billion-context 0.1.74 → 0.1.75-pr.335.311

This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
package/dist/index.js CHANGED
@@ -47856,7 +47856,7 @@ function parseCompressSettings(v2) {
47856
47856
  }
47857
47857
  out[key] = typeof obj[key] === "string" ? obj[key].trim() : obj[key];
47858
47858
  }
47859
- for (const key of ["nudgeGrowthTokens", "preserveRecentMessages", "preserveRecentTokens", "minCompressRange"]) {
47859
+ for (const key of ["nudgeGrowthTokens", "preserveRecentMessages", "preserveRecentTokens", "minCompressRange", "minCompressRangeChars"]) {
47860
47860
  takeNumber(key);
47861
47861
  }
47862
47862
  if ("tiers" in obj) {
@@ -47922,6 +47922,7 @@ function resolveContextLimitValue(raw, nativeLimit) {
47922
47922
  function mergeCompress(global2, provider, model) {
47923
47923
  const pick = (k2) => model?.[k2] ?? provider?.[k2] ?? global2?.[k2];
47924
47924
  const promptLevels = [global2?.prompts, provider?.prompts, model?.prompts].filter(Boolean);
47925
+ const rangeOf = (s3) => s3?.minCompressRangeChars ?? s3?.minCompressRange;
47925
47926
  return {
47926
47927
  modelContextLimit: pick("modelContextLimit"),
47927
47928
  maxContextLimit: pick("maxContextLimit"),
@@ -47929,7 +47930,7 @@ function mergeCompress(global2, provider, model) {
47929
47930
  nudgeGrowthTokens: pick("nudgeGrowthTokens"),
47930
47931
  preserveRecentMessages: pick("preserveRecentMessages"),
47931
47932
  preserveRecentTokens: pick("preserveRecentTokens"),
47932
- minCompressRange: pick("minCompressRange"),
47933
+ minCompressRangeChars: rangeOf(model) ?? rangeOf(provider) ?? rangeOf(global2),
47933
47934
  tiers: pick("tiers"),
47934
47935
  prompts: promptLevels.length > 0 ? Object.assign({}, ...promptLevels) : void 0,
47935
47936
  acknowledgePromptsRisk: pick("acknowledgePromptsRisk")
@@ -47983,7 +47984,7 @@ function applyCompressSettings(base, limit, s3) {
47983
47984
  preserveRecentTokens: s3.preserveRecentTokens ?? base.preserveRecentTokens,
47984
47985
  compress: {
47985
47986
  ...base.compress,
47986
- minCompressRange: s3.minCompressRange ?? base.compress.minCompressRange
47987
+ minCompressRange: s3.minCompressRangeChars ?? s3.minCompressRange ?? base.compress.minCompressRange
47987
47988
  }
47988
47989
  };
47989
47990
  }
@@ -48008,12 +48009,22 @@ import path3 from "path";
48008
48009
  // src/fetch-util.ts
48009
48010
  var MAX_REQUEST_BYTES = 100 * 1024 * 1024;
48010
48011
  var UPSTREAM_TIMEOUT_MS = 10 * 60 * 1e3;
48012
+ var liveUpstreamTimers = /* @__PURE__ */ new Set();
48011
48013
  async function fetchWithTimeout(url, opts, timeoutMs = UPSTREAM_TIMEOUT_MS, externalSignal) {
48012
48014
  const controller = new AbortController();
48013
- let timer2 = setTimeout(() => controller.abort(), timeoutMs);
48015
+ const armTimer = () => {
48016
+ const t = setTimeout(() => {
48017
+ liveUpstreamTimers.delete(t);
48018
+ controller.abort();
48019
+ }, timeoutMs);
48020
+ liveUpstreamTimers.add(t);
48021
+ return t;
48022
+ };
48023
+ let timer2 = armTimer();
48014
48024
  const rearm = () => {
48015
48025
  clearTimeout(timer2);
48016
- timer2 = setTimeout(() => controller.abort(), timeoutMs);
48026
+ liveUpstreamTimers.delete(timer2);
48027
+ timer2 = armTimer();
48017
48028
  };
48018
48029
  let onExternalAbort = null;
48019
48030
  if (externalSignal) {
@@ -48025,6 +48036,7 @@ async function fetchWithTimeout(url, opts, timeoutMs = UPSTREAM_TIMEOUT_MS, exte
48025
48036
  }
48026
48037
  const cleanup = () => {
48027
48038
  clearTimeout(timer2);
48039
+ liveUpstreamTimers.delete(timer2);
48028
48040
  if (onExternalAbort && externalSignal) externalSignal.removeEventListener("abort", onExternalAbort);
48029
48041
  };
48030
48042
  try {
@@ -48154,6 +48166,7 @@ var CACHE_FILE = path3.join(cacheDir(), "models-dev.json");
48154
48166
  var TTL_MS = 24 * 60 * 60 * 1e3;
48155
48167
  var cache2 = null;
48156
48168
  var loading = null;
48169
+ var warnedConflicts = /* @__PURE__ */ new Set();
48157
48170
  var snapshotReg = null;
48158
48171
  var snapshotMs = 0;
48159
48172
  function bundledSnapshotRegistry() {
@@ -48326,15 +48339,22 @@ function registryLookup(reg, model, host) {
48326
48339
  }
48327
48340
  if (provider === void 0) {
48328
48341
  const suffix = `/${model}`;
48329
- let agreed;
48342
+ let max;
48343
+ const distinct = /* @__PURE__ */ new Set();
48344
+ const parts = [];
48330
48345
  for (const key of Object.keys(reg)) {
48331
48346
  if (!key.endsWith(suffix)) continue;
48332
48347
  const ctx = reg[key].limit?.context;
48333
48348
  if (typeof ctx !== "number" || ctx <= 0) continue;
48334
- if (agreed === void 0) agreed = ctx;
48335
- else if (agreed !== ctx) return void 0;
48349
+ if (max === void 0 || ctx > max) max = ctx;
48350
+ distinct.add(ctx);
48351
+ parts.push(`${key}=${ctx}`);
48336
48352
  }
48337
- return agreed;
48353
+ if (max !== void 0 && distinct.size > 1 && !warnedConflicts.has(model)) {
48354
+ warnedConflicts.add(model);
48355
+ log("warn", `[acp-registry] conflicting context windows for "${model}" (${parts.join(", ")}) \u2014 using max ${max}`);
48356
+ }
48357
+ return max;
48338
48358
  }
48339
48359
  return void 0;
48340
48360
  }
@@ -48459,11 +48479,14 @@ ${summary}` }]
48459
48479
  }
48460
48480
  return { items, replaced, dropped };
48461
48481
  }
48462
- function codexCompactGate(session, effectiveLimit, transformOk) {
48463
- if (!transformOk || effectiveLimit <= 0) return false;
48482
+ function codexCompactGatePre(session, effectiveLimit) {
48483
+ if (effectiveLimit <= 0) return false;
48464
48484
  if (session.stats.lastInputTokens >= effectiveLimit * 0.9) return false;
48465
48485
  return session.state.blocks.some((b2) => b2.active);
48466
48486
  }
48487
+ function codexCompactGate(session, effectiveLimit, transformOk) {
48488
+ return transformOk && codexCompactGatePre(session, effectiveLimit);
48489
+ }
48467
48490
  function buildTriggerForgeBody(summaryText, usage, stream2) {
48468
48491
  const compactionItem = {
48469
48492
  type: "compaction",
@@ -49940,7 +49963,13 @@ function buildSession(parsed) {
49940
49963
  metadata: parsed.metadata ?? {},
49941
49964
  state: mergeState(parsed.state),
49942
49965
  createdAt: parsed.createdAt ?? Date.now(),
49943
- lastSeen: Date.now(),
49966
+ // #404: lastSeen reflects the on-disk savedAt (the last real
49967
+ // activity), NOT the restore moment — a restart must not fabricate
49968
+ // activity for every session (broke fallback=latest ties, panel
49969
+ // freshness, and eviction ordering). Consumers that need "has this
49970
+ // session been used since boot" read the restored flag.
