diff --git a/CONFIGURATION.md b/CONFIGURATION.md
index a09f5f8d..946da782 100644
--- a/CONFIGURATION.md
+++ b/CONFIGURATION.md
@@ -421,7 +421,7 @@ Environment variables take precedence over the config file. They are useful for
 | `ACP_LOG` | Set to `0` to disable request logging. |
 | `ACP_AUTO_UPDATE` | Set to `0` to disable auto-update checks. |
 | `ACP_UPDATE_TAG` | Dist-tag channel the auto-updater follows (default `latest`, e.g. `dev`). File-config key: `updateTag`. A `pr-N` preview tag is only followed when explicitly configured. |
-| `BILI_HOST_USAGE_CREDIT` | `#408` host-usage backfill mode (file-config key: `hostUsageCredit`). `auto` (default) = the uncompressed-baseline backfill is armed for plain proxy clients (the bili-launched pi/omp extensions are exempted — their host-side compaction is cancelled, so the baseline drives nothing there). `off` = never backfill — the usage reported to the host is the actually-forwarded (folded) request, matching `[acp-usage] input=`. Use `off` for plain anthropic proxy clients (e.g. ZCode) whose UI would otherwise show the cumulative, drifting baseline as inflated context (#648). |
+| ~~`BILI_HOST_USAGE_CREDIT`~~ / ~~`hostUsageCredit`~~ | **Removed in #660.** Used to select the host-facing usage mode. The #408 uncompressed-baseline backfill is gone entirely — every host now reports the actually-forwarded (post-fold) request as provider-measured (matches `[acp-usage] input=`). Old values left in env or the config file are ignored; remove them. See the "Bug history lesson" section of PR #691. |
 | `ACP_PROVIDERS` | Path to an external `providers.json` (legacy / shared file). |
 | `BILI_REPLAY_RETRY_BASE_MS` | Base backoff delay (ms) for acp-loop replay retries after a transient upstream rejection (default `1500`; set `0` to disable the delay). See #189. |
 | `BILI_REPLAY_RETRY_MAX` | Total attempts for acp-loop replay retries (default `3`; set `1` to disable retries entirely — legacy fail-fast behavior). See #189. |
diff --git a/CONFIGURATION.zh-CN.md b/CONFIGURATION.zh-CN.md
index b31421c3..1cfced12 100644
--- a/CONFIGURATION.zh-CN.md
+++ b/CONFIGURATION.zh-CN.md
@@ -419,6 +419,7 @@
 | `ACP_LOG` | 设为 `0` 关闭请求日志。 |
 | `ACP_AUTO_UPDATE` | 设为 `0` 禁用自动更新检查。 |
 | `ACP_UPDATE_TAG` | 自动更新跟随的 dist-tag 通道（默认 `latest`，如 `dev`）。文件配置键：`updateTag`。`pr-N` 预览 tag 仅在显式配置时才会被跟随。 |
+| ~~`BILI_HOST_USAGE_CREDIT`~~ / ~~`hostUsageCredit`~~ | **#660 已移除。** 曾用于选择宿主可见的用量模式。#408 的未折叠基线回补（backfill）已整体删除 —— 所有宿主现在统一上报“实际转发（后折叠）请求”的 provider 实测用量，与 `[acp-usage] input=` 一致。遗留该环境变量 / 配置键的旧值会被忽略，请删除。教训详见 PR #691 的 “Bug 历史教训” 一节。 |
 | `ACP_PROVIDERS` | 指向外部 `providers.json` 的路径（旧版 / 共享文件）。 |
 | `BILI_REPLAY_RETRY_BASE_MS` | acp-loop 回放重试的基础退避延迟（毫秒）：上游瞬时拒绝后重试（默认 `1500`；设 `0` 关闭延迟）。见 #189。 |
 | `BILI_REPLAY_RETRY_MAX` | acp-loop 回放重试的总次数（默认 `3`；设 `1` 彻底关闭重试 —— 旧版 fail-fast 行为）。见 #189。 |
diff --git a/src/config.ts b/src/config.ts
index e2c6e2a5..670ea1c4 100644
--- a/src/config.ts
+++ b/src/config.ts
@@ -316,15 +316,6 @@ export type ProxyOptions = {
     autoUpdate: boolean;
     /** Dist-tag channel the auto-updater follows (default "latest"). */
     updateTag: string;
-    /** #408 host-usage backfill mode. "auto" (default) = the uncompressed-
-     *  baseline backfill is armed for plain proxy clients (the bili-launched
-     *  pi/omp extensions are exempted — their host compaction is cancelled, so
-     *  the baseline drives nothing on the host side). "off" = never backfill —
-     *  the usage reported to the host is the actually-forwarded (folded)
-     *  request, matching [acp-usage] input= (#648: plain anthropic proxy
-     *  clients like ZCode otherwise show a cumulative, drifting baseline that
-     *  overstates real context pressure). */
-    hostUsageCredit: "auto" | "off";
     logFile?: string;
     /** MITM transparent-proxy mode. When enabled, an HTTP CONNECT handler is
      *  attached so clients that only know how to set HTTP_PROXY (ZCode with a
@@ -462,7 +453,6 @@ export function loadOptions(env: NodeJS.ProcessEnv = process.env): ProxyOptions
         passthrough: passthrough.enabled,
         passthroughSource: passthrough.source,
         autoUpdate: (env.ACP_AUTO_UPDATE ?? (fileConfig.autoUpdate === false ? "0" : "1")) !== "0",
-        hostUsageCredit: parseHostUsageCredit(env.BILI_HOST_USAGE_CREDIT ?? fileConfig.hostUsageCredit),
         updateTag: (env.ACP_UPDATE_TAG ?? fileConfig.updateTag ?? "latest").trim() || "latest",
         logFile: env.ACP_LOG_FILE !== undefined ? (env.ACP_LOG_FILE || undefined) : fileConfig.logFile,
         mitm: {
@@ -497,7 +487,6 @@ type FileConfig = {
     autoUpdate?: boolean;
     /** Dist-tag channel the auto-updater follows (default "latest"). */
     updateTag?: string;
-    hostUsageCredit?: "auto" | "off";
     upstreamProxy?: string;
     upstreamProxyMode?: string;
     logFile?: string;
@@ -609,10 +598,6 @@ export function parseUpstreamProxyMode(value: string | undefined): UpstreamProxy
     return value === "manual" || value === "auto" ? value : "direct";
 }
 
-export function parseHostUsageCredit(value: string | undefined): "auto" | "off" {
-    return value === "off" ? "off" : "auto";
-}
-
 export function parseCompressSettings(v: unknown): (CompressSettings & { injectTool?: boolean; injectNudge?: boolean }) | undefined {
     if (!v || typeof v !== "object" || Array.isArray(v)) return undefined;
     const obj = v as Record<string, unknown>;
diff --git a/src/exit-matrix.ts b/src/exit-matrix.ts
index 42eddbaf..114b7f7e 100644
--- a/src/exit-matrix.ts
+++ b/src/exit-matrix.ts
@@ -1,5 +1,4 @@
 import { emitStreamError, emitPreflightError } from "./stream-error.js";
-import { backfillHostUsage } from "./util.js";
 import { pipePluginChatWithStrip, pipePluginResponsesWithStrip, pipePluginJson } from "./plugin.js";
 import { startServer } from "./server.js";
 
@@ -66,7 +65,7 @@ export const ERROR_DELIVERY: Record<WireExitId, ExitCell> = {
     "plugin-json": {
         implementer: [pipePluginJson],
         contract: "upstream non-2xx/fetch failure → `{error: formatUpstreamError}` JSON; client abort mid-body → clean end, timer cleared",
-        coveredBy: ["tests/issue411-abort-usage.test.ts", "tests/host-usage-backfill.test.ts"],
+        coveredBy: ["tests/issue411-abort-usage.test.ts", "tests/host-usage-postfold.test.ts"],
     },
 };
 
@@ -108,48 +107,48 @@ export const ABORT_PROPAGATION: Record<WireExitId, ExitCell> = {
     },
 };
 
