema, "output_schema_json")?;
}
if self.tools.is_empty() && self.tool_choice == ToolChoice::Required {
return Err(ContractError::Invalid("tool_choice"));
}
let mut tool_names = HashSet::new();
for tool in &self.tools {
require_provider_name(&tool.name, "tool.name")?;
require_content(&tool.description, "tool.description")?;
require_content(&tool.parameters_json, "tool.parameters_json")?;
require_json_object(&tool.parameters_json, "tool.parameters_json")?;
if !tool_names.insert(tool.name.as_str()) {
return Err(ContractError::DuplicateTool(tool.name.clone()));
}
}
for item in &self.input {
match item {
InputItem::UserText { text } | InputItem::AssistantText { text } => {
require_content(text, "input.text")?;
}
InputItem::ToolCall {
call_id,
name,
arguments,
} => {
require_provider_call_id(call_id, "input.call_id")?;
require_provider_name(name, "input.tool_name")?;
require_content(arguments, "input.arguments")?;
}
InputItem::ToolResult { call_id, output } => {
require_provider_call_id(call_id, "input.call_id")?;
require_content(output, "input.output")?;
}
}
}
Ok(())
}
}
#[derive(Debug, Clone, Copy, PartialEq, Eq, Serialize, Deserialize)]
#[serde(rename_all = "snake_case")]
pub enum ToolChoice {
Auto,
Required,
}
#[derive(Debug, Clone, Copy, PartialEq, Eq, Serialize, Deserialize)]
#[serde(rename_all = "snake_case")]
pub enum ReasoningEffort {
Low,
Medium,
High,
Max,
}
#[derive(Debug, Clone, Copy, PartialEq, Eq, Serialize, Deserialize)]
#[serde(rename_all = "snake_case")]
pub enum CacheRetention {
Short,
Long,
}
/// Why a provider stopped producing output.
#[derive(Debug, Clone, Copy, PartialEq, Eq, Serialize, Deserialize)]
#[serde(rename_all = "snake_case")]
pub enum StopReason {
EndTurn,
ToolUse,
MaxOutputTokens,
StopSequence,
Refusal,
ContextWindow,
}
/// What kind of failure a provider reported. `ResultSize` is the connector's own
/// stand-in for a result whose sealed frame cannot be carried.
#[derive(Debug, Clone, Copy, PartialEq, Eq, Serialize, Deserialize)]
#[serde(rename_all = "snake_case")]
pub enum FailureKind {
ResultSize,
InvalidRequest,
Authentication,
RateLimited,
Unavailable,
Timeout,
Provider,
}
/// Token accounting for one completed request. Cache reads and writes are
/// counted apart from plain input; `reasoning_tokens` is absent when the
/// provider does not report it.
#[derive(Debug, Clone, PartialEq, Eq, Serialize, Deserialize)]
pub struct Usage {
pub input_tokens: u64,
pub cache_read_tokens: u64,
pub cache_write_tokens: u64,
pub output_tokens: u64,
pub reasoning_tokens: Option<u64>,
}
#[derive(Debug, Clone, PartialEq, Eq, Serialize, Deserialize)]
pub enum InputItem {
UserText {
text: String,
},
AssistantText {
text: String,
},
ToolCall {
call_id: String,
name: String,
arguments: String,
},
ToolResult {
call_id: String,
output: String,
},
}
#[derive(Debug, Clone, PartialEq, Eq, Serialize, Deserialize)]
pub struct ToolDefinition {
pub name: String,
pub description: String,
pub parameters_json: String,
}
#[derive(Debug, Clone, PartialEq, Eq, Serialize, Deserialize)]
pub struct TransportInvocation {
pub schema_id: String,
pub caller_runtime_id: String,
pub expires_at_unix_ms: u64,
pub native_request_sha256: [u8; 32],
pub provider_request_sha256: [u8; 32],
pub request: ProviderRequest,
}
impl TransportInvocation {
pub fn new(
caller_runtime_id: impl Into<String>,
expires_at_unix_ms: u64,
native_request_sha256: [u8; 32],
request: ProviderRequest,
) -> Result<Self, ContractError> {
let provider_request_sha256 = provider_request_sha256(&request)?;
Ok(Self {
schema_id: INVOCATION_SCHEMA_ID.to_string(),
caller_runtime_id: caller_runtime_id.into(),
expires_at_unix_ms,
native_request_sha256,
provider_request_sha256,
request,
})
}
pub fn validate(&self, now_unix_ms: u64, max_expiry_skew_ms: u64) -> Result<(), ContractError> {
if self.schema_id != INVOCATION_SCHEMA_ID {
return Err(ContractError::Schema);
}
require_id(&self.caller_runtime_id, "caller_runtime_id")?;