49971
+ lastSeen: parsed.savedAt ?? Date.now(),
49972
+ restored: true,
49944
49973
  blockContents,
49945
49974
  lastMessages: Array.isArray(parsed.messages) ? parsed.messages : void 0,
49946
49975
  inFlight: 0,
@@ -49995,7 +50024,8 @@ async function initSessions() {
49995
50024
  await store.migrateLegacyIds();
49996
50025
  const loaded = await store.loadAll();
49997
50026
  if (loaded.size > MAX_SESSIONS) {
49998
- const entries = [...loaded.entries()].sort((a, b2) => (b2[1].createdAt ?? 0) - (a[1].createdAt ?? 0));
50027
+ const freshness = (s3) => Math.max(s3.createdAt ?? 0, s3.lastSeen ?? 0);
50028
+ const entries = [...loaded.entries()].sort((a, b2) => freshness(b2[1]) - freshness(a[1]));
49999
50029
  for (const [id, s3] of entries) {
50000
50030
  if (sessions.size >= MAX_SESSIONS) break;
50001
50031
  sessions.set(id, s3);
@@ -50008,6 +50038,7 @@ function getSession(id, meta) {
50008
50038
  const existing = sessions.get(id);
50009
50039
  if (existing) {
50010
50040
  existing.lastSeen = Date.now();
50041
+ existing.restored = false;
50011
50042
  if (meta?.protocol && !existing.meta.protocol) existing.meta.protocol = meta.protocol;
50012
50043
  if (meta?.upstreamOrigin && !existing.meta.upstreamOrigin) existing.meta.upstreamOrigin = meta.upstreamOrigin;
50013
50044
  if (meta?.label && !existing.meta.label) existing.meta.label = meta.label;
@@ -50017,6 +50048,7 @@ function getSession(id, meta) {
50017
50048
  const reloaded = store.loadSync(id, meta);
50018
50049
  if (reloaded) {
50019
50050
  reloaded.lastSeen = Date.now();
50051
+ reloaded.restored = false;
50020
50052
  reloaded.persisted = true;
50021
50053
  sessions.set(id, reloaded);
50022
50054
  return reloaded;
@@ -50106,6 +50138,58 @@ function reconcileNativeCompactionBoundary(session) {
50106
50138
  markDirty(session);
50107
50139
  return true;
50108
50140
  }
50141
+ function markCompactionBoundary(session) {
50142
+ session.metadata.compactionBoundary = {
50143
+ at: Date.now(),
50144
+ pending: true
50145
+ };
50146
+ markDirty(session);
50147
+ }
50148
+ function readPreCompactionArchive(session) {
50149
+ const raw = session.metadata.preCompactionArchive;
50150
+ if (raw && typeof raw === "object" && !Array.isArray(raw)) {
50151
+ return raw;
50152
+ }
50153
+ return {};
50154
+ }
50155
+ function preCompactionArchiveOf(session) {
50156
+ return readPreCompactionArchive(session);
50157
+ }
50158
+ function applyCompactionArchive(session, activeBefore, liveRawIds, log2) {
50159
+ const boundary = session.metadata.compactionBoundary;
50160
+ if (!boundary || typeof boundary !== "object" || !boundary.pending) {
50161
+ return;
50162
+ }
50163
+ const activeAfter = new Set(session.state.blocks.filter((b2) => b2.active).map((b2) => b2.blockId));
50164
+ const deactivated = [...activeBefore].filter((id) => !activeAfter.has(id));
50165
+ if (deactivated.length > 0) {
50166
+ const archive = readPreCompactionArchive(session);
50167
+ const at2 = Date.now();
50168
+ for (const id of deactivated) {
50169
+ archive[id] = { at: at2, reason: "content replaced by client native compaction summary" };
50170
+ }
50171
+ session.metadata.preCompactionArchive = archive;
50172
+ }
50173
+ const { byRaw, byRef } = session.state.messageRefs;
50174
+ const prunedByRaw = {};
50175
+ for (const [rawId, ref] of Object.entries(byRaw)) {
50176
+ if (liveRawIds.has(rawId)) prunedByRaw[rawId] = ref;
50177
+ }
50178
+ const prunedByRef = {};
50179
+ for (const [ref, rawId] of Object.entries(byRef)) {
50180
+ if (liveRawIds.has(rawId)) prunedByRef[ref] = rawId;
50181
+ }
50182
+ session.state.messageRefs.byRaw = prunedByRaw;
50183
+ session.state.messageRefs.byRef = prunedByRef;
50184
+ session.metadata.compactionBoundary = {
50185
+ ...boundary,
50186
+ pending: false,
50187
+ archivedAt: Date.now(),
50188
+ archivedBlocks: deactivated
50189
+ };
50190
+ markDirty(session);
50191
+ log2("info", `[${session.id}] native compaction boundary: archived ${deactivated.length} pre-compaction block(s)${deactivated.length > 0 ? ` (${deactivated.join(", ")})` : ""}; pruned ref maps to ${prunedByRaw.length} live raw id(s)`);
50192
+ }
50109
50193
  function evictOldest() {
50110
50194
  let oldestId;
50111
50195
  let oldestSeen = Infinity;
@@ -50182,6 +50266,10 @@ function resolveDecompress(args, ctx) {
50182
50266
  const blockId = rawBlockId.trim();
50183
50267
  const block = ctx.core.decompress(blockId, ctx.session.state);
50184
50268
  if (!block) return `[Block ${blockId} not found]`;
50269
+ const archived = preCompactionArchiveOf(ctx.session);
50270
+ if (archived[blockId] !== void 0) {
50271
+ return `[decompress FAILED: block ${blockId} is a pre-compaction archive \u2014 its content was in the history BEFORE the client's native compaction and is no longer reachable (replaced by the client's compaction summary). decompress is unavailable for archived blocks.]`;
50272
+ }
50185
50273
  const full = args.full === true;
50186
50274