-export const HOST_USAGE_BACKFILL: Record<WireExitId, ExitCell> = {
+export const HOST_USAGE_PASSTHROUGH: Record<WireExitId, ExitCell> = {
     "proxy-openai-sse": {
-        implementer: [backfillHostUsage],
-        contract: "usage frame from the upstream terminal event is credited to the session (input/cached/output), fallback samples on truncation",
-        coveredBy: ["tests/host-usage-backfill.test.ts"],
+        implementer: [startServer],
+        contract: "provider-measured (post-fold) usage credited to the internal ledger and forwarded to the host verbatim — no baseline backfill (#660)",
+        coveredBy: ["tests/host-usage-postfold.test.ts"],
     },
     "proxy-anthropic-sse": {
-        implementer: [backfillHostUsage],
-        contract: "message_start/message_delta usage frames credited",
-        coveredBy: ["tests/host-usage-backfill.test.ts"],
+        implementer: [startServer],
+        contract: "message_start/message_delta usage frames credited, forwarded verbatim",
+        coveredBy: ["tests/host-usage-postfold.test.ts"],
     },
     "proxy-responses-sse": {
-        implementer: [backfillHostUsage],
-        contract: "response.completed usage frame credited (issue #589 frame shape)",
-        coveredBy: ["tests/issue589-usage-frame.test.ts", "tests/host-usage-backfill.test.ts"],
+        implementer: [startServer],
+        contract: "response.completed usage frame credited (issue #589 frame shape), forwarded verbatim",
+        coveredBy: ["tests/issue589-usage-frame.test.ts", "tests/host-usage-postfold.test.ts"],
     },
     "proxy-json": {
-        implementer: [backfillHostUsage],
-        contract: "non-stream usage object credited; no session → skipped (title-gen must not clobber lastInputTokens)",
-        coveredBy: ["tests/host-usage-backfill.test.ts"],
+        implementer: [startServer],
+        contract: "non-stream usage object credited; no session → skipped (title-gen must not clobber lastInputTokens); forwarded verbatim",
+        coveredBy: ["tests/host-usage-postfold.test.ts"],
     },
     "plugin-chat-sse": {
-        implementer: [pipePluginChatWithStrip, backfillHostUsage],
-        contract: "sniffed usage credited unless the call is session-less (#460 title-gen skip)",
-        coveredBy: ["tests/host-usage-backfill.test.ts", "tests/issue411-abort-usage.test.ts"],
+        implementer: [pipePluginChatWithStrip],
+        contract: "sniffed usage credited unless the call is session-less (#460 title-gen skip); stream bytes untouched",
+        coveredBy: ["tests/host-usage-postfold.test.ts", "tests/issue411-abort-usage.test.ts"],
     },
     "plugin-responses-sse": {
-        implementer: [pipePluginResponsesWithStrip, backfillHostUsage],
+        implementer: [pipePluginResponsesWithStrip],
         contract: "sniffed usage credited; verbatim variant skips accounting by design",
-        coveredBy: ["tests/host-usage-backfill.test.ts"],
+        coveredBy: ["tests/host-usage-postfold.test.ts"],
     },
     "plugin-json": {
-        implementer: [pipePluginJson, backfillHostUsage],
-        contract: "JSON usage object credited",
-        coveredBy: ["tests/host-usage-backfill.test.ts"],
+        implementer: [pipePluginJson],
+        contract: "JSON usage object credited; response body forwarded verbatim",
+        coveredBy: ["tests/host-usage-postfold.test.ts"],
     },
 };
 
 export const EXIT_CONCERNS = {
     errorDelivery: ERROR_DELIVERY,
     abortPropagation: ABORT_PROPAGATION,
-    hostUsageBackfill: HOST_USAGE_BACKFILL,
+    hostUsagePassthrough: HOST_USAGE_PASSTHROUGH,
 } as const;
 
 export type ExitConcernId = keyof typeof EXIT_CONCERNS;
diff --git a/src/loop/adapter-anthropic.ts b/src/loop/adapter-anthropic.ts
index dd13f09a..a813d04b 100644
--- a/src/loop/adapter-anthropic.ts
+++ b/src/loop/adapter-anthropic.ts
@@ -118,7 +118,7 @@ function buildTextDeltaEvent(index: number, text: string): Buffer {
     );
 }
 
-export function createAnthropicAdapter(requestBody: Record<string, unknown>, originalSystem?: AnthropicRequestBody["system"], hostCredit = 0): CompressLoopAdapter {
+export function createAnthropicAdapter(requestBody: Record<string, unknown>, originalSystem?: AnthropicRequestBody["system"]): CompressLoopAdapter {
     const model = (requestBody.model as string) ?? undefined;
     let messageId: string | undefined;
     let clientIndex = 0;
@@ -248,26 +248,10 @@ export function createAnthropicAdapter(requestBody: Record<string, unknown>, ori
                     if (typeof u.input_tokens === "number") roundInput = u.input_tokens;
                     if (typeof u.cache_read_input_tokens === "number") roundCached = u.cache_read_input_tokens;
                     if (round === 1) {
-                        // #408: this raw message_start (post-fold input_tokens)
-                        // reaches the host verbatim — add the prepare-time
-                        // credit back so the host anchors on the uncompressed
-                        // baseline.
-                        let chunk = rawBuf;
-                        if (hostCredit > 0 && typeof u.input_tokens === "number") {
-                            const patched = structuredClone(data);
-                            const pmsg = patched["message"] as Record<string, unknown> | undefined;
-                            const pu = (pmsg?.["usage"] ?? {}) as Record<string, unknown>;
-                            if (typeof pu.input_tokens === "number") {
-                                pu.input_tokens += hostCredit;
-                                const out = eventStr
-                                    .split("\n")
-                                    .map((l) => (l.startsWith("data:") ? `data: ${JSON.stringify(patched)}` : l))
-                                    .join("\n");
-                                chunk = Buffer.from(out + "\n\n", "utf8");
-                            }
-                        }
+                        // The raw message_start (with the provider's measured
+                        // usage) reaches the host verbatim — no rewriting.
                         messageStartForwarded = true;
-                        yield { kind: "meta", chunk, firstRoundOnly: true } as ParsedStreamEvent;
+                        yield { kind: "meta", chunk: rawBuf, firstRoundOnly: true } as ParsedStreamEvent;
                     }
                 } else if (type === "ping") {
                     yield { kind: "meta", chunk: rawBuf } as ParsedStreamEvent;
diff --git a/src/loop/adapter-openai.ts b/src/loop/adapter-openai.ts
index 8f5f60da..13b2d788 100644
--- a/src/loop/adapter-openai.ts
+++ b/src/loop/adapter-openai.ts
@@ -116,28 +116,7 @@ function stripFinishReasonChunk(buf: Buffer): Buffer {
     }
 }
 
-function patchUsageChunk(eventStr: string, parsed: Record<string, unknown>, u: Record<string, unknown>, hostCredit: number): Buffer {
-    const pu = typeof u.prompt_tokens === "number" ? u.prompt_tokens : undefined;
-    const tu = typeof u.total_tokens === "number" ? u.total_tokens : undefined;
-    if (hostCredit > 0 && (pu !== undefined || tu !== undefined)) {
-        const patched = {
-            ...parsed,
-            usage: {
-                ...u,
-                ...(pu !== undefined ? { prompt_tokens: pu + hostCredit } : {}),
-                ...(tu !== undefined ? { total_tokens: tu + hostCredit } : {}),
-            },
-        };
-        const out = eventStr
-            .split("\n")
-            .map((l) => (l.startsWith("data:") ? `data: ${JSON.stringify(patched)}` : l))
-            .join("\n");
-        return Buffer.from(out + "\n\n", "utf8");
-    }
-    return Buffer.from(eventStr + "\n\n", "utf8");
-}
-
-export function createOpenaiAdapter(requestBody: Record<string, unknown>, clientSystem?: string, hostCredit = 0, absorbName?: string): CompressLoopAdapter {
+export function createOpenaiAdapter(requestBody: Record<string, unknown>, clientSystem?: string, absorbName?: string): CompressLoopAdapter {
     const model = (requestBody.model as string) ?? "unknown";
     let responseId = `chatcmpl-proxy-${Date.now()}`;
     let toolIndex = 0;
@@ -352,10 +331,10 @@ export function createOpenaiAdapter(requestBody: Record<string, unknown>, client
                         } as ParsedStreamEvent;
                         // #589: include_usage clients (dsh, OpenAI SDK) read usage
                         // from this trailing empty-choices frame; raw tool-call rounds
-                        // must forward it (with the prepare-time credit), not swallow
-                        // it into the internal ledger.
+                        // must forward it verbatim, not swallow it into the internal
+                        // ledger.
                         if (sawRealToolCall) {
-                            yield { kind: "meta", chunk: patchUsageChunk(eventStr, parsed, u, hostCredit) } as ParsedStreamEvent;
+                            yield { kind: "meta", chunk: rawBuf } as ParsedStreamEvent;
                         }
                     }
                     continue;
@@ -376,11 +355,9 @@ export function createOpenaiAdapter(requestBody: Record<string, unknown>, client
                         cachedTokens: typeof pd?.cached_tokens === "number" ? pd.cached_tokens : undefined,
                     } as ParsedStreamEvent;
                     if (sawRealToolCall) {
-                        // #408: this raw finish chunk (with the provider's
-                        // post-fold usage) reaches the host verbatim — add the
-                        // prepare-time credit back so the host anchors on the
-                        // uncompressed baseline.
-                        const chunk = patchUsageChunk(eventStr, parsed, u ?? {}, hostCredit);
+                        // The raw finish chunk (provider-measured usage) reaches
+                        // the host verbatim — no rewriting.
+                        const chunk = rawBuf;
                         // This verbatim chunk IS the round's authoritative completion
                         // (suppressCompletion); write it once and never fall through
                         // to the text/reasoning branches (which would re-emit the same
diff --git a/src/loop/core.ts b/src/loop/core.ts
index fab8c27d..f65f45db 100644
--- a/src/loop/core.ts
+++ b/src/loop/core.ts
@@ -177,9 +177,6 @@ function recordUsage(
     // re-sends the unfolded history, so its usage report over-reports the
     // context the NEXT request will actually carry (see stream.ts applyRanges).
     ctx.session.stats.lastInputTokens = Math.max(0, total - (ctx.session.stats.compressCreditTokens ?? 0));
-    // #408: remember the input-side total reported to the host AFTER the
-    // prepare-time fold backfill (uncompressed baseline), for the /acp panel.
-    ctx.session.hostContextTokens = total + (ctx.session.hostCreditTokens ?? 0);
     if (typeof cached === "number") {
         ctx.session.stats.cachedTokens += cached;
         ctx.session.stats.cacheSamples += 1;
@@ -362,16 +359,6 @@ export async function* runCompressLoop(
             ) {
                 recordUsage(ctx, usage, round);
             }
-            // #408: the provider measured the FOLDED (post-compress) view; add
-            // the prepare-time credit back so the completion event the host
-            // anchors on reports the uncompressed baseline. recordUsage above
-            // already ran on the un-backfilled numbers (internal ledger stays
-            // post-fold).
-            const hostCredit = ctx.session.hostCreditTokens ?? 0;
-            if (hostCredit > 0 && typeof usage.inputTokens === "number") {
-                usage.inputTokens += hostCredit;
-            }
-
             let resolvedText = assistantText;
             let allCalls = calls;
             if (ctx.textProtocol && assistantText.length > 0 && adapter.extractTextTriggers) {
diff --git a/src/loop/index.ts b/src/loop/index.ts
index f41cf473..c8af79e2 100644
--- a/src/loop/index.ts
+++ b/src/loop/index.ts
@@ -25,11 +25,10 @@ export function pickAdapter(
     responsesProjection?: ResponsesProjection,
     anthropicSystem?: AnthropicRequestBody["system"],
     openaiSystem?: string,
-    hostCredit = 0,
     absorbName?: string,
 ): CompressLoopAdapter {
     if (protocol === "responses") return createResponsesAdapter(textProtocol, responsesProjection, absorbName);
-    if (protocol === "openai") return createOpenaiAdapter(requestBody, openaiSystem, hostCredit, absorbName);
-    if (protocol === "anthropic") return createAnthropicAdapter(requestBody, anthropicSystem, hostCredit);
+    if (protocol === "openai") return createOpenaiAdapter(requestBody, openaiSystem, absorbName);
+    if (protocol === "anthropic") return createAnthropicAdapter(requestBody, anthropicSystem);
     throw new Error(`[acp-loop] unknown protocol: ${protocol}`);
 }
diff --git a/src/plugin.ts b/src/plugin.ts
index 3b357ea0..87e2d823 100644
--- a/src/plugin.ts
+++ b/src/plugin.ts
@@ -13,7 +13,7 @@ import { log as loggerLog } from "./logger.js";
 import { degenerateTurnWarning } from "./degenerate-turn.js";
 import { noteWeakOverflow } from "./weak-overflow.js";
 import { warnCacheCollapse } from "./cache-warn.js";
-import { backfillHostUsage, promptInputTotal, type WireProtocol } from "./util.js";
+import { promptInputTotal, type WireProtocol } from "./util.js";
 import { stateDir } from "./paths.js";
 
