UNPKG

@bitkyc08/opencodex

Version:

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

216 lines (205 loc) • 11.4 kB
import { carryProxyError, isProxyError, markProxyError } from "../lib/client-error-body"; import { sharedPoolContinuityDenial } from "../lib/spend-reservation-ledger"; import { markResponseNonReplayable } from "../lib/upstream-retry"; /** * The one place that knows how this proxy refuses a turn on its own workflow budget. * * It is a module rather than two inline blocks because the two call sites -- the HTTP admission * check in `src/server/index.ts` and the pre-dispatch ceiling check in * `src/server/responses/core.ts` -- had drifted into saying different things about the same * refusal, and because the non-obvious part below has to be stated once and not twice. * * The HTTP admission DECISION lives here too, for the same reason: the lane derivation and the * refusal it produces are two halves of one answer, and keeping them in one file is what stops * a denial from losing the detail that makes it legible on the way out. */ import { formatErrorResponse } from "../bridge"; import { addFinalRequestLog, markLocalRequestLogRefusal, type RequestLogContext, } from "./request-log"; import { WORKFLOW_LOCAL_REFUSAL_HEADER, admitWorkflowTurn, workflowSendCeilingReached, workflowDenialSummary, recordWorkflowRefusalEvent, type WorkflowDenial, type WorkflowDecision, type WorkflowLane, type WorkflowSpendDenialDetail, } from "../lib/workflow-budget"; import { currentWorkflowBudgetPolicy } from "../lib/workflow-budget-config"; import { SEND_BUDGET_EXHAUSTED_CODE } from "../lib/errors"; /** * What a caller needs to hand over for the refusal to become a row on `/api/logs`. * * The HTTP admission check refuses before the body is parsed, so its `logCtx` still carries the * `unknown` model and provider the caller seeded it with. That is the honest record -- this * request genuinely never resolved either -- and it is the same placeholder the native * passthrough path already writes. Skipping the row entirely was the worse option: an operator * reading the logs saw no trace at all of a request the proxy had refused. */ export interface WorkflowRefusalLog { readonly requestId: string; readonly start: number; readonly logCtx: RequestLogContext; /** Receiving HTTP route's notification after this refusal's final-log call. */ readonly onLogged?: () => void; } /** * Build the 429 for a refusal this proxy made itself. * * The status and type arguments below do not reach the client: `classifyError` rewrites every * 429 to `rate_limit_error` / `rate_limit_exceeded`, so the body is shaped exactly like a * provider rate limit. That is a deliberate wire contract -- changing it would change how every * client retries -- which leaves two places to carry the truth. The message names the ceiling * that fired and distinguishes admission from refusing a later send, and the header carries * the machine-readable name. The in-process proxy marker establishes its local provenance. * * The row is where an operator actually looks, so it gets the same treatment #4639 established: * `terminalSource: "synthetic"`, a local reason, and an error code naming the ceiling. Pass * `logCtx` when the caller is inside a turn that will write its own row, or `refusalLog` when * the refusal happens before any row exists and this is the only chance to write one. */ export function workflowRefusalResponse( reason: WorkflowDenial, logCtx?: RequestLogContext, refusalLog?: WorkflowRefusalLog, rootId?: string, spend?: WorkflowSpendDenialDetail, ): Response { // The denial detail, when the caller has it, is what turns "a token ceiling refused this" // into "the account ceiling of 100,000 is spent". It reaches the message, the event ring and // the log row through this one function, so the three cannot drift into saying different // things about the same refusal -- which is what put this module here in the first place. const summary = workflowDenialSummary(reason, spend); // Only a caller that decided the refusal ITSELF passes a root id. admitWorkflowTurn already // records its own denials, so passing one there would double-count them. if (rootId) recordWorkflowRefusalEvent(rootId, reason, Date.now(), spend); const recordOn = logCtx ?? refusalLog?.logCtx; const sentPreviously = (recordOn?.activeAttempt?.sendCount ?? 0) > 0 || recordOn?.attempts?.some(attempt => attempt.sendCount > 0) === true; const message = sentPreviously ? summary.message .replace("refused the request locally", "refused an additional send locally") .replace("no provider was contacted", "no additional provider request was sent") .replace("No provider was contacted", "No additional provider request was sent") : summary.message; if (recordOn) { markLocalRequestLogRefusal(recordOn, summary.code); // A locally assigned code wins in addFinalRequestLog, so this is what names the ceiling in // the logs column rather than the generic rate-limit classification a 429 would get. recordOn.errorCode = summary.code; } if (refusalLog) { addFinalRequestLog(refusalLog.requestId, refusalLog.start, refusalLog.logCtx, 429, { closeReason: "terminal", }); refusalLog.onLogged?.