  const cached = ctx.session.blockContents.get(blockId);
50187
50275
  let body;
@@ -50971,7 +51059,7 @@ async function saveProviders(){const button=byId("save-providers");busy(button,t
50971
51059
  async function saveCompress(){const button=byId("save-compress");busy(button,true,"\u4FDD\u5B58\u4E2D\u2026");try{const text=byId("compress-json").value.trim();const compress=text?JSON.parse(text):null;await json("/__bili/config",{method:"PUT",headers:{"content-type":"application/json"},body:JSON.stringify({compress})});toast("\u538B\u7F29\u53C2\u6570\u5DF2\u70ED\u66F4\u65B0")}catch(error){toast(String(error),true)}finally{busy(button,false)}}
50972
51060
  async function loadOverrides(){try{const data=await loadConfig();const providers=data.providers||{};const box=byId("upstream-overrides");const entries=Object.entries(providers);if(entries.length===0){box.innerHTML='<p class="status">\u6682\u65E0 Provider\u3002\u5728"\u9AD8\u7EA7\u8BBE\u7F6E"\u9875\u6DFB\u52A0 Provider \u540E\u53EF\u5728\u6B64\u6309 URL \u914D\u7F6E\u4EE3\u7406\u3002</p>';return}box.innerHTML='<table><thead><tr><th>Provider URL</th><th>\u4E0A\u6E38\u4EE3\u7406\uFF08\u7A7A=\u7EE7\u627F\u5168\u5C40\uFF09</th></tr></thead><tbody>'+entries.map(([url,route])=>'<tr><td class="mono">'+escapeHtml(url)+'</td><td><input class="mono override-proxy" data-url="'+escapeHtml(url)+'" value="'+escapeHtml((route&&route.proxy)||"")+'" placeholder="\u7EE7\u627F\u5168\u5C40"></td></tr>').join("")+'</tbody></table>'}catch(error){byId("upstream-overrides").textContent=String(error)}}
50973
51061
  async function saveOverrides(){const button=byId("save-overrides");busy(button,true,"\u4FDD\u5B58\u4E2D\u2026");try{const data=await loadConfig();const providers=data.providers||{};document.querySelectorAll(".override-proxy").forEach((input)=>{const url=input.dataset.url;if(!url)return;if(!providers[url])providers[url]={};const val=input.value.trim();if(val)providers[url].proxy=val;else delete providers[url].proxy});await json("/__bili/config",{method:"PUT",headers:{"content-type":"application/json"},body:JSON.stringify({providers})});await json("/__bili/config/reload",{method:"POST"});toast("\u6309 URL \u8986\u76D6\u5DF2\u70ED\u66F4\u65B0");await loadOverrides()}catch(error){toast(String(error),true)}finally{busy(button,false)}}
50974
- async function loadSessions(){try{const data=await json("/__bili/stats");const rows=(data.sessions||[]).map((item)=>"<tr><td>"+escapeHtml(item.title||"\u2014")+"</td><td>"+escapeHtml(item.protocol||"\u2014")+"</td><td class=mono>"+escapeHtml(item.label||"\u2014")+"</td><td>"+escapeHtml(item.requests)+"</td><td>"+escapeHtml(item.contextTokens)+"</td><td>"+escapeHtml(new Date(item.lastSeen).toLocaleString())+"</td></tr>").join("");byId("sessions-body").innerHTML=rows||'<tr><td colspan="6">\u6682\u65E0\u4F1A\u8BDD</td></tr>'}catch(error){toast(String(error),true);throw error}}
51062
+ async function loadSessions(){try{const data=await json("/__bili/stats");const rows=(data.sessions||[]).map((item)=>"<tr><td>"+escapeHtml(item.title||"\u2014")+"</td><td>"+escapeHtml(item.protocol||"\u2014")+"</td><td class=mono>"+escapeHtml(item.label||"\u2014")+"</td><td>"+escapeHtml(item.requests)+"</td><td>"+escapeHtml(item.contextTokens)+"</td><td>"+escapeHtml(new Date(item.lastSeen).toLocaleString())+(item.restored?" <span class=muted>(\u6062\u590D\uFF0C\u5F00\u673A\u540E\u672A\u6D3B\u52A8)</span>":"")+"</td></tr>").join("");byId("sessions-body").innerHTML=rows||'<tr><td colspan="6">\u6682\u65E0\u4F1A\u8BDD</td></tr>'}catch(error){toast(String(error),true);throw error}}
50975
51063
  async function refreshSessions(){const button=byId("refresh-sessions");busy(button,true,"\u5237\u65B0\u4E2D\u2026");try{await loadSessions();toast("\u4F1A\u8BDD\u5217\u8868\u5DF2\u5237\u65B0")}catch{}finally{busy(button,false)}}
50976
51064
  document.querySelectorAll(".copy-btn").forEach((btn)=>btn.addEventListener("click",async()=>{busy(btn,true,"\u590D\u5236\u4E2D\u2026");try{await navigator.clipboard.writeText(btn.dataset.copy||"");toast("\u5DF2\u590D\u5236")}catch(e){toast(String(e),true)}finally{busy(btn,false)}}));byId("save-upstream").addEventListener("click",saveUpstream);byId("test-upstream").addEventListener("click",testUpstream);byId("save-providers").addEventListener("click",saveProviders);byId("save-compress").addEventListener("click",saveCompress);byId("save-overrides").addEventListener("click",saveOverrides);byId("refresh-sessions").addEventListener("click",refreshSessions);
50977
51065
  Promise.all([loadConfig(),loadUpstream()]).catch((error)=>toast(String(error),true));setInterval(()=>{if(byId("page-sessions").classList.contains("active"))loadSessions().catch(()=>{})},5000);
@@ -50981,7 +51069,7 @@ Promise.all([loadConfig(),loadUpstream()]).catch((error)=>toast(String(error),tr
50981
51069
  var WEB_STYLES = `
50982
51070
  :root{--bg:#0f1115;--panel:#171a21;--panel2:#1e222c;--line:#2b313d;--text:#e7eaf0;--muted:#8f98a8;--blue:#6ea8fe;--green:#55c98b;--yellow:#e9b949;--red:#ef6b73;--radius:10px}