 // The proxy's own version, read from package.json at runtime (works in both dev
@@ -449,7 +449,7 @@ export function handlePluginStatus(conversationId: string, res: import("node:htt
         const systemPromptTokens = typeof sysTokRaw === "number" && Number.isFinite(sysTokRaw) && sysTokRaw > 0 ? sysTokRaw : 0;
         panel = buildStatusPanel({
             version: `billion-context@${PROXY_VERSION}`,
-            tokenCount: session.hostContextTokens ?? session.stats.lastInputTokens,
+            tokenCount: session.stats.lastInputTokens,
             systemPromptTokens,
             state: session.state,
             nudge,
@@ -469,8 +469,7 @@ export function handlePluginStatus(conversationId: string, res: import("node:htt
         label: session.meta.label ?? null,
         pluginAgent: session.metadata.pluginAgent ?? null,
         contextLimit: typeof limit === "number" ? limit : null,
-        contextTokens: session.hostContextTokens ?? session.stats.lastInputTokens,
-        hostCredit: session.hostCreditTokens ?? 0,
+        contextTokens: session.stats.lastInputTokens,
         inputTokens: session.stats.inputTokens,
         outputTokens: session.stats.outputTokens,
         cachedTokens: session.stats.cachedTokens,
@@ -626,14 +625,12 @@ export function applyUsageSample(session: Session, sample: UsageSample, protocol
         // compress tool results shrink the next request, not this report.
         session.stats.lastInputTokens = Math.max(0, total - (session.stats.compressCreditTokens ?? 0));
         warnCacheCollapse(session, total, sample.cachedTokens ?? 0);
-        // #408: host-facing baseline = this report + prepare-time fold credit.
-        session.hostContextTokens = total + (session.hostCreditTokens ?? 0);
         // #695: per-request parity with the wire path's [acp-usage] — without
         // this, post-fold cache cliffs cannot be attributed from logs.
         const hit = sample.cachedTokens === undefined || total <= 0 ? undefined : Math.round((100 * (sample.cachedTokens ?? 0)) / total);
         const foldNew = session.stats.pendingFoldUsage === true;
         if (foldNew) session.stats.pendingFoldUsage = false;
-        loggerLog("info", `[${session.id}] [plugin] [acp-usage] input=${total} cached=${sample.cachedTokens ?? "n/a"}${hit === undefined ? "" : ` (cache hit ${hit}%)`} ctx=${session.hostContextTokens}${foldNew ? " fold=new" : ""}`);
+        loggerLog("info", `[${session.id}] [plugin] [acp-usage] input=${total} cached=${sample.cachedTokens ?? "n/a"}${hit === undefined ? "" : ` (cache hit ${hit}%)`}${foldNew ? " fold=new" : ""}`);
     }
     if (sample.outputTokens !== undefined) session.stats.outputTokens += sample.outputTokens;
 }
@@ -674,7 +671,6 @@ export async function pipePluginChatWithStrip(
     const decoder = new TextDecoder("utf-8");
     let buf = "";
     const acc: UsageSample = {};
-    const credit = session?.hostCreditTokens ?? 0;
     const onDrop = (snippet: string) => {
         loggerLog("warn", `[tag-echo] stripped model-emitted render tag (plugin passthrough): ${snippet.slice(0, 80).replace(/\n/g, " ")}`);
         log?.(`[tag-echo] stripped model-emitted render tag from plugin passthrough text`);
@@ -898,23 +894,8 @@ export async function pipePluginChatWithStrip(
                     // the uncompressed baseline. Patch `ev` BEFORE the tag-echo
                     // processors run and only rebuild when they return the event
                     // verbatim — otherwise their render-tag stripping is lost.
-                    // message_delta input_tokens is normally absent (or a 0 echo)
-                    // — only patch a real echo.
-                    let backfilled = false;
-                    if (credit > 0 && protocol) {
-                        const usage =
-                            ev["type"] === "message_start"
-                                ? ((ev["message"] as Record<string, unknown> | undefined)?.["usage"] as Record<string, unknown> | undefined)
-                                : (ev["usage"] as Record<string, unknown> | undefined);
-                        const deltaEcho = ev["type"] === "message_delta" && (num(usage?.["input_tokens"]) ?? 0) <= 0;
-                        if (usage && !deltaEcho && backfillHostUsage(protocol, usage, credit)) backfilled = true;
-                    }
                     const out = protocol === "anthropic" ? processAnthropic(ev, rawEvent) : processOpenai(ev, rawEvent);
-                    if (backfilled && out === rawEvent + "\n\n") {
-                        await write(rebuildEvent(rawEvent, ev));
-                    } else if (out.length > 0) {
-                        await write(out);
-                    }
+                    if (out.length > 0) await write(out);
                 }
             }
             if (res.destroyed || res.writableEnded) break;
@@ -1092,17 +1073,6 @@ export async function pipePluginResponsesWithStrip(
                         let evOut = ev;
                         let rebuild = containsRenderTagText(jsonStr);
                         if (rebuild) evOut = stripResponsesText(ev);
-                        if (type === "response.completed") {
-                            // #408: acc already holds the pre-backfill sample
-                            // (usageFromSseEvent ran above) — the internal
-                            // ledger stays post-fold, the host gets the
-                            // uncompressed baseline.
-                            const credit = session?.hostCreditTokens ?? 0;
-                            const usage = (evOut["response"] as Record<string, unknown> | undefined)?.["usage"] as Record<string, unknown> | undefined;
-                            if (credit > 0 && usage && backfillHostUsage("responses", usage, credit)) {
-                                rebuild = true;
-                            }
-                        }
                         const out = rebuild ? rebuildEvent(rawEvent, evOut) : rawEvent + "\n\n";
                         await write(flushTail(out));
                         continue;
@@ -1217,11 +1187,6 @@ export async function pipePluginJson(
                         num(usage["cache_read_input_tokens"]),
                 }, protocol);
                 markDirty(session);
-                // #408: backfill mutates json.usage in place — reserialize below.
-                const credit = session.hostCreditTokens ?? 0;
-                if (credit > 0 && protocol && backfillHostUsage(protocol, usage, credit)) {
-                    mutated = true;
-                }
             }
         }
     } catch { /* non-JSON body — forward verbatim */ }
diff --git a/src/server.ts b/src/server.ts
index 941657c6..f770b697 100644
--- a/src/server.ts
+++ b/src/server.ts
@@ -73,7 +73,7 @@ import { flushPrefixAffinity, hydratePrefixAffinity, scheduleAffinityPersist } f
 import { consumePluginRegisterFor, flushConversations, handlePluginCompact, handlePluginManifest, handlePluginRegister, handlePluginStatus, handlePluginTool, loadConversations, pipePluginChatWithStrip, pipePluginJson, pipePluginResponsesWithStrip, pluginAgentHeader, pluginConversationHeader, pluginReportedContextWindow, recordPluginSession, rememberPluginMessages, takePendingPluginRegister } from "./plugin.js";
 import { setupMitm, readMitmUpstream } from "./mitm.js";
 import type { BiliMessage } from "acp-kernel/wire";
-import { backfillHostUsage, isLoopbackAddress, inspectContextOverflow, reserveOutputHeadroom, shouldReserveOutputHeadroom, systemToUser, usageTotals, type WireProtocol } from "./util.js";
+import { isLoopbackAddress, inspectContextOverflow, reserveOutputHeadroom, shouldReserveOutputHeadroom, systemToUser, usageTotals, type WireProtocol } from "./util.js";
 import { resolveConfirmedLimit, resolveLearnedLimit, resolveSpeculativeLimit, retractStaleLearnedLimits } from "./weak-overflow.js";
 import { BILI_TUNNEL_HEADER, checkTunnelDestination, tunnelAllowlistFromEnv } from "./tunnel-guard.js";
 
