UNPKG

@bitkyc08/opencodex

Version:

Universal provider proxy for OpenAI Codex & Claude Code — use any LLM with Codex CLI/App/SDK and Claude Code

241 lines (228 loc) • 12.6 kB
/** * The operator opt-in that overrides the stage table's refusal, and the one request-wide * allowance both ambiguous stages claim from. * * `resendPermission` answers `refused-ambiguous` for a native Responses send that died with * the caller having observed nothing -- before the response head, or after it while the SSE * body carried only control events. This module is the one place that answer is overridden, * and it is narrow on three axes at once: the provider has to opt in, the request has to be * one whose second send cannot do more than run the same inference again, and the whole * logical request gets a fixed number of replacements no matter how many legs ask. * * The judgment is made on the inbound body the client sent, which is already parsed. It is * conservative for the outbound request: the proxy expands `previous_response_id` and lowers * hosted tools into client execution, so every hazard that reaches the wire was visible here, * and a hazard visible here may already have been removed. A cheap fail-closed answer beats * re-parsing a multi-megabyte outbound body on every send. * * One removal is certain enough to judge on directly: a `previous_response_id` this proxy * expanded from its own replay state. The native adapter strips that id from every send, so the * wire carries the complete input and no continuation handle (#8453). * * Native passthrough judges BOTH: this inbound body, plus the physical outbound body and adapter * substitution metadata through `createPhysicalReplayGuard`, because a rebuild after the grant * was claimed can reintroduce a hazard the inbound body never had. */ import type { OcxParsedRequest, OcxProviderConfig } from "../../types"; import { resetReplayPolicyFor } from "../../providers/key-failover"; import { authorizeResendForRecovery, type AmbiguousResendAllowance, type ResendDecision } from "../../lib/request-resend-gate"; import type { RequestFailureStage } from "../../lib/request-failure-model"; import type { AdapterRequest } from "../../adapters/base"; import { parseOutboundRequestBody } from "./outbound-request-body"; type PhysicalReplaySnapshot = { body: BodyInit | null | undefined; hosted: boolean; providerExecuted: boolean }; export class UnsafePhysicalReplacement extends Error {} /** Captures the attempt that actually left, and fences any replacement after queued rebuilding. */ export function createPhysicalReplayGuard(providerExecuted: () => boolean) { let last: PhysicalReplaySnapshot | undefined; // A judgment is a pure function of its snapshot, and the outbound body may be megabytes. let lastSafe: boolean | undefined; let pending = false; const snapshot = (request: AdapterRequest, body: BodyInit | null | undefined): PhysicalReplaySnapshot => ({ body, hosted: (request.hostedToolSubstitutions?.length ?? 0) > 0, providerExecuted: providerExecuted(), }); const safe = (value: PhysicalReplaySnapshot | undefined): boolean => value !== undefined && !value.hosted && !value.providerExecuted && typeof value.body === "string" && selfContainedResponsesBody(parseOutboundRequestBody(value.body)); return { isSelfContained: (request: AdapterRequest) => { if (!(lastSafe ??= safe(last))) return false; const current = snapshot(request, request.body); // The body that physically left was judged above; only fresh substitution flags remain. return current.body === last!.body ? !current.hosted && !current.providerExecuted : safe(current); }, authorize: () => { pending = true; }, record: (request: AdapterRequest, body: BodyInit | null | undefined, next?: () => void) => { const physical = snapshot(request, body); // The authorization covers this one send, whether it leaves or fails to: a refused rebuild or // an exhausted send budget must not leave the next ordinary send judged as a replacement. try { if (pending && !safe(physical)) throw new UnsafePhysicalReplacement("The rebuilt request cannot be safely replayed"); next?.(); } finally { pending = false; } last = physical; lastSafe = undefined; }, }; } /** Input items a client owns end to end: replaying them re-runs nothing but the model. */ const CLIENT_INPUT_ITEM_TYPES: ReadonlySet<string> = new Set([ "message", "reasoning", "compaction", "function_call", "function_call_output", "custom_tool_call", "custom_tool_call_output", "tool_search_call", ]); const MESSAGE_ROLES: ReadonlySet<string> = new Set(["user", "assistant", "system", "developer"]); /** Bounded traversal: a catalog is operator data, not a reason to walk forever. */ const MAX_TOOL_ENTRIES = 4096; const MAX_TOOL_DEPTH = 4; function record(value: unknown): value is Record<string, unknown> { return value !== null && typeof value === "object" && !Array.isArray(value); } /** * True when every tool in the catalog is executed by the client. Hosted tools (`web_search`, * `mcp`, `code_interpreter`, ...) run on the origin during the turn, so an unknown or hosted * type fails the whole catalog rather than being skipped: a tool this proxy does not * recognise is a tool it cannot vouch for. */ function clientExecutedTools(tools: unknown, budget: { remaining: number }, depth = 0): boolean { if (!Array.isArray(tools) || depth > MAX_TOOL_DEPTH) return false; return tools.every(tool => { budget.remaining -= 1; if (budget.remaining < 0 || !record(tool)) return false; if (tool.type === "function" || tool.type === "custom") return true; if (tool.type === "tool_search") return tool.execution === "client"; return tool.type === "namespace" && typeof tool.name === "string" && clientExecutedTools(tool.tools, budget, depth + 1); }); } /** * A Responses body whose second send can only repeat the inference: nothing stored, no * server-side continuation state, complete input, and only client-executed tools. Deferred * tool declarations inside `input` are checked by the same rule as the root catalog, so a * hosted tool cannot ride in through `additional_tools` or a `tool_search_output`. */ export function selfContainedResponsesBody(body: unknown): boolean { if (!record(body)) return false; if (body.store !