(); } const refusal = formatErrorResponse( 429, reason === "workflow-sends-exhausted" ? "workflow_budget_exhausted" : "queue_capacity_exceeded", message, ); refusal.headers.set(WORKFLOW_LOCAL_REFUSAL_HEADER, summary.code); // Without this a browser dashboard cannot read the header at all: the data plane never sets // Access-Control-Expose-Headers, so a cross-origin reader sees only the CORS-safelisted ones. refusal.headers.set("Access-Control-Expose-Headers", WORKFLOW_LOCAL_REFUSAL_HEADER); // No Combo target has an allowance after a root send denial; do not classify it as quota. if (reason === "workflow-sends-exhausted") markResponseNonReplayable(refusal); return markProxyError(refusal, reason === "workflow-spend-exhausted" && spend && Object.hasOwn({ root: 1, identity: 1, pool: 1 }, spend.scope) ? (spend.includesUnboundPoolHistory ? "workflow_unbound" : "workflow") : undefined); } /** Preserve only a locally constructed workflow refusal across response projections. */ export function carryWorkflowRefusal(source: Response, target: Response): Response { const code = isProxyError(source) ? source.headers.get(WORKFLOW_LOCAL_REFUSAL_HEADER) : null; if (!code) return target; target.headers.set(WORKFLOW_LOCAL_REFUSAL_HEADER, code); const exposed = target.headers.get("Access-Control-Expose-Headers"); if (!exposed?.split(",").some(name => name.trim().toLowerCase() === WORKFLOW_LOCAL_REFUSAL_HEADER)) { target.headers.set("Access-Control-Expose-Headers", exposed ? `${exposed}, ${WORKFLOW_LOCAL_REFUSAL_HEADER}` : WORKFLOW_LOCAL_REFUSAL_HEADER); } return carryProxyError(source, target); } /** Distinguish an exhausted live root from the request's independent send allowance. */ export function workflowSendCeilingRefusalResponse( logCtx: RequestLogContext, rootId?: string, recordedOn: Pick<RequestLogContext, "localTerminalReason"> = logCtx, ): Response | undefined { if (!workflowSendCeilingReached(rootId, currentWorkflowBudgetPolicy())) return undefined; // Observer-backed admission already records its denial; legacy budgets need the event here. return workflowRefusalResponse("workflow-sends-exhausted", logCtx, undefined, recordedOn.localTerminalReason === "workflow_sends_exhausted" ? undefined : rootId); } /** Render the approved conservative-overlay explanation after a dispatch reservation refuses. */ export function unboundPoolSpendRefusalMessage( logCtx: Pick<RequestLogContext, "spendRefusalDetail">, ): string | undefined { const detail = logCtx.spendRefusalDetail; if (!detail?.includesUnboundPoolHistory) return undefined; return workflowDenialSummary("workflow-spend-exhausted", detail).message; } /** HTTP form for pre-commit send-budget catches; the tracker already recorded the event. */ export function unboundPoolSpendRefusalResponse(logCtx: RequestLogContext): Response | undefined { const detail = logCtx.spendRefusalDetail; if (!detail?.includesUnboundPoolHistory) return undefined; return workflowRefusalResponse("workflow-spend-exhausted", logCtx, undefined, undefined, detail); } /** Preserve affinity when the proxy, rather than the serving credential, refuses a send. */ export function adapterSendBudgetRefusal(logCtx: RequestLogContext, rootId: string | undefined, message: string): Response { logCtx.terminalSource = "synthetic"; return unboundPoolSpendRefusalResponse(logCtx) ?? workflowSendCeilingRefusalResponse(logCtx, rootId) ?? markProxyError(formatErrorResponse(429, SEND_BUDGET_EXHAUSTED_CODE, message)); } /** * Admit one HTTP turn against its root workflow budget. * * A fan-out shares the conversation it serves. Without a reserve, a worker burst takes every * slot under its own root and the interactive turn that started it waits behind its own * children. A request that names a parent is treated as that fan-out; a top-level request is * the conversation and may use the reserved slots. */ /** Read-only alias validation: keep replay corruption distinct from conflicting pool owners. */ export function poolContinuityRefusalReason(): WorkflowDenial | undefined { const denial = sharedPoolContinuityDenial(); if (!denial) return undefined; return denial.reason === "pool-history-unresolved" ? "workflow-pool-history-unresolved" : "workflow-spend-undurable"; } export function admitHttpWorkflowTurn(headers: Headers): WorkflowDecision | undefined { const rootId = headers.get("x-codex-parent-thread-id")?.trim() || undefined; const continuityRefusal = poolContinuityRefusalReason(); if (continuityRefusal) { recordWorkflowRefusalEvent(rootId, continuityRefusal); return { admitted: false, reason: continuityRefusal, rootId: rootId ?? "" }; } const threadId = headers.get("thread-id")?.trim() || undefined; const lane: WorkflowLane = rootId !== undefined && threadId !== undefined && threadId !== rootId ? "worker" : "interactive"; return admitWorkflowTurn(rootId, lane, currentWorkflowBudgetPolicy(), threadId); } /** * The refusal for a decision {@link admitHttpWorkflowTurn} declined. * * It exists so the denial's own detail survives the trip to the wire. A token-ceiling denial * arrives carrying the scope that refused and the number it refused against, and a caller that * forwarded only `reason` would answer with a 429 that names no ceiling -- which is the * illegible refusal this whole path is built to avoid. No root id is passed on: the admission * check already recorded its own event, and passing one would count the refusal twice. */ export function workflowDecisionRefusalResponse( decision: Extract<WorkflowDecision, { admitted: false }>, logCtx?: RequestLogContext, refusalLog?: WorkflowRefusalLog, ): Response { const spend = decision.spendScope !== undefined && decision.spendLimit !== undefined ? { scope: decision.spendScope, limit: decision.spendLimit, ...(decision.spendProjected !== undefined ? { projected: decision.spendProjected } : {}), ...(decision.spendIncludesUnboundPoolHistory ? { includesUnboundPoolHistory: true } : {}), } : undefined; return workflowRefusalResponse(decision.reason, logCtx, refusalLog, undefined, spend); }