@@ -1916,54 +1916,6 @@ function diagNudge(turn: { nudge?: { shouldInject: boolean; reason: string; cont
     return `[${sessionId}] nudge ${inject}: usage=${pct} (${tokenCount}/${limit}), growth=${growth}/${floor} (ref=${ref}, interval=${interval}), pendingT1=${pendingT1}/${interval}${modelTag}, reason="${n.reason.slice(0, 120)}"`;
 }
 
-// #408/#590/#623/#645/#648: hosts that get the #408 uncompressed-baseline usage
-// backfill. pi's AND omp's bili extensions cancel the host's NATIVE compaction
-// so ACP owns compression — pi cancels auto-compaction, omp cancels ALL
-// compaction (its session_before_compact event carries no reason field, so
-// manual /compact can't be preserved). Codex is exempted the same way (#645):
-// the backfilled baseline is a virtual number the model never receives — it
-// drifts turn-to-turn (real post-fold usage + a character-based estimate of
-// the folded-out tokens, so the metric can decrement with no compress), it
-// exceeds the window (user saw 1315/950k = 138%), and it drives nothing in
-// codex: codex's auto-compact keys off total_tokens, which the backfill never
-// touches. Detected by UA (same signal as native-compact interception) because
-// bili-launched codex sessions carry pluginAgent "mcp" (shared with claude).
-// With the host's compaction off (pi/omp) or the backfill inert (codex),
-// reporting the baseline only puts a >100% footer that mismatches the folded
-// request actually forwarded (#590 pi 302.7%, #623 omp 205%, #645 codex 138%).
-// Plain proxy clients keep the #408 behavior by default (their native
-// compaction stays live and consumes the baseline); the `hostUsageCredit`
-// config option (#648) additionally lets such a client opt out entirely
-// ("off") — ZCode and similar plain anthropic clients otherwise show the
-// cumulative, drifting baseline as inflated context in their UI.
-function armHostUsageCredit(
-    session: Session,
-    originalMessages: CoreMessage[],
-    processedMessages: CoreMessage[],
-    headers: http.IncomingHttpHeaders,
-    hostUsageCredit: "auto" | "off",
-    log: (level: string, msg: string) => void,
-): void {
-    session.hostCreditTokens = 0;
-    // #648: "off" disables the #408 uncompressed-baseline backfill — the host
-    // sees the actually-forwarded (folded) request, matching [acp-usage]
-    // input=. Plain proxy clients (ZCode) otherwise show a cumulative,
-    // drifting baseline that overstates real context pressure.
-    if (hostUsageCredit === "off") return;
-    if (session.metadata.pluginAgent === "pi" || session.metadata.pluginAgent === "omp") return;
-    if (isCodexClient(headers)) return;
-    // #408: tokens folded out of the forwarded view vs the host's own (unfolded)
-    // view — added back into the usage reported to the host so its anchor
-    // reflects the uncompressed baseline. Same estimator both sides, so
-    // systematic error cancels in the difference.
-    session.hostCreditTokens = processedMessages.length > 0
-        ? Math.max(0, estimateCoreMessages(originalMessages) - estimateCoreMessages(processedMessages))
-        : 0;
-    if (session.hostCreditTokens > 0) {
-        log("info", `[${session.id}] host usage backfill armed: +${session.hostCreditTokens} tok (forwarded view is folded); host usage will report the uncompressed baseline`);
-    }
-}
-
 // Zero-baseline sessions are judged conservatively ONLY when they arrived
 // anonymously (prefix-affinity forks/reloads, #553): they carry the full raw
 // history but no measurement yet, so feeding 0 blinds the nudge (usage 0%,
@@ -1993,7 +1945,6 @@ function prepareAnthropic(
     const sessionId = session.id;
     const stream = parsed.stream === true;
     ++session.stats.requests;
-    session.hostCreditTokens = 0;
     const injectTools = opts.compress.injectTool && !pluginMode;
     const stripReasoning = (msgs: BiliMessage[]): BiliMessage[] => withReasoningDrop(msgs, reasoning, log, sessionId, isStrictReasoningEcho(session, upstreamOrigin));
 
@@ -2103,7 +2054,6 @@ function prepareAnthropic(
     // identity chain (#268), not part of the Anthropic Messages API — strip it
     // so the real upstream never sees a field it doesn't know.
     delete (rebuilt as Record<string, unknown>).prompt_cache_key;
-    armHostUsageCredit(session, originalMessages, processedMessages, req.headers, opts.hostUsageCredit, log);
     return { body: JSON.stringify(rebuilt), session, processedMessages, originalMessages, anthropicSystem: parsed.system, protocol: "anthropic", stream, compressInjected: injectTools, pluginMode, nudge, prompts, renderTags: "text-only" } as Prepared;
 }
 