if self.native_request_sha256 == [0; 32] {
return Err(ContractError::Invalid("native_request_sha256"));
}
self.request.validate()?;
if self.provider_request_sha256 != provider_request_sha256(&self.request)? {
return Err(ContractError::ProviderDigest);
}
if self.expires_at_unix_ms < now_unix_ms
|| self.expires_at_unix_ms > now_unix_ms.saturating_add(max_expiry_skew_ms)
{
return Err(ContractError::Expiry);
}
Ok(())
}
pub fn request_id(&self) -> &str {
&self.request.request_id
}
}
#[derive(Debug, Clone, PartialEq, Eq, Serialize, Deserialize)]
pub struct TransportResult {
pub schema_id: String,
pub request_id: String,
pub caller_runtime_id: String,
pub disposition: TransportDisposition,
}
#[derive(Debug, Clone, PartialEq, Eq, Serialize, Deserialize)]
pub enum TransportDisposition {
Refused(Refusal),
Transported {
events: Vec<TransportEvent>,
receipt: Box<TransportReceipt>,
},
}
impl TransportResult {
pub fn refused(invocation: &TransportInvocation, reason: Refusal) -> Self {
Self {
schema_id: RESULT_SCHEMA_ID.to_string(),
request_id: invocation.request_id().to_string(),
caller_runtime_id: invocation.caller_runtime_id.clone(),
disposition: TransportDisposition::Refused(reason),
}
}
pub fn transported(
invocation: &TransportInvocation,
events: Vec<TransportEvent>,
receipt: TransportReceipt,
) -> Self {
Self {
schema_id: RESULT_SCHEMA_ID.to_string(),
request_id: invocation.request_id().to_string(),
caller_runtime_id: invocation.caller_runtime_id.clone(),
disposition: TransportDisposition::Transported {
events,
receipt: Box::new(receipt),
},
}
}
pub fn validate_against(&self, invocation: &TransportInvocation) -> Result<(), ContractError> {
if self.schema_id != RESULT_SCHEMA_ID {
return Err(ContractError::Schema);
}
if self.request_id != invocation.request_id()
|| self.caller_runtime_id != invocation.caller_runtime_id
{
return Err(ContractError::Identity);
}
if let TransportDisposition::Transported { events, receipt } = &self.disposition {
for (expected, event) in events.iter().enumerate() {
if event.sequence != expected as u64 {
return Err(ContractError::EventSequence);
}
event.validate()?;
}
receipt.validate_against(invocation)?;
}
Ok(())
}
}
#[derive(Debug, Clone, PartialEq, Eq, Serialize, Deserialize)]
pub enum TransportEventPayload {
TextDelta {
text: String,
},
ToolCall {
call_id: String,
name: String,
arguments: String,
},
}
#[derive(Debug, Clone, PartialEq, Eq, Serialize, Deserialize)]
pub struct TransportEvent {
pub sequence: u64,
pub payload: TransportEventPayload,
}
impl TransportEvent {
fn validate(&self) -> Result<(), ContractError> {
match &self.payload {
TransportEventPayload::TextDelta { text } => require_content(text, "event.text"),
TransportEventPayload::ToolCall {
call_id,
name,
arguments,
} => {
require_id(call_id, "event.call_id")?;
require_id(name, "event.tool_name")?;
require_content(arguments, "event.arguments")
}
}
}
}
#[derive(Debug, Clone, PartialEq, Eq, Serialize, Deserialize)]
pub struct TransportReceipt {
pub schema_id: String,
pub request_id: String,
pub caller_runtime_id: String,
pub native_request_sha256: [u8; 32],
pub provider_request_sha256: [u8; 32],
pub model: String,
pub provider: String,
pub outcome: TransportOutcome,
}
impl TransportReceipt {
pub fn validate_against(&self, invocation: &TransportInvocation) -> Result<(), ContractError> {
if self.schema_id != RECEIPT_SCHEMA_ID {
return Err(ContractError::Schema);
}
if self.request_id != invocation.request_id()
|| self.caller_runtime_id != invocation.caller_runtime_id
|| self.model != invocation.request.model
{
return Err(ContractError::Identity);
}
if self.provider_request_sha256 != invocation.provider_request_sha256 {
return Err(ContractError::ProviderDigest);
}
if self.native_request_sha256 != invocation.native_request_sha256 {
return Err(ContractError::NativeDigest);
}
require_id(&self.provider, "receipt.provider")?;
match &self.outcome {
TransportOutcome::Completed {
provider_response_id: Some(value),
..
