UNPKG

@bitkyc08/opencodex

Version:

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

141 lines (133 loc) • 7.35 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. */ import type { OcxProviderConfig } from "../../types"; import { resetReplayPolicyFor } from "../../providers/key-failover"; import type { AmbiguousResendAllowance } from "../../lib/request-resend-gate"; /** 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); }); } /** * 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, and the caller memoizes one answer * across every leg of the request. */ 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), }; }