@@ -2232,7 +2182,6 @@ function prepareOpenai(
     const sessionId = session.id;
     const stream = parsed.stream === true;
     ++session.stats.requests;
-    session.hostCreditTokens = 0;
     let openaiSystemText = "";
     const stripReasoning = (msgs: BiliMessage[]): BiliMessage[] => withReasoningDrop(msgs, reasoning, log, sessionId, isStrictReasoningEcho(session, upstreamOrigin));
     let openaiOutboundSystem: string | undefined;
@@ -2360,7 +2309,6 @@ function prepareOpenai(
     if (stream && (rebuilt as Record<string, unknown>).stream_options === undefined) {
         (rebuilt as Record<string, unknown>).stream_options = { include_usage: true };
     }
-    armHostUsageCredit(session, originalMessages, processedMessages, req.headers, opts.hostUsageCredit, log);
     // #532: title-gen side requests carry their own tiny system and would
     // clobber the conversation's measured overhead — skip them.
     if (!isTitleGen && openaiOutboundSystem !== undefined) {
@@ -2389,7 +2337,6 @@ function prepareResponses(
     const sessionId = session.id;
     const stream = parsed.stream === true;
     ++session.stats.requests;
-    session.hostCreditTokens = 0;
     const stripReasoning = (msgs: BiliMessage[]): BiliMessage[] => withReasoningDrop(msgs, reasoning, log, sessionId, isStrictReasoningEcho(session, upstreamOrigin));
     if (reconcileNativeCompactionBoundary(session)) {
         log("info", `[${sessionId}] reconciled ACP state after native Responses compact boundary`);
@@ -2617,7 +2564,6 @@ function prepareResponses(
         });
         log("info", `[${sessionId}] responses forward tools=[${fwdTools.join(",")}] injectTool=${injectTools}${pluginMode ? " (plugin mode: wire injection suppressed)" : ""} NO_INJECT_TOOL=${!!process.env.ACP_NO_INJECT_TOOL} NO_COMPRESS_PROMPT=${!!process.env.ACP_NO_COMPRESS_PROMPT}`);
     }
-    armHostUsageCredit(session, originalMessages, processedMessages, req.headers, opts.hostUsageCredit, log);
     // #532: measure the outbound developer(system)+tools overhead for the panel.
     // On this wire the system rides the injected developer message outside the
     // fold space, so counting devContent + tools does not double-count the
@@ -3895,7 +3841,7 @@ async function forward(
                 ? `\n\n---\n\n${buildAbsorbSystemPrompt(absorbToolName(loopConfig))}`
                 : "";
             const systemPrompt = (textProtocol ? buildCompressHybridSystemPrompt(prepared.prompts ?? defaultPrompts) : buildCompressSystemPrompt(prepared.prompts ?? defaultPrompts)) + absorbSection;
-            const adapter = pickAdapter(prepared.protocol, parsedReq, textProtocol, prepared.responsesProjection, prepared.anthropicSystem, prepared.openaiSystemText, prepared.session.hostCreditTokens ?? 0, absorbActive ? absorbToolName(loopConfig) : undefined);
+            const adapter = pickAdapter(prepared.protocol, parsedReq, textProtocol, prepared.responsesProjection, prepared.anthropicSystem, prepared.openaiSystemText, absorbActive ? absorbToolName(loopConfig) : undefined);
             const refreshFolded = (current: CoreMessage[]): CoreMessage[] => {
                 // #422: mirror the prepare's fold with the post-compress state so
                 // the re-request shows the compression the model just performed.
@@ -4004,13 +3950,6 @@ async function forward(
                     const out = u.completion_tokens ?? u.output_tokens;
                     if (typeof out === "number") prepared.session.stats.outputTokens += out;
                 }
-                // #408: the provider measured the folded view — add the
-                // prepare-time credit back so the host anchors on the
-                // uncompressed baseline.
-                const credit = prepared.session.hostCreditTokens ?? 0;
-                if (credit > 0 && backfillHostUsage(prepared.protocol, u, credit)) {
-                    prepared.session.hostContextTokens = (typeof total === "number" ? total : 0) + credit;
-                }
                 if (prepared.protocol === "openai") {
                     rewriteOpenaiJsonResponse(json, ctx);
                 } else if (prepared.protocol === "responses") {
diff --git a/src/session.ts b/src/session.ts
index 5db8c5bc..9892cb5f 100644
--- a/src/session.ts
+++ b/src/session.ts
@@ -146,19 +146,6 @@ export type Session = {
      *  retry callbacks to correlate a transient upstream rejection with the
      *  rewrite that preceded it (#189). A fresh process has none. */
     lastCompress?: LastCompressInfo;
-    /** In-memory only (NOT persisted): tokens folded out of THIS request's
-     *  forwarded view vs the host's own (unfolded) view, computed in prepare*
-     *  as est(originalMessages) − est(processedMessages). Usage recorders add
-     *  this back into the input-side usage field before forwarding to the host,
-     *  so the host's usage anchor carries the uncompressed baseline instead of
-     *  the post-fold value (#408). Overwritten each prepare(); 0 when nothing
-     *  was folded this request. */
-    hostCreditTokens?: number;
-    /** In-memory only (NOT persisted): last input-side usage total reported to
-     *  the host AFTER the hostCreditTokens backfill (uncompressed baseline).
-     *  Feeds the /acp panel tokenCount so it matches what the host footer
-     *  shows (#408). 0/undefined until the first backfilled usage lands. */
-    hostContextTokens?: number;
     /** Promise chain for per-session serialization. Two concurrent requests
      *  sharing a session id would interleave processTurn / stream-rewriter
      *  mutations on session.state, corrupting it. withSessionLock chains each
@@ -314,8 +301,6 @@ export function resetSessionCompression(session: Session): void {
     session.blockContents.clear();
     session.stats.lastInputTokens = 0;
     session.stats.contextTokens = 0;
-    session.hostCreditTokens = 0;
-    session.hostContextTokens = 0;
     session.metadata.nativeCompactionAt = Date.now();
     markDirty(session);
 }
diff --git a/src/util.ts b/src/util.ts
index 40383761..0c138e67 100644
--- a/src/util.ts
+++ b/src/util.ts
@@ -111,34 +111,6 @@ export function promptInputTotal(
     return input + (splitSemantics && typeof cached === "number" ? cached : 0);
 }
 
-/** #408: add the prepare-time fold credit back into a usage object's
- *  input-side fields, in place, so the host's usage anchor reports the
- *  uncompressed baseline instead of the post-fold value the provider measured.
- *  Field names per protocol: Anthropic/Responses `input_tokens` (TOTAL there),
- *  OpenAI `prompt_tokens` (+ `total_tokens` to keep the sum consistent).
- *  Returns true when anything was patched. */
-export function backfillHostUsage(
-    protocol: WireProtocol,
-    usage: Record<string, unknown>,
-    credit: number,
-): boolean {
-    if (!Number.isFinite(credit) || credit <= 0) return false;
-    let patched = false;
-    const add = (key: string): void => {
-        if (typeof usage[key] === "number" && Number.isFinite(usage[key])) {
-            usage[key] = (usage[key] as number) + credit;
-            patched = true;
-        }
-    };
-    if (protocol === "openai") {
-        add("prompt_tokens");
-        add("total_tokens");
-    } else {
-        add("input_tokens");
-    }
-    return patched;
-}
-
 /** Result of inspecting an upstream response for a "context too long" error. */
 export interface ContextOverflowInfo {
     /** True if the response looks like an upstream context-overflow error. */
diff --git a/tests/host-usage-backfill.test.ts b/tests/host-usage-postfold.test.ts
similarity index 78%
rename from tests/host-usage-backfill.test.ts
rename to tests/host-usage-postfold.test.ts
index 073e3968..a28f5a51 100644
--- a/tests/host-usage-backfill.test.ts
+++ b/tests/host-usage-postfold.test.ts
@@ -9,14 +9,14 @@ import type { Config, CoreMessage } from "acp-kernel";
 import { createCore, createInitialState, defaultConfig } from "acp-kernel";
 import { listSessions, _resetSessionsForTest, type Session } from "../src/session.ts";
 import { runCompressLoop, createResponsesAdapter, createOpenaiAdapter, createAnthropicAdapter } from "../src/loop/index.ts";
-import { backfillHostUsage, promptInputTotal, usageTotals } from "../src/util.ts";
+import { promptInputTotal, usageTotals } from "../src/util.ts";
 import { pipePluginChatWithStrip, pipePluginResponsesWithStrip, pipePluginJson, _resetPluginStateForTest } from "../src/plugin.ts";
 import { SessionStore, _setStoreForTest } from "../src/persist.ts";
 import { startServer } from "../src/server.ts";
 import type { ProxyOptions } from "../src/config.ts";
 import { _setForTest as setRegistryForTest } from "../src/registry.ts";
 
-function makeSession(id: string, hostCreditTokens?: number): Session {
+function makeSession(id: string): Session {
     return {
         id,
         meta: {},
@@ -28,11 +28,10 @@ function makeSession(id: string, hostCreditTokens?: number): Session {
         blockContents: new Map(),
         inFlight: 0,
         persisted: false,
-        ...(hostCreditTokens !== undefined ? { hostCreditTokens } : {}),
     };
 }
 
-function makeCtx(id: string, messages: CoreMessage[], hostCreditTokens?: number): {
+function makeCtx(id: string, messages: CoreMessage[]): {
     core: ReturnType<typeof createCore>;
     config: Config;
     messages: CoreMessage[];
@@ -43,7 +42,7 @@ function makeCtx(id: string, messages: CoreMessage[], hostCreditTokens?: number)
         core: createCore(),
         config: defaultConfig(200000),
         messages,
-        session: makeSession(id, hostCreditTokens),
+        session: makeSession(id),
         log: () => {},
     };
 }
@@ -107,69 +106,7 @@ function jsonFilesUnder(dir: string): string[] {
     return out;
 }
 