50983
51071
  *{box-sizing:border-box}body{margin:0;background:var(--bg);color:var(--text);font:14px/1.5 -apple-system,BlinkMacSystemFont,"Segoe UI",sans-serif;min-height:100vh}.app{display:grid;grid-template-columns:220px 1fr;min-height:100vh}.sidebar{border-right:1px solid var(--line);padding:22px 14px;background:#12151b;position:sticky;top:0;height:100vh}.brand{font-size:18px;font-weight:700;padding:0 10px 22px}.brand span{color:var(--blue)}.version{display:block;color:var(--muted);font-size:11px;font-weight:400;margin-top:2px}.nav{display:grid;gap:4px}.nav button{appearance:none;border:0;background:transparent;color:var(--muted);text-align:left;padding:10px 12px;border-radius:7px;cursor:pointer;font:inherit}.nav button:hover,.nav button.active{background:var(--panel2);color:var(--text)}.main{padding:32px;max-width:1050px;width:100%}.page{display:none}.page.active{display:block}h1{font-size:23px;margin:0 0 4px}h2{font-size:15px;margin:0}.lead{color:var(--muted);margin:0 0 24px}.grid{display:grid;grid-template-columns:repeat(2,minmax(0,1fr));gap:14px}.card{background:var(--panel);border:1px solid var(--line);border-radius:var(--radius);padding:18px;margin-bottom:14px}.card-head{display:flex;justify-content:space-between;align-items:center;margin-bottom:16px}.badge{border-radius:99px;padding:3px 9px;font-size:12px;background:var(--panel2);color:var(--muted)}.badge.ok{color:var(--green)}.badge.warn{color:var(--yellow)}.badge.err{color:var(--red)}.kv{display:grid;grid-template-columns:135px 1fr;gap:10px 14px;margin:0}.kv dt{color:var(--muted)}.kv dd{margin:0;word-break:break-all}.mono{font-family:"Cascadia Code",Consolas,monospace;font-size:12px}.actions{display:flex;gap:9px;flex-wrap:wrap;margin-top:18px}.btn{border:1px solid var(--line);background:var(--panel2);color:var(--text);border-radius:7px;padding:8px 13px;cursor:pointer;font:inherit}.btn:hover{border-color:#4b5668}.btn.primary{background:var(--blue);border-color:var(--blue);color:#0b1220;font-weight:600}.btn.danger{color:var(--red)}.btn:disabled{opacity:.55;cursor:wait}.field{display:grid;gap:6px;margin:12px 0}.field label{color:var(--muted);font-size:12px}input,textarea,select{width:100%;background:#101319;color:var(--text);border:1px solid var(--line);border-radius:7px;padding:9px 10px;font:inherit}textarea{min-height:250px;font-family:"Cascadia Code",Consolas,monospace;font-size:12px;resize:vertical}.modes{display:flex;gap:18px;margin:12px 0}.modes label{display:flex;gap:7px;align-items:center;color:var(--muted)}.modes input{width:auto}.notice{border-left:3px solid var(--yellow);padding:10px 12px;background:rgba(233,185,73,.08);color:#d8c185;margin-top:14px}.status{color:var(--muted);white-space:pre-wrap}.status.ok{color:var(--green)}.status.err{color:var(--red)}table{width:100%;border-collapse:collapse}th,td{text-align:left;border-bottom:1px solid var(--line);padding:9px 8px;font-size:12px}th{color:var(--muted);font-weight:500}.toast{position:fixed;right:24px;bottom:24px;background:var(--green);color:#07140d;padding:10px 15px;border-radius:8px;opacity:0;transform:translateY(8px);transition:.2s;pointer-events:none}.toast.show{opacity:1;transform:none}.toast.err{background:var(--red);color:#1b0608}
50984
- .section-title{font-size:17px;margin:28px 0 4px}.hint{color:var(--muted);margin:0 0 14px;font-size:13px}.btn.small{padding:4px 10px;font-size:12px}.config-block{background:var(--panel2);border:1px solid var(--line);border-radius:6px;padding:10px 14px;margin:8px 0 12px;overflow-x:auto;font-size:13px}.config-block code{font-family:"Cascadia Code",Consolas,monospace;color:var(--text)}.fork-ribbon{position:fixed;top:0;right:0;background:#1a1a2e;color:#fff;padding:6px 16px;font-size:12px;text-decoration:none;border-radius:0 0 0 8px;z-index:100;opacity:.8;transition:.2s}.fork-ribbon:hover{opacity:1}
51072
+ .section-title{font-size:17px;margin:28px 0 4px}.hint{color:var(--muted);margin:0 0 14px;font-size:13px}.muted{color:var(--muted);font-size:11px}.btn.small{padding:4px 10px;font-size:12px}.config-block{background:var(--panel2);border:1px solid var(--line);border-radius:6px;padding:10px 14px;margin:8px 0 12px;overflow-x:auto;font-size:13px}.config-block code{font-family:"Cascadia Code",Consolas,monospace;color:var(--text)}.fork-ribbon{position:fixed;top:0;right:0;background:#1a1a2e;color:#fff;padding:6px 16px;font-size:12px;text-decoration:none;border-radius:0 0 0 8px;z-index:100;opacity:.8;transition:.2s}.fork-ribbon:hover{opacity:1}
50985
51073
  @media(max-width:760px){.app{grid-template-columns:1fr}.sidebar{height:auto;position:static;border-right:0;border-bottom:1px solid var(--line)}.nav{display:flex;overflow:auto}.main{padding:20px}.grid{grid-template-columns:1fr}.kv{grid-template-columns:110px 1fr}}
50986
51074
  `;
50987
51075
 
@@ -51010,7 +51098,7 @@ function renderPage(origin, version2) {
51010
51098