} => require_id(value, "receipt.provider_response_id")?,
TransportOutcome::Failed { message, .. } => {
require_content(message, "receipt.failure_message")?;
}
TransportOutcome::Completed {
provider_response_id: None,
..
} => {}
}
Ok(())
}
}
#[derive(Debug, Clone, PartialEq, Eq, Serialize, Deserialize)]
pub enum TransportOutcome {
Completed {
provider_response_id: Option<String>,
stop_reason: StopReason,
usage: Usage,
},
Failed {
kind: FailureKind,
message: String,
},
}
#[derive(Debug, Clone, PartialEq, Eq, Serialize, Deserialize)]
pub enum Refusal {
Expired,
IdentitySubstitution,
ProviderDigestSubstitution,
ModelNotAllowed,
OutputBound,
PayloadBound,
Capacity,
InFlight,
Indeterminate,
ReplayConflict,
Malformed,
}
pub fn canonical_provider_request_bytes(
request: &ProviderRequest,
) -> Result<Vec<u8>, ContractError> {
request.validate()?;
rmp_serde::to_vec(request).map_err(|_| ContractError::Encoding)
}
pub fn provider_request_sha256(request: &ProviderRequest) -> Result<[u8; 32], ContractError> {
Ok(Sha256::digest(canonical_provider_request_bytes(request)?).into())
}
pub fn canonical_invocation_bytes(
invocation: &TransportInvocation,
) -> Result<Vec<u8>, ContractError> {
invocation.request.validate()?;
if invocation.provider_request_sha256 != provider_request_sha256(&invocation.request)? {
return Err(ContractError::ProviderDigest);
}
rmp_serde::to_vec(invocation).map_err(|_| ContractError::Encoding)
}
pub(crate) fn require_id(value: &str, field: &'static str) -> Result<(), ContractError> {
if value.trim().is_empty() || value.trim() != value {
return Err(ContractError::Invalid(field));
}
Ok(())
}
fn require_content(value: &str, field: &'static str) -> Result<(), ContractError> {
if value.is_empty() {
return Err(ContractError::Invalid(field));
}
Ok(())
}
fn require_json_object(value: &str, field: &'static str) -> Result<(), ContractError> {
if !matches!(
serde_json::from_str::<serde_json::Value>(value),
Ok(serde_json::Value::Object(_))
) {
return Err(ContractError::Invalid(field));
}
Ok(())
}
fn require_provider_name(value: &str, field: &'static str) -> Result<(), ContractError> {
if value.is_empty()
|| value.len() > 64
|| !value
.bytes()
.all(|byte| byte.is_ascii_alphanumeric() || matches!(byte, b'_' | b'-'))
{
return Err(ContractError::Invalid(field));
}
Ok(())
}
fn require_provider_call_id(value: &str, field: &'static str) -> Result<(), ContractError> {
if value.is_empty() || value.len() > 64 || !value.is_ascii() {
return Err(ContractError::Invalid(field));
}
Ok(())
}
#[derive(Debug, Error, PartialEq, Eq)]
pub enum ContractError {
#[error("unexpected schema")]
Schema,
#[error("invalid {0}")]
Invalid(&'static str),
#[error("duplicate tool {0:?}")]
DuplicateTool(String),
#[error("request identity substitution")]
Identity,
#[error("provider request digest substitution")]
ProviderDigest,
#[error("native request digest substitution")]
NativeDigest,
#[error("invocation expiry is outside the admitted horizon")]
Expiry,
#[error("transport event sequence is not contiguous")]
EventSequence,
#[error("canonical MessagePack encoding failed")]
Encoding,
}
#[cfg(test)]
mod tests {
use super::*;
use crate::fixtures::*;