-test("backfillHostUsage: openai patches prompt_tokens + total_tokens", () => {
-    const u: Record<string, unknown> = { prompt_tokens: 60000, completion_tokens: 5, total_tokens: 60005 };
-    assert.equal(backfillHostUsage("openai", u, 40000), true);
-    assert.equal(u.prompt_tokens, 100000);
-    assert.equal(u.total_tokens, 100005);
-    assert.equal(u.completion_tokens, 5);
-});
-
-test("backfillHostUsage: openai prompt_tokens only (no total_tokens invented)", () => {
-    const u: Record<string, unknown> = { prompt_tokens: 60000 };
-    assert.equal(backfillHostUsage("openai", u, 40000), true);
-    assert.equal(u.prompt_tokens, 100000);
-    assert.equal("total_tokens" in u, false);
-});
-
-test("backfillHostUsage: anthropic + responses patch input_tokens", () => {
-    const a: Record<string, unknown> = { input_tokens: 60000, output_tokens: 5 };
-    assert.equal(backfillHostUsage("anthropic", a, 40000), true);
-    assert.equal(a.input_tokens, 100000);
-    assert.equal(a.output_tokens, 5);
-    const r: Record<string, unknown> = { input_tokens: 60000 };
-    assert.equal(backfillHostUsage("responses", r, 40000), true);
-    assert.equal(r.input_tokens, 100000);
-});
-
-test("backfillHostUsage: credit <= 0 or missing input field is a no-op", () => {
-    const u: Record<string, unknown> = { prompt_tokens: 60000 };
-    assert.equal(backfillHostUsage("openai", u, 0), false);
-    assert.equal(u.prompt_tokens, 60000);
-    const v: Record<string, unknown> = { completion_tokens: 5 };
-    assert.equal(backfillHostUsage("openai", v, 40000), false);
-    assert.equal(v.completion_tokens, 5);
-});
-
-test("#408: responses loop — host completion carries uncompressed baseline, internal ledger stays post-fold", async () => {
-    const ctx = makeCtx("loop-resp", [textMsg("raw_1", "user", "hello")], 40000);
-    const sse = (type: string, data: unknown): string => `event: ${type}\ndata: ${JSON.stringify(data)}\n\n`;
-    const body =
-        sse("response.created", { type: "response.created", response: { id: "resp_1", status: "in_progress" } }) +
-        sse("response.completed", {
-            type: "response.completed",
-            response: { id: "resp_1", status: "completed", output: [], usage: { input_tokens: 60000, output_tokens: 5 } },
-        });
-    const chunks: Buffer[] = [];
-    for await (const chunk of runCompressLoop(
-        streamOf([body]),
-        { ...ctx, protocol: "responses" },
-        { model: "m", input: [] },
-        { url: "https://upstream.test/v1/responses", headers: { authorization: "Bearer t" } },
-        createResponsesAdapter(),
-        "",
-    )) {
-        chunks.push(chunk);
-    }
-    const out = Buffer.concat(chunks).toString("utf8");
-    assert.ok(out.includes('"input_tokens":100000'), `expected backfilled input_tokens in completed event, got: ${out}`);
-    assert.ok(!out.includes('"input_tokens":60000'), "raw folded input_tokens must not reach the host");
-    assert.equal(ctx.session.hostContextTokens, 100000);
-    assert.equal(ctx.session.stats.lastInputTokens, 60000);
-    assert.equal(ctx.session.stats.inputTokens, 60000);
-});
-
-test("#408: responses loop — no credit leaves usage untouched (control)", async () => {
+test("#660: responses loop — provider usage forwarded verbatim (internal ledger credits post-fold)", async () => {
     const ctx = makeCtx("loop-resp-ctrl", [textMsg("raw_1", "user", "hello")]);
     const sse = (type: string, data: unknown): string => `event: ${type}\ndata: ${JSON.stringify(data)}\n\n`;
     const body =
@@ -191,12 +128,11 @@ test("#408: responses loop — no credit leaves usage untouched (control)", asyn
     }
     const out = Buffer.concat(chunks).toString("utf8");
     assert.ok(out.includes('"input_tokens":60000'), out);
-    assert.equal(ctx.session.hostContextTokens, 60000);
     assert.equal(ctx.session.stats.lastInputTokens, 60000);
 });
 
-test("#408: openai adapter — raw finish chunk with usage carries the backfill on real tool calls", async () => {
-    const adapter = createOpenaiAdapter({ model: "m" }, undefined, 40000);
+test("#660: openai adapter — terminal usage chunk reaches the host untouched on real tool calls", async () => {
+    const adapter = createOpenaiAdapter({ model: "m" });
     const chunk = (o: unknown): string => `data: ${JSON.stringify(o)}\n\n`;
     const stream = streamOf([
         chunk({ id: "c1", object: "chat.completion.chunk", created: 1, model: "m", choices: [{ index: 0, delta: { tool_calls: [{ index: 0, id: "call_1", type: "function", function: { name: "get_weather", arguments: "" } }] } }] }),
@@ -207,12 +143,12 @@ test("#408: openai adapter — raw finish chunk with usage carries the backfill
     for await (const ev of adapter.parseStream(stream, 1)) {
         if (ev.kind === "meta") meta += ev.chunk.toString("utf8");
     }
-    assert.ok(meta.includes('"prompt_tokens":90000'), `expected backfilled prompt_tokens in raw finish chunk: ${meta}`);
-    assert.ok(meta.includes('"total_tokens":90010'), meta);
+    assert.ok(meta.includes('"prompt_tokens":50000'), `usage chunk must reach the host unmodified: ${meta}`);
+    assert.ok(!meta.includes('"prompt_tokens":90'), meta);
 });
 
-test("#408: anthropic adapter — first-round message_start meta carries backfilled input_tokens", async () => {
-    const adapter = createAnthropicAdapter({ model: "m" }, undefined, 40000);
+test("#660: anthropic adapter — first-round message_start reaches the host untouched", async () => {
+    const adapter = createAnthropicAdapter({ model: "m" });
     const stream = streamOf([
         `event: message_start\ndata: ${JSON.stringify({ type: "message_start", message: { id: "msg_1", type: "message", role: "assistant", content: [], usage: { input_tokens: 60000, output_tokens: 1 } } })}\n\n`,
         `event: message_stop\ndata: ${JSON.stringify({ type: "message_stop" })}\n\n`,
@@ -221,51 +157,12 @@ test("#408: anthropic adapter — first-round message_start meta carries backfil
     for await (const ev of adapter.parseStream(stream, 1)) {
         if (ev.kind === "meta") meta += ev.chunk.toString("utf8");
     }
-    assert.ok(meta.includes('"input_tokens":100000'), `expected backfilled input_tokens in message_start: ${meta}`);
+    assert.ok(meta.includes('"input_tokens":60000'), `message_start usage must reach the host unmodified: ${meta}`);
 });
 
 before(_resetPluginStateForTest);
 
-test("#408: pipePluginChatWithStrip — openai final usage chunk backfilled, ledger post-fold", async () => {
-    await withTempStore("pipe-openai", async (_dir, store) => {
-        _setStoreForTest(store);
-        const session = makeSession("pipe-oai", 40000);
-        const chunks: Buffer[] = [];
-        const res = makeRes(chunks);
-        const stream = streamOf([
-            `data: ${JSON.stringify({ id: "c1", object: "chat.completion.chunk", choices: [{ index: 0, delta: { content: "Hi" } }] })}\n\n`,
-            `data: ${JSON.stringify({ id: "c1", object: "chat.completion.chunk", choices: [], usage: { prompt_tokens: 60000, completion_tokens: 5, total_tokens: 60005 } })}\n\n`,
-            "data: [DONE]\n\n",
-        ]);
-        await pipePluginChatWithStrip(stream, res, "openai", session);
-        const out = chunks.join("");
-        assert.ok(out.includes('"prompt_tokens":100000'), out);
-        assert.ok(out.includes('"total_tokens":100005'), out);
-        assert.equal(session.stats.lastInputTokens, 60000);
-        assert.equal(session.hostContextTokens, 100000);
-    });
-});
-
-test("#408: pipePluginChatWithStrip — anthropic message_start backfilled, zero message_delta untouched", async () => {
-    await withTempStore("pipe-anthropic", async (_dir, store) => {
-        _setStoreForTest(store);
-        const session = makeSession("pipe-ant", 40000);
-        const chunks: Buffer[] = [];
-        const res = makeRes(chunks);
-        const stream = streamOf([
-            `event: message_start\ndata: ${JSON.stringify({ type: "message_start", message: { id: "msg_1", usage: { input_tokens: 60000, cache_read_input_tokens: 1000 } } })}\n\n`,
-            `event: message_delta\ndata: ${JSON.stringify({ type: "message_delta", usage: { input_tokens: 0, output_tokens: 10 } })}\n\n`,
-        ]);
-        await pipePluginChatWithStrip(stream, res, "anthropic", session);
-        const out = chunks.join("");
-        assert.ok(out.includes('"input_tokens":100000'), out);
-        assert.ok(out.includes('"input_tokens":0'), "zero message_delta input_tokens must stay 0");
-        assert.equal(session.stats.lastInputTokens, 61000);
-        assert.equal(session.hostContextTokens, 101000);
-    });
-});
-
-test("#408: pipePluginChatWithStrip — no credit leaves bytes verbatim (control)", async () => {
+test("#660: pipePluginChatWithStrip forwards usage frames verbatim", async () => {
     await withTempStore("pipe-ctrl", async (_dir, store) => {
         _setStoreForTest(store);
         const session = makeSession("pipe-ctrl");
@@ -279,7 +176,6 @@ test("#408: pipePluginChatWithStrip — no credit leaves bytes verbatim (control
         const out = chunks.join("");
         assert.ok(out.includes('"prompt_tokens":60000'), out);
         assert.ok(!out.includes('"prompt_tokens":100000'), out);
-        assert.equal(session.hostContextTokens, 60000);
     });
 });
 
@@ -326,15 +222,14 @@ test("#408: pipePluginChatWithStrip — split-semantics openai usage keeps lastI
         ]);
         await pipePluginChatWithStrip(stream, res, "openai", session);
         assert.equal(session.stats.lastInputTokens, 26284);
-        assert.equal(session.hostContextTokens, 26284);
         assert.equal(session.stats.cachedTokens, 26278);
     });
 });
 
-test("#408: pipePluginResponsesWithStrip — response.completed usage backfilled", async () => {
+test("#660: pipePluginResponsesWithStrip — response.completed usage forwarded verbatim", async () => {
     await withTempStore("pipe-resp", async (_dir, store) => {
         _setStoreForTest(store);
-        const session = makeSession("pipe-resp", 40000);
+        const session = makeSession("pipe-resp");
         const chunks: Buffer[] = [];
         const res = makeRes(chunks);
         const stream = streamOf([
@@ -342,16 +237,15 @@ test("#408: pipePluginResponsesWithStrip — response.completed usage backfilled
         ]);
         await pipePluginResponsesWithStrip(stream, res, session);
         const out = chunks.join("");
-        assert.ok(out.includes('"input_tokens":100000'), out);
+        assert.ok(out.includes('"input_tokens":60000'), out);
         assert.equal(session.stats.lastInputTokens, 60000);
-        assert.equal(session.hostContextTokens, 100000);
     });
 });
 