  <div class="card"><div class="card-head"><h2>ZCode\uFF08\u7F16\u7A0B\u5957\u9910\uFF09</h2><span class="badge">MITM</span><button class="btn small copy-btn" data-copy="${caPathEsc}">\u590D\u5236 CA \u8DEF\u5F84</button></div><dl class="kv"><dt>\u65B9\u5F0F</dt><dd>MITM \u900F\u660E\u4EE3\u7406</dd><dt>\u8BBE\u7F6E</dt><dd>Settings \u2192 Network \u2192 HTTP Proxy = <span class="mono">${origin}</span></dd><dt>CA \u8BC1\u4E66</dt><dd class="mono">${caPathEsc}${caReady ? "" : "\uFF08\u5C1A\u672A\u751F\u6210\uFF0C\u542F\u52A8\u4EE3\u7406\u540E\u81EA\u52A8\u521B\u5EFA\uFF09"}</dd></dl><p class="hint">CA \u8DEF\u5F84\u662F\u672C\u673A bili \u5B9E\u9645\u751F\u6210\u7684\u7EDD\u5BF9\u8DEF\u5F84\uFF0C\u70B9"\u590D\u5236 CA \u8DEF\u5F84"\u540E\u76F4\u63A5\u7C98\u8D34\u8FDB ZCode\u3002\u26A0\uFE0F Windows \u4E0A ZCode \u4E0D\u4F1A\u5C55\u5F00 <span class="mono">~</span>\uFF0C\u8BF7\u52FF\u624B\u586B <span class="mono">~/\u2026</span> \u5F62\u5F0F\u7684\u8DEF\u5F84\u3002</p></div></section>
51011
51099
  <section id="page-upstream" class="page"><h1>\u4E0A\u6E38\u7F51\u7EDC</h1><p class="lead">\u63A7\u5236 bili \u5982\u4F55\u8BBF\u95EE\u771F\u5B9E Provider\uFF0C\u4E0D\u4F1A\u6539\u53D8\u5BA2\u6237\u7AEF\u7684\u672C\u5730\u8DEF\u7531\u5730\u5740\u3002</p><div class="card"><div class="card-head"><h2>\u5168\u5C40\u4EE3\u7406</h2></div><div class="modes"><label><input type="radio" name="proxy-mode" value="direct" checked>\u76F4\u8FDE\uFF08\u9ED8\u8BA4\uFF09</label><label><input type="radio" name="proxy-mode" value="manual">\u624B\u52A8\u4EE3\u7406</label><label><input type="radio" name="proxy-mode" value="auto">\u81EA\u52A8\uFF08\u8DDF\u968F\u7CFB\u7EDF\uFF09</label></div><div class="field"><label>HTTP / HTTPS Proxy</label><input id="proxy-url" class="mono" placeholder="http://127.0.0.1:7897"></div><dl class="kv"><dt>\u5F53\u524D\u6765\u6E90</dt><dd id="upstream-source">\u76F4\u8FDE</dd><dt>\u6709\u6548\u4EE3\u7406</dt><dd id="upstream-effective" class="mono">direct</dd><dt>\u7CFB\u7EDF PAC</dt><dd id="upstream-pac" class="mono">\u2014</dd><dt>\u72B6\u6001</dt><dd id="upstream-state" class="status">\u5C1A\u672A\u6D4B\u8BD5</dd></dl><div class="actions"><button id="save-upstream" class="btn primary">\u4FDD\u5B58\u5E76\u70ED\u66F4\u65B0</button><button id="test-upstream" class="btn">\u6D4B\u8BD5\u8FDE\u63A5</button></div></div><div class="card"><div class="card-head"><h2>\u6309 URL \u8986\u76D6</h2></div><p class="status">\u4E3A\u7279\u5B9A Provider \u5355\u72EC\u8BBE\u7F6E\u4E0A\u6E38\u4EE3\u7406\uFF0C\u8986\u76D6\u5168\u5C40\u8BBE\u7F6E\u3002\u7559\u7A7A = \u7EE7\u627F\u5168\u5C40\u3002</p><div id="upstream-overrides" class="status">\u52A0\u8F7D\u4E2D</div><div class="actions"><button id="save-overrides" class="btn primary">\u4FDD\u5B58\u8986\u76D6\u5E76\u70ED\u66F4\u65B0</button></div></div></section>
51012
51100
  <section id="page-sessions" class="page"><div class="card-head"><div><h1>\u4F1A\u8BDD</h1><p class="lead">ACP \u538B\u7F29\u72B6\u6001\u4E0E\u4E0A\u6E38\u7528\u91CF\u3002</p></div><button id="refresh-sessions" class="btn">\u5237\u65B0</button></div><div class="card"><table><thead><tr><th>\u6807\u9898</th><th>\u534F\u8BAE</th><th>\u6807\u8BC6</th><th>\u8BF7\u6C42</th><th>\u4E0A\u4E0B\u6587</th><th>\u6700\u540E\u6D3B\u52A8</th></tr></thead><tbody id="sessions-body"></tbody></table></div></section>
51013
- <section id="page-settings" class="page"><h1>\u9AD8\u7EA7\u8BBE\u7F6E</h1><p class="lead">\u76F4\u63A5\u7F16\u8F91 providers JSON\uFF08\u6DFB\u52A0\u65B0 Provider\u3001\u914D\u7F6E\u6A21\u578B\u4E0A\u4E0B\u6587\u7A97\u53E3\u3001\u6309 URL \u4EE3\u7406\u7B49\uFF09\uFF1B\u4FDD\u5B58\u540E\u7ACB\u5373\u70ED\u66F4\u65B0\u3002</p><div class="card"><div class="field"><label>providers JSON</label><textarea id="providers-json">{}</textarea></div><div class="actions"><button id="save-providers" class="btn primary">\u4FDD\u5B58\u5E76\u5E94\u7528</button></div></div><div class="card"><div class="card-head"><h2>\u538B\u7F29\u53C2\u6570\uFF08\u5168\u5C40\uFF09</h2></div><p class="status">\u5168\u5C40 compress \u914D\u7F6E\uFF08modelContextLimit / maxContextLimit / emergencyThresholdPercent / nudgeGrowthTokens / preserveRecentMessages / preserveRecentTokens / minCompressRange / tiers / injectTool / injectNudge\uFF09\u3002\u4FDD\u5B58\u540E\u70ED\u66F4\u65B0\uFF0C\u65E0\u9700\u91CD\u542F\u3002\u7F6E\u7A7A = \u6062\u590D\u9ED8\u8BA4\u3002</p><div class="field"><label>compress JSON</label><textarea id="compress-json" placeholder='{"modelContextLimit": 200000}'></textarea></div><div class="actions"><button id="save-compress" class="btn primary">\u4FDD\u5B58\u5E76\u70ED\u66F4\u65B0</button></div></div></section></main></div><div id="toast" class="toast"></div><script>${WEB_CLIENT}</script></body></html>`;
51101