#[test]
fn provider_digest_is_stable_and_content_addressed() {
let original = request();
let same = original.clone();
assert_eq!(
provider_request_sha256(&original).unwrap(),
provider_request_sha256(&same).unwrap()
);
let mut changed = original.clone();
changed.instructions.push_str(" Changed.");
assert_ne!(
provider_request_sha256(&original).unwrap(),
provider_request_sha256(&changed).unwrap()
);
}
#[test]
fn provider_contract_refuses_transport_side_fixup_pressure() {
let mut request = request();
request.output_schema_json = Some(r#"{"type":"object"}"#.to_string());
assert_eq!(request.validate(), Ok(()));
request.output_schema_json = Some("[]".to_string());
assert_eq!(
request.validate(),
Err(ContractError::Invalid("output_schema_json"))
);
request.output_schema_json = None;
request.input.push(InputItem::ToolResult {
call_id: "x".repeat(65),
output: "evidence".to_string(),
});
assert_eq!(
request.validate(),
Err(ContractError::Invalid("input.call_id"))
);
}
#[test]
fn invocation_refuses_provider_request_substitution() {
let mut invocation = invocation();
invocation.request.model = "substituted".to_string();
assert_eq!(
invocation.validate(1_000, 2_000),
Err(ContractError::ProviderDigest)
);
}
#[test]
fn invocation_refuses_expired_or_blank_native_provenance() {
let mut invocation = invocation();
assert_eq!(
invocation.validate(2_001, 2_000),
Err(ContractError::Expiry)
);
invocation.expires_at_unix_ms = 2_500;
invocation.native_request_sha256 = [0; 32];
assert_eq!(
invocation.validate(2_001, 2_000),
Err(ContractError::Invalid("native_request_sha256"))
);
}
#[test]
fn typed_tool_round_survives_messagepack() {
let invocation = invocation();
let bytes = rmp_serde::to_vec(&invocation).unwrap();
let decoded: TransportInvocation = rmp_serde::from_slice(&bytes).unwrap();
assert_eq!(decoded, invocation);
assert!(matches!(
decoded.request.tools.as_slice(),
[ToolDefinition { name, .. }] if name == "read_source"
));
}
#[test]
fn transported_result_binds_identity_digest_and_event_order() {
let invocation = invocation();
let result = TransportResult::transported(
&invocation,
vec![TransportEvent {
sequence: 0,
payload: TransportEventPayload::ToolCall {
call_id: "call-1".to_string(),
name: "read_source".to_string(),
arguments: "{}".to_string(),
},
}],
receipt(&invocation),
);
assert_eq!(result.validate_against(&invocation), Ok(()));
let mut reordered = result;
let TransportDisposition::Transported { events, .. } = &mut reordered.disposition else {
unreachable!()
};
events[0].sequence = 1;
assert_eq!(
reordered.validate_against(&invocation),
Err(ContractError::EventSequence)
);
}
#[test]
fn receipt_refuses_caller_substitution() {
let invocation = invocation();
let mut substituted_caller = receipt(&invocation);
substituted_caller.caller_runtime_id = "ghostlight-dungeon-yggdrasil".to_string();
assert_eq!(
substituted_caller.validate_against(&invocation),
Err(ContractError::Identity)
);
let mut substituted_native_basis = receipt(&invocation);
substituted_native_basis.native_request_sha256 = [8; 32];
assert_eq!(
substituted_native_basis.validate_against(&invocation),
Err(ContractError::NativeDigest)
);
}
}