-test("#408: pipePluginJson — openai JSON usage backfilled", async () => {
+test("#660: pipePluginJson — openai JSON usage forwarded verbatim", async () => {
     await withTempStore("pipe-json", async (_dir, store) => {
         _setStoreForTest(store);
-        const session = makeSession("pipe-json", 40000);
+        const session = makeSession("pipe-json");
         const chunks: Buffer[] = [];
         const res = makeRes(chunks);
         const body = JSON.stringify({
@@ -363,10 +257,9 @@ test("#408: pipePluginJson — openai JSON usage backfilled", async () => {
         await pipePluginJson(streamOf([body]), res, session, "openai");
         const out = chunks.join("");
         const json = JSON.parse(out) as { usage: Record<string, unknown> };
-        assert.equal(json.usage.prompt_tokens, 100000);
-        assert.equal(json.usage.total_tokens, 100005);
+        assert.equal(json.usage.prompt_tokens, 60000);
+        assert.equal(json.usage.total_tokens, 60005);
         assert.equal(session.stats.lastInputTokens, 60000);
-        assert.equal(session.hostContextTokens, 100000);
     });
 });
 
@@ -409,7 +302,7 @@ test("#408: persist — flat v1 negative lastInputTokens clamps to 0 on load", a
     });
 });
 
-test("#408: prepareOpenai arms the credit — host sees backfilled usage after a compress fold", async () => {
+test("#408/#660: prepareOpenai — post-fold provider usage reaches the host verbatim", async () => {
     _setStoreForTest(new SessionStore({ enabled: false }));
     setRegistryForTest({});
     const upstreamBodies: string[] = [];
@@ -490,8 +383,8 @@ test("#408: prepareOpenai arms the credit — host sees backfilled usage after a
         // turn 5: the model folds m00003..m00004 (u2 + a2, outside the kernel's
         // protected zone of the last 5 messages). The range must NOT include
         // m00001 — the kernel never prunes the first user message, so folding
-        // it would leave the big content in the forwarded view and the credit
-        // (est(original) − est(processed)) would stay ~0.
+        // it would leave the big content in the forwarded view and the fold
+        // assertion below would pass vacuously.
         history.push({ role: "user", content: "t5" });
         const r2 = await post();
         assert.ok(!r2.includes('"name":"compress"'), `compress tool call must be suppressed from the host: ${r2}`);
@@ -503,9 +396,7 @@ test("#408: prepareOpenai arms the credit — host sees backfilled usage after a
         const r3 = await post();
         const m = r3.match(/"prompt_tokens":(\d+)/);
         assert.ok(m, `turn 6 usage chunk missing: ${r3}`);
-        const prompt = Number(m[1]);
-        assert.ok(prompt > 100, `host-facing prompt_tokens must be backfilled above the post-fold 100 (got ${prompt})`);
-        assert.ok(prompt >= 100 + 200, `backfill must carry a meaningful share of the folded ~1400-token range (got ${prompt})`);
+        assert.equal(Number(m[1]), 100, `host-facing prompt_tokens must be the provider-measured post-fold value — no baseline backfill (#660): ${r3}`);
         assert.ok(upstreamBodies.length >= 7, `expected 7 upstream requests (turn5 has a compress round-trip), got ${upstreamBodies.length}`);
         assert.ok(!upstreamBodies[6]!.includes("SENTINEL_FOLD_GONE"), "turn-6 upstream body must not carry the folded u2 content — fold must have happened");
     } finally {
@@ -514,7 +405,7 @@ test("#408: prepareOpenai arms the credit — host sees backfilled usage after a
     }
 });
 
-test("#590: pi plugin mode reports folded usage — host backfill suppressed", async () => {
+test("#590: pi plugin mode reports folded usage verbatim", async () => {
     _setStoreForTest(new SessionStore({ enabled: false }));
     setRegistryForTest({});
     _resetPluginStateForTest();
@@ -561,8 +452,7 @@ test("#590: pi plugin mode reports folded usage — host backfill suppressed", a
     const url = `http://127.0.0.1:${proxyPort}/bili/http://127.0.0.1:${relayPort}/v1/messages`;
     // Sizing copied from plugin-protocol.test.ts: the compressed head
     // (m00001..m00002) exceeds minCompressibleChars while the protected-zone
-    // walk exhausts itself on the tail — so the fold is real and large
-    // enough that an ungated backfill would be plainly visible.
+    // walk exhausts itself on the tail — so the fold is real and non-vacuous.
     const headFiller = "lorem ipsum dolor sit amet consectetur adipiscing elit sed do eiusmod tempor incididunt ut labore et dolore magna aliqua. ".repeat(28);
     const tailFiller = "enim ad minim veniam quis nostrud exercitation ullamco laboris nisi ut aliquip ex ea commodo consequat duis aute irure dolor in reprehenderit in voluptate. ".repeat(28);
     type AnthropicMessage = { role: string; content: string | Array<Record<string, unknown>> };
@@ -626,17 +516,16 @@ test("#590: pi plugin mode reports folded usage — host backfill suppressed", a
         // this would pass vacuously with no credit to suppress.
         assert.equal(upstreamBodies.length, 2);
         assert.ok(!upstreamBodies[1]!.includes("SENTINEL_FOLD_GONE"), "post-fold upstream body must not carry the folded head content");
-        assert.equal(inputTokensOf(r2), 100, "pi plugin mode must report the folded request's own usage — no uncompressed-baseline backfill (#590)");
+        assert.equal(inputTokensOf(r2), 100, "pi plugin mode must report the folded request's own usage verbatim (#590)");
     } finally {
         await new Promise<void>((resolve, reject) => proxy.close((e) => (e ? reject(e) : resolve())));
         await new Promise<void>((resolve, reject) => relay.close((e) => (e ? reject(e) : resolve())));
     }
 });
 
-test("#623: omp plugin mode reports folded usage — host backfill suppressed", async () => {
-    // Mirrors the #590 pi e2e, binding the session as omp. The wire is
-    // incidental — armHostUsageCredit's pluginAgent gate is protocol-agnostic;
-    // reusing the proven pi fixture guarantees a real fold (not a vacuous pass).
+test("#623: omp plugin mode reports folded usage verbatim", async () => {
+    // Mirrors the #590 pi e2e, binding the session as omp; reusing the proven
+    // pi fixture guarantees a real fold (not a vacuous pass).
     _setStoreForTest(new SessionStore({ enabled: false }));
     setRegistryForTest({});
     _resetPluginStateForTest();
@@ -742,20 +631,17 @@ test("#623: omp plugin mode reports folded usage — host backfill suppressed",
         ]);
         assert.equal(upstreamBodies.length, 2);
         assert.ok(!upstreamBodies[1]!.includes("SENTINEL_FOLD_GONE"), "post-fold upstream body must not carry the folded head content");
-        assert.equal(inputTokensOf(r2), 100, "omp plugin mode must report the folded request's own usage — no uncompressed-baseline backfill (#623)");
+        assert.equal(inputTokensOf(r2), 100, "omp plugin mode must report the folded request's own usage verbatim (#623)");
     } finally {
         await new Promise<void>((resolve, reject) => proxy.close((e) => (e ? reject(e) : resolve())));
         await new Promise<void>((resolve, reject) => relay.close((e) => (e ? reject(e) : resolve())));
     }
 });
 
-// #648: ZCode — a plain proxy client on the anthropic wire (no x-bili-plugin
-// header, no special UA) — must be able to opt out of the #408
-// uncompressed-baseline backfill via hostUsageCredit: "off", reporting the
-// folded request's own usage (matching [acp-usage] input=). The control test
-// pins the other side of the gate: an identical plain client on the default
-// (hostUsageCredit: "auto") still gets the #408 backfill. The fold is real
-// (the relay emits a compress tool_use), not a vacuous pass.
+// #648/#660: ZCode — a plain proxy client on the anthropic wire (no
+// x-bili-plugin header, no special UA). Every host sees the folded request's
+// own provider-measured usage (#660). The fold is real (the relay emits a
+// compress tool_use), not a vacuous pass.
 
 const ZCODE_CONV_648 = "zcode-usage-648";
 
@@ -803,7 +689,7 @@ function zcodeConversation(): Array<{ role: string; content: string }> {
     return history;
 }
 
-async function withZCodeHarness(hostUsageCredit: "auto" | "off", fn: (h: { proxy: http.Server; upstream: http.Server; bodies: string[]; url: string }) => Promise<void>): Promise<void> {
+async function withZCodeHarness(fn: (h: { proxy: http.Server; upstream: http.Server; bodies: string[]; url: string }) => Promise<void>): Promise<void> {
     const bodies: string[] = [];
     const upstream = http.createServer((req, res) => {
         const chunks: Buffer[] = [];
@@ -842,7 +728,6 @@ async function withZCodeHarness(hostUsageCredit: "auto" | "off", fn: (h: { proxy
         debug: false,
         passthrough: false,
         autoUpdate: false,
-        hostUsageCredit,
         mitm: { enabled: false, domains: [] },
     } as ProxyOptions);
     await once(proxy, "listening");
@@ -879,8 +764,8 @@ async function setupZCodeCompressedSession(h: { bodies: string[]; url: string })
     return h.bodies.length;
 }
 