+ <section id="page-settings" class="page"><h1>\u9AD8\u7EA7\u8BBE\u7F6E</h1><p class="lead">\u76F4\u63A5\u7F16\u8F91 providers JSON\uFF08\u6DFB\u52A0\u65B0 Provider\u3001\u914D\u7F6E\u6A21\u578B\u4E0A\u4E0B\u6587\u7A97\u53E3\u3001\u6309 URL \u4EE3\u7406\u7B49\uFF09\uFF1B\u4FDD\u5B58\u540E\u7ACB\u5373\u70ED\u66F4\u65B0\u3002</p><div class="card"><div class="field"><label>providers JSON</label><textarea id="providers-json">{}</textarea></div><div class="actions"><button id="save-providers" class="btn primary">\u4FDD\u5B58\u5E76\u5E94\u7528</button></div></div><div class="card"><div class="card-head"><h2>\u538B\u7F29\u53C2\u6570\uFF08\u5168\u5C40\uFF09</h2></div><p class="status">\u5168\u5C40 compress \u914D\u7F6E\uFF08modelContextLimit / maxContextLimit / emergencyThresholdPercent / nudgeGrowthTokens / preserveRecentMessages / preserveRecentTokens / minCompressRangeChars\uFF08\u5355\u4F4D\uFF1A\u5B57\u7B26\uFF0C\u65E7\u540D minCompressRange \u517C\u5BB9\uFF09 / tiers / injectTool / injectNudge\uFF09\u3002\u4FDD\u5B58\u540E\u70ED\u66F4\u65B0\uFF0C\u65E0\u9700\u91CD\u542F\u3002\u7F6E\u7A7A = \u6062\u590D\u9ED8\u8BA4\u3002</p><div class="field"><label>compress JSON</label><textarea id="compress-json" placeholder='{"modelContextLimit": 200000}'></textarea></div><div class="actions"><button id="save-compress" class="btn primary">\u4FDD\u5B58\u5E76\u70ED\u66F4\u65B0</button></div></div></section></main></div><div id="toast" class="toast"></div><script>${WEB_CLIENT}</script></body></html>`;
51014
51102
  }
51015
51103
 
51016
51104
  // src/web/api.ts
@@ -51427,6 +51515,8 @@ function handleAcpStatus(args, ctx) {
51427
51515
  const nudge = ctx.nudge;
51428
51516
  const ranges = nudge?.compressibleRanges ?? [];
51429
51517
  const protectedRanges = nudge?.protectedRanges ?? [];
51518
+ const archive = preCompactionArchiveOf(ctx.session);
51519
+ const archivedIds = Object.keys(archive);
51430
51520
  const extra = [];
51431
51521
  if (nudge) {
51432
51522
  extra.push("");
@@ -51436,6 +51526,13 @@ function handleAcpStatus(args, ctx) {
51436
51526
  extra.push("");
51437
51527
  extra.push(formatRanges(ranges, protectedRanges));
51438
51528
  }
51529
+ if (archivedIds.length > 0) {
51530
+ extra.push("");
51531
+ extra.push(`PRE-COMPACTION ARCHIVE \u2014 ${archivedIds.length} block(s): content was replaced by the client's native compaction summary, so it is no longer in the session history and decompress is unavailable.`);
51532
+ for (const id of archivedIds) {
51533
+ extra.push(` ${id} \u2014 ${archive[id].reason}`);
51534
+ }
51535
+ }
51439
51536
  return extra.length > 0 ? `${base}
51440
51537
  ${extra.join("\n")}` : base;
51441
51538
  }
@@ -54239,6 +54336,32 @@ function handlePluginRegister(payload, res) {
54239
54336
  queuePluginRegister(conversationId2, agent, parsed.identity === true);
54240
54337
  res.end(JSON.stringify({ ok: true, conversationId: conversationId2, agent }));
54241
54338
  }
54339
+ function handlePluginCompact(payload, res) {
54340
+ let parsed;
54341
+ try {
54342
+ parsed = JSON.parse(payload);
54343
+ } catch {
54344
+ res.writeHead(400, { "content-type": "application/json" });
54345
+ res.end(JSON.stringify({ ok: false, error: "invalid JSON body" }));
54346
+ return;
54347
+ }
54348
+ const conversationId2 = typeof parsed.conversationId === "string" ? parsed.conversationId.trim() : "";
54349
+ if (!conversationId2) {
54350
+ res.writeHead(400, { "content-type": "application/json" });
54351
+ res.end(JSON.stringify({ ok: false, error: "conversationId is required" }));
54352
+ return;
54353
+ }
54354
+ const entry = conversations.get(conversationId2);
54355
+ const session = entry ? peekSession(entry.sessionId) : void 0;
54356
+ if (!entry || !session) {
54357
+ res.writeHead(404, { "content-type": "application/json" });
54358
+ res.end(JSON.stringify({ ok: false, error: "unknown plugin conversation (no model request has arrived with this conversation id yet)" }));
54359
+ return;
54360
+ }
54361
+ markCompactionBoundary(session);
54362
+ entry.lastSeen = Date.now();
54363
+ res.end(JSON.stringify({ ok: true, conversationId: conversationId2 }));
54364
+ }
54242
54365
  function handlePluginManifest(res) {
54243
54366
  res.writeHead(200, { "content-type": "application/json" });
54244
54367
  res.end(JSON.stringify({
@@ -54262,10 +54385,14 @@ function handlePluginStatus(conversationId2, res, fallbackLatest = false) {
54262
54385
  let session = entry ? peekSession(entry.sessionId) : void 0;
54263
54386
  let viaFallback = false;
54264
54387
  if ((!entry || !session) && fallbackLatest) {
54265
- const latest = listSessions().filter((s3) => typeof s3.lastSeen === "number").sort((a, b2) => (b2.lastSeen ?? 0) - (a.lastSeen ?? 0))[0];
54388
+ const latest = listSessions().filter((s3) => s3.restored !== true).sort((a, b2) => (b2.lastSeen ?? 0) - (a.lastSeen ?? 0))[0];
54266
54389
  if (latest) {
54267
54390
  session = latest;
54268
54391
  viaFallback = true;
54392
+ } else {
54393
+ res.writeHead(404, { "content-type": "application/json" });
54394
+ res.end(JSON.stringify({ ok: false, error: "no session with activity since boot \u2014 issue a model request or pass the conversation id" }));
54395