== false || body.background === true) return false; if (body.previous_response_id != null || body.conversation != null || Object.hasOwn(body, "stream_id")) return false; const input = body.input; if (typeof input !== "string" && !Array.isArray(input)) return false; const budget = { remaining: MAX_TOOL_ENTRIES }; if (body.tools !== undefined && !clientExecutedTools(body.tools, budget)) return false; if (typeof input === "string") return true; return input.every(item => { if (!record(item)) return false; if (item.type === "additional_tools" || item.type === "tool_search_output") { return clientExecutedTools(item.tools, budget); } if (item.type === undefined) return typeof item.role === "string" && MESSAGE_ROLES.has(item.role); return typeof item.type === "string" && CLIENT_INPUT_ITEM_TYPES.has(item.type); }); } /** * The native passthrough body to judge: the inbound body, minus a `previous_response_id` the * proxy already expanded into `input`. * * `expandPreviousResponseInput` keeps the id on the expanded body because Kiro and Cursor read * their conversation ids from it, so the inbound body always names it. The native Responses * adapter (and the Azure adapter that wraps it) removes it whenever `_previousResponseInputExpanded` * is set, which is the flag read here. Judging the id that never reaches the wire refused every * chained Codex turn, so an operator's `retryOnReset` opt-in applied only to first turns. * * An id that was NOT expanded is forwarded to the destination and still refuses: that send * continues state stored on the origin. */ export function nativePassthroughReplayJudgmentBody( parsed: Pick<OcxParsedRequest, "_rawBody" | "_previousResponseInputExpanded">, ): unknown { const body = parsed._rawBody; if (parsed._previousResponseInputExpanded !== true || !record(body) || !Object.hasOwn(body, "previous_response_id")) { return body; } const { previous_response_id: _expandedLocally, ...wire } = body; return wire; } /** * Log why an ambiguous reset was not replaced. Operators who opted in otherwise saw only the * terminal 429 and could not tell a missing policy from a request judged not replayable or an * allowance already spent (#8453). Host label, stage and reason only: no body, no credential. */ export function warnAmbiguousResendRefusal(decision: ResendDecision, hostLabel: string): void { if (decision.allowed || !decision.refusal.startsWith("ambiguous-")) return; console.warn(`[upstream-retry] ambiguous reset not replaced (${hostLabel}, ${decision.stage}): ${decision.refusal}`); } /** * One gate answer for every ambiguous row of a lane -- pre-header, post-header SSE, and a Codex * WebSocket that died under its create frame -- asked through `authorizeResendForRecovery` and * with its refusal reason logged. The allowance and host label are read per ask, because the * recovery loop reassigns the provider row and the request URL. */ export function ambiguousResendAuthorizer( allowance: () => AmbiguousResendAllowance | undefined, hostLabel: () => string, ): (stage: RequestFailureStage) => boolean { return stage => { const decision = authorizeResendForRecovery(stage, "connection-reset", allowance()); warnAmbiguousResendRefusal(decision, hostLabel()); return decision.allowed; }; } /** * Chat Completions tools are executed by the client by spec -- the origin only emits the call. * Anything else (`web_search`, `code_interpreter`, a vendor hosted tool) runs on the origin during * the turn, so an unknown or hosted entry fails the whole catalog rather than being skipped. */ function clientExecutedChatTools(tools: unknown, budget: { remaining: number }): boolean { if (!Array.isArray(tools) || tools.length > budget.remaining) return false; budget.remaining -= tools.length; return tools.every(tool => record(tool) && tool.type === "function"); } /** * A Chat Completions body whose second send can only repeat the inference. * * The lane is stateless by construction -- the proxy rebuilds the whole `messages` array on every * turn, and `previous_response_id` is not part of the Chat wire -- so the only hazards left are * server-side storage, which would record a second completion, and a hosted tool the origin would * run a second time. */ export function selfContainedChatBody(body: unknown): boolean { if (!record(body)) return false; if (body.store === true) return false; if (body.previous_response_id != null) return false; // Hosted execution requested outside the `tools` catalog. Judged on the inbound body like every // other hazard here, which is conservative for the outbound request by design: a body that asks // for a hosted search is refused rather than assumed harmless because a later stage might drop // the field. if (body.web_search_options !== undefined) return false; if (!Array.isArray(body.messages)) return false; return body.tools === undefined || clientExecutedChatTools(body.tools, { remaining: MAX_TOOL_ENTRIES }); } /** * The allowance for ONE logical request, or nothing when the provider did not opt in. * * Built once per request and handed to every leg. `claim` spends the request's counter, which * lives on the execution budget and is therefore shared with a combo child's derived scope -- * that sharing is the reason the pre-header helper takes a callback instead of a number. * * The body judgment is carried rather than applied here, because the gate has to be able to * say WHY it refused: "the operator granted nothing" and "this request cannot be replayed" are * different operator problems, and folding them together is what made the old refusal a single * undifferentiated no. * * `selfContained` is a predicate rather than a body, and the getter below is why: a provider * that never opted in must not pay to walk the input array. */ export function ambiguousResendAllowanceFor( provider: Pick<OcxProviderConfig, "retryOnReset">, requestIsSelfContained: () => boolean, claim: (limit: number) => boolean, ): AmbiguousResendAllowance | undefined { const policy = resetReplayPolicyFor(provider); if (policy === null) return undefined; return { get selfContained(): boolean { return requestIsSelfContained(); }, claim: () => claim(policy.replacements), }; }