-test("#648: ZCode (anthropic wire, hostUsageCredit off) reports folded usage — host backfill suppressed", async () => {
-    await withZCodeHarness("off", async (h) => {
+test("#648/#660: ZCode (anthropic wire, plain proxy client) reports folded usage verbatim", async () => {
+    await withZCodeHarness(async (h) => {
         const afterSetup = await setupZCodeCompressedSession(h);
         const r2 = await fetch(h.url, {
             method: "POST",
@@ -891,33 +776,13 @@ test("#648: ZCode (anthropic wire, hostUsageCredit off) reports folded usage —
         const raw = await r2.text();
         assert.equal(h.bodies.length, afterSetup + 1, "post-fold turn forwarded to upstream exactly once");
         assert.ok(!h.bodies[h.bodies.length - 1]!.includes("SENTINEL_FOLD_GONE"), "post-fold upstream body must not carry the folded head content");
-        assert.equal(zcodeInputTokensOf(raw), 1000, "hostUsageCredit off must report the folded request's own usage — no uncompressed-baseline backfill (#648)");
+        assert.equal(zcodeInputTokensOf(raw), 1000, "host usage must be the folded request's own provider-measured value — no baseline backfill (#648/#660)");
     });
 });
 
-test("#648 control: plain client (anthropic wire, hostUsageCredit auto) still gets the #408 backfill", async () => {
-    await withZCodeHarness("auto", async (h) => {
-        const afterSetup = await setupZCodeCompressedSession(h);
-        const r2 = await fetch(h.url, {
-            method: "POST",
-            headers: { "content-type": "application/json", "x-acp-session": ZCODE_CONV_648 },
-            body: JSON.stringify({ model: "claude-test", max_tokens: 1024, stream: true, system: "You are a test assistant.", messages: zcodeConversation() }),
-        });
-        assert.equal(r2.status, 200);
-        const raw = await r2.text();
-        assert.equal(h.bodies.length, afterSetup + 1, "post-fold turn forwarded to upstream exactly once");
-        assert.ok(!h.bodies[h.bodies.length - 1]!.includes("SENTINEL_FOLD_GONE"), "post-fold upstream body must not carry the folded head content");
-        assert.ok(zcodeInputTokensOf(raw) > 1000, "plain proxy client with hostUsageCredit auto must still see the uncompressed baseline (#408)");
-    });
-});
-
-// #645: codex — a plain proxy client on the responses wire identified by UA —
-// must report the folded request's own usage; the #408 uncompressed-baseline
-// backfill is suppressed (virtual number the model never receives, drifts
-// turn-to-turn, exceeds the window: 1315/950k). The control test pins the
-// other side of the gate: an identical non-codex client still gets the
-// backfill. Harness mirrors codex-compact-e2e.test.ts (real fold, not a
-// vacuous pass).
+// #645/#660: codex — a plain proxy client on the responses wire identified by
+// UA. Every host sees the folded request's own provider-measured usage
+// (#660). Harness mirrors codex-compact-e2e.test.ts (real fold, not a vacuous pass).
 
 const CODEX_UA_645 = "codex_cli_rs/0.1.0 (linux x86_64)";
 const CODEX_CONV_645 = "codex-usage-645";
@@ -1033,7 +898,7 @@ async function setupCodexCompressedSession(h: { bodies: string[]; url: string },
     return h.bodies.length;
 }
 
-test("#645: codex (responses wire, UA) reports folded usage — host backfill suppressed", async () => {
+test("#645/#660: codex UA client (responses wire) reports folded usage verbatim", async () => {
     await withCodexHarness(async (h) => {
         const afterSetup = await setupCodexCompressedSession(h, CODEX_UA_645);
         const r2 = await fetch(h.url, {
@@ -1045,22 +910,6 @@ test("#645: codex (responses wire, UA) reports folded usage — host backfill su
         const raw = await r2.text();
         assert.equal(h.bodies.length, afterSetup + 1, "post-fold turn forwarded to upstream exactly once");
         assert.ok(!h.bodies[h.bodies.length - 1]!.includes("SENTINEL_FOLD_GONE"), "post-fold upstream body must not carry the folded head content");
-        assert.equal(completedUsageOf(raw).input_tokens, 1000, "codex must report the folded request's own usage — no uncompressed-baseline backfill (#645)");
-    });
-});
-
-test("#645 control: non-codex plain client (responses wire) still gets the #408 backfill", async () => {
-    await withCodexHarness(async (h) => {
-        const afterSetup = await setupCodexCompressedSession(h);
-        const r2 = await fetch(h.url, {
-            method: "POST",
-            headers: { "content-type": "application/json" },
-            body: JSON.stringify({ model: "gpt-resp", stream: true, session_id: CODEX_CONV_645, instructions: "You are the test coding agent.", input: codexConversation() }),
-        });
-        assert.equal(r2.status, 200);
-        const raw = await r2.text();
-        assert.equal(h.bodies.length, afterSetup + 1, "post-fold turn forwarded to upstream exactly once");
-        assert.ok(!h.bodies[h.bodies.length - 1]!.includes("SENTINEL_FOLD_GONE"), "post-fold upstream body must not carry the folded head content");
-        assert.ok(completedUsageOf(raw).input_tokens > 1000, "plain proxy client must still see the uncompressed baseline (#408)");
+        assert.equal(completedUsageOf(raw).input_tokens, 1000, "codex (UA) must report the folded request's own provider-measured usage — no baseline backfill (#645/#660)");
     });
 });
diff --git a/tests/issue589-usage-frame.test.ts b/tests/issue589-usage-frame.test.ts
index fac0eaab..75d8fe92 100644
--- a/tests/issue589-usage-frame.test.ts
+++ b/tests/issue589-usage-frame.test.ts
@@ -31,27 +31,20 @@ async function collect(adapter: ReturnType<typeof createOpenaiAdapter>, events:
     return { meta, kinds };
 }
 
-test("#589: raw tool-call round forwards the trailing usage-only frame, patched with credit", async () => {
-    const adapter = createOpenaiAdapter({ model: "m" }, undefined, 40000);
+test("#589: raw tool-call round forwards the trailing usage-only frame verbatim", async () => {
+    const adapter = createOpenaiAdapter({ model: "m" });
     const { meta } = await collect(adapter, [chunk(toolDelta), chunk(finishChunk), chunk(usageFrame), "data: [DONE]\n\n"]);
-    const usageAt = meta.indexOf('"prompt_tokens":90000');
+    const usageAt = meta.indexOf('"prompt_tokens":50000');
     const finishAt = meta.indexOf('"finish_reason":"tool_calls"');
     const doneAt = meta.indexOf("[DONE]");
     assert.ok(usageAt >= 0, `expected patched usage frame in client stream: ${meta}`);
-    assert.ok(meta.includes('"total_tokens":90010'), meta);
+    assert.ok(meta.includes('"total_tokens":50010'), meta);
     assert.ok(finishAt >= 0 && usageAt > finishAt, `usage frame must come after the finish chunk: ${meta}`);
     assert.ok(doneAt >= 0 && usageAt < doneAt, `usage frame must come before [DONE]: ${meta}`);
 });
 
-test("#589: trailing usage-only frame forwarded verbatim when no credit", async () => {
-    const adapter = createOpenaiAdapter({ model: "m" }, undefined, 0);
-    const { meta } = await collect(adapter, [chunk(toolDelta), chunk(finishChunk), chunk(usageFrame), "data: [DONE]\n\n"]);
-    assert.ok(meta.includes('"prompt_tokens":50000'), meta);
-    assert.ok(!meta.includes('"prompt_tokens":90000'), meta);
-});
-
 test("#589: non-tool rounds keep swallowing the trailing frame (rebuild embeds usage)", async () => {
-    const adapter = createOpenaiAdapter({ model: "m" }, undefined, 40000);
+    const adapter = createOpenaiAdapter({ model: "m" });
     const textDelta = { id: "c1", object: "chat.completion.chunk", created: 1, model: "m", choices: [{ index: 0, delta: { content: "hi there" } }] };
     const finishText = { id: "c1", object: "chat.completion.chunk", created: 1, model: "m", choices: [{ index: 0, delta: {}, finish_reason: "stop" }] };
     const { meta } = await collect(adapter, [chunk(textDelta), chunk(finishText), chunk(usageFrame), "data: [DONE]\n\n"]);
@@ -59,7 +52,7 @@ test("#589: non-tool rounds keep swallowing the trailing frame (rebuild embeds u
 });
 
 test("#589: usage-only frame without usage field is not forwarded", async () => {
-    const adapter = createOpenaiAdapter({ model: "m" }, undefined, 40000);
+    const adapter = createOpenaiAdapter({ model: "m" });
     const emptyChoices = { id: "c1", object: "chat.completion.chunk", created: 1, model: "m", choices: [] };
     const { meta } = await collect(adapter, [chunk(toolDelta), chunk(finishChunk), chunk(emptyChoices), "data: [DONE]\n\n"]);
     assert.ok(!meta.includes("chat.completion.chunk\" }"), meta);