+ return;
54269
54396
  }
54270
54397
  }
54271
54398
  if (!session) {
@@ -54472,6 +54599,12 @@ data: ${JSON.stringify({ type: "content_block_delta", index, delta: { type: delt
54472
54599
  return new Promise((r) => res.once("drain", () => r()));
54473
54600
  }
54474
54601
  };
54602
+ const settleUsage = () => {
54603
+ if (session && (acc.inputTokens !== void 0 || acc.outputTokens !== void 0 || acc.cachedTokens !== void 0)) {
54604
+ applyUsageSample(session, acc, protocol);
54605
+ markDirty(session);
54606
+ }
54607
+ };
54475
54608
  const pushField = (field, index, text) => {
54476
54609
  const s3 = filterFor(field, index);
54477
54610
  const clean = s3.filter.push(text);
@@ -54577,10 +54710,14 @@ data: ${JSON.stringify({ type: "content_block_delta", index, delta: { type: delt
54577
54710
  }
54578
54711
  const rest = flushTails();
54579
54712
  if (rest.length > 0 && !res.destroyed && !res.writableEnded) await write(rest);
54580
- if (session && (acc.inputTokens !== void 0 || acc.outputTokens !== void 0 || acc.cachedTokens !== void 0)) {
54581
- applyUsageSample(session, acc, protocol);
54582
- markDirty(session);
54713
+ settleUsage();
54714
+ } catch (e) {
54715
+ settleUsage();
54716
+ if (res.destroyed || res.writableEnded) {
54717
+ log2?.("client aborted mid-stream");
54718
+ return;
54583
54719
  }
54720
+ throw e;
54584
54721
  } finally {
54585
54722
  reader.releaseLock();
54586
54723
  res.end();
@@ -54614,6 +54751,12 @@ async function pipePluginResponsesWithStrip(stream2, res, session, log2) {
54614
54751
  return new Promise((r) => res.once("drain", () => r()));
54615
54752
  }
54616
54753
  };
54754
+ const settleUsage = () => {
54755
+ if (session && (acc.inputTokens !== void 0 || acc.outputTokens !== void 0 || acc.cachedTokens !== void 0)) {
54756
+ applyUsageSample(session, acc, "responses");
54757
+ markDirty(session);
54758
+ }
54759
+ };
54617
54760
  let lastDeltaMeta = null;
54618
54761
  const flushTail = (after) => {
54619
54762
  const tail = tagFilter.flush();
@@ -54686,10 +54829,14 @@ async function pipePluginResponsesWithStrip(stream2, res, session, log2) {
54686
54829
  const rest = flushTail("");
54687
54830
  if (rest.length > 0) await write(rest);
54688
54831
  }
54689
- if (session && (acc.inputTokens !== void 0 || acc.outputTokens !== void 0 || acc.cachedTokens !== void 0)) {
54690
- applyUsageSample(session, acc, "responses");
54691
- markDirty(session);
54832
+ settleUsage();
54833
+ } catch (e) {
54834
+ settleUsage();
54835
+ if (res.destroyed || res.writableEnded) {
54836
+ log2?.("client aborted mid-stream");
54837
+ return;
54692
54838
  }
54839
+ throw e;
54693
54840
  } finally {
54694
54841
  reader.releaseLock();
54695
54842
  res.end();
@@ -54715,10 +54862,16 @@ function rebuildEvent(rawEvent, ev) {
54715
54862
  async function pipePluginJson(stream2, res, session, protocol) {
54716
54863
  const reader = stream2.getReader();
54717
54864
  const chunks = [];
54718
- for (; ; ) {
54719
- const { done, value } = await reader.read();
54720
- if (done) break;
54721
- if (value && value.length > 0) chunks.push(Buffer.from(value));
54865
+ try {
54866
+ for (; ; ) {
54867
+ const { done, value } = await reader.read();
54868
+ if (done) break;
54869
+ if (value && value.length > 0) chunks.push(Buffer.from(value));
54870
+ }
54871
+ } catch (e) {
54872
+ if (!res.writableEnded) res.end();
54873
+ if (res.destroyed || res.writableEnded) return;
54874
+ throw e;
54722
54875
  }
54723
54876
  reader.releaseLock();
54724
54877
  const text = Buffer.concat(chunks).toString("utf8");
@@ -56369,7 +56522,10 @@ async function startServer(opts) {
56369
56522
  await handle(req, res, opts, core, config, log2, instanceId, instanceStartedAt);
56370
56523
  } catch (err2) {
56371
56524
  const msg = String(err2);
56372
- log2("error", msg);
56525
+ const e = err2;
56526
+ const clientAbort = (e?.name === "AbortError" || /abort/i.test(String(e?.message ?? ""))) && (res.destroyed || res.writableEnded);
56527
+ if (clientAbort) log2("info", `client aborted mid-stream: ${msg}`);
56528
+ else log2("error", msg);
56373
56529
  if (!res.headersSent) {
56374
56530
  const status = msg.includes("exceeds") ? 413 : 502;
56375
56531
  res.writeHead(status, { "content-type": "application/json" });
@@ -56665,6 +56821,17 @@ async function handle(req, res, opts, core, config, log2, instanceId, instanceSt
56665
56821
  return;
56666
56822
  }
56667
56823
  }
56824
+ if (req.method === "POST" && req.url === "/__bili/plugin/compact") {
56825
+ try {
56826
+ const body = await readBody(req);
56827
+ handlePluginCompact(body.toString("utf8"), res);
56828
+ return;
56829
+ } catch (err2) {
56830
+ res.writeHead(err2 instanceof BodyTooLargeError ? 413 : 400, { "content-type": "application/json" });
56831
+ res.end(JSON.stringify({ ok: false, error: String(err2) }));
56832
+ return;
56833
+ }
56834
+ }
56668
56835
  if (isAdminPath) {
56669
56836
  res.writeHead(404, { "content-type": "application/json" });
56670
56837
  res.end(JSON.stringify({ error: { type: "not_found", message: "no such management endpoint" } }));
@@ -56894,6 +57061,7 @@ ${bodyBuffer.toString("utf8")}`);
56894
57061
  nativeFromFallback = false;
56895
57062
  log2("info", `[${session.id}] self-healed context window: ${resolved} \u2192 ${learnedLimit} (learned from an upstream overflow)`);
56896
57063
  }
57064
+ const nativeWindow = reqConfig.modelContextLimit;
56897
57065
  if (shouldReserveOutputHeadroom(protocol)) {
56898
57066
  const p2 = parsed;
56899
57067
  const rawMax = p2.max_tokens ?? p2.max_completion_tokens ?? p2.max_output_tokens;
@@ -56908,7 +57076,22 @@ ${bodyBuffer.toString("utf8")}`);
56908
57076
  acquireInFlight(session);
56909
57077
  try {
56910
57078
  await withSessionLock(session, async () => {
56911
- const runPrepare = () => countTokens ? prepareCountTokens(parsed, core, reqConfig, log2, session) : protocol === "anthropic" ? prepareAnthropic(parsed, req, opts, core, reqConfig, reqPrompts, log2, session, pluginMode) : protocol === "openai" ? prepareOpenai(parsed, req, opts, core, reqConfig, reqPrompts, log2, session, pluginMode) : responsesCompact ? prepareResponsesCompact(bodyBuffer, parsed, session, req, core, reqConfig, log2) : prepareResponses(parsed, req, opts, core, reqConfig, reqPrompts, log2, session, responsesIdentity, pluginMode, upstreamOrigin);
57079
+ const runPrepare = () => countTokens ? prepareCountTokens(parsed, core, reqConfig, log2, session) : protocol === "anthropic" ? prepareAnthropic(parsed, req, opts, core, reqConfig, reqPrompts, log2, session, pluginMode) : protocol === "openai" ? prepareOpenai(parsed, req, opts, core, reqConfig, reqPrompts, log2, session, pluginMode, nativeWindow) : responsesCompact ? prepareResponsesCompact(bodyBuffer, parsed, session, req, core, reqConfig, log2) : prepareResponses(parsed, req, opts, core, reqConfig, reqPrompts, log2, session, responsesIdentity, pluginMode, upstreamOrigin, nativeWindow);
57080
+ const isCodexCompactTrigger = protocol === "responses" && !responsesCompact && isCodexClient(req.headers) && hasCompactionTrigger(parsed.input);
57081
+ if (isCodexCompactTrigger) {
57082
+ const mode = codexCompactMode();
57083
+ const gatePre = codexCompactGatePre(session, reqConfig.modelContextLimit);
57084
+ if (mode === "intercept" && gatePre) prepared = runPrepare();
57085
+ if (prepared?.codexForge) {
57086
+ await forward(req, res, opts, prepared.body, prepared, core, reqConfig, log2, route, instanceId, affinity);
57087
+ rememberPluginMessages(sessionId, prepared.processedMessages, prepared.originalMessages, prepared.nudge);
57088
+ return;
57089
+ }
57090
+ const why = mode !== "intercept" ? "BILI_CODEX_COMPACT=pass" : !gatePre ? "gate preconditions not met" : "transform/forge failed";
57091
+ log2("info", `[${session.id}] codex compaction_trigger request not intercepted (${why}) \u2014 forwarding verbatim (no preflight, no rebuild, no window clamp)`);
57092
+ await forward(req, res, opts, bodyBuffer, null, core, reqConfig, log2, route, instanceId, affinity);
57093
+ return;
57094
+ }
56912
57095
  prepared = runPrepare();
56913
57096
  if (!countTokens && !responsesCompact) {
56914
57097
  const outcome = await preflightCompressIfNeeded(
@@ -56992,7 +57175,7 @@ function diagNudge(turn, sessionId, tokenCount, limit, model, willInject) {
56992
57175
  const interval = b2["nudgeGrowthTokens"] ?? 0;
56993
57176
  const pendingT1 = b2["pendingT1"] ?? 0;
56994
57177
  const ref = b2["growthReference"] ?? 0;
56995
- const inject = !n.shouldInject ? "idle" : willInject ? `INJECT T${n.tier ?? "?"}` : `ARMED-SUPPRESSED T${n.tier ?? "?"}`;
57178
+ const inject = willInject ? n.shouldInject ? `INJECT T${n.tier ?? "?"}` : `INJECT-ESC T${n.tier ?? "?"}` : n.shouldInject ? `ARMED-SUPPRESSED T${n.tier ?? "?"}` : "idle";
56996
57179
  const modelTag = model ? ` model=${model}` : "";
56997
57180
  return `[${sessionId}] nudge ${inject}: usage=${pct2} (${tokenCount}/${limit}), growth=${growth}/${floor} (ref=${ref}, interval=${interval}), pendingT1=${pendingT1}/${interval}${modelTag}, reason="${n.reason.slice(0, 120)}"`;
56998
57181
  }
@@ -57016,6 +57199,7 @@ function prepareAnthropic(parsed, req, opts, core, config, prompts, log2, sessio
57016
57199
  originalMessages = msgs;
57017
57200
  extractSystem(parsed.system);
57018
57201
  const tokenCount = session.stats.lastInputTokens;
57202
+ const activeBefore = new Set(session.state.blocks.filter((b2) => b2.active).map((b2) => b2.blockId));
57019
57203
  const turn = core.processTurn({ messages: msgs, state: session.state, config, tokenCount, renderTags: "text-only" });
57020
57204
  session.state = turn.state;
57021
57205
  session.stats.compressCreditTokens = 0;
@@ -57027,9 +57211,10 @@ function prepareAnthropic(parsed, req, opts, core, config, prompts, log2, sessio
57027
57211
  if (t) session.meta.title = t;
57028
57212
  }
57029
57213
  log2("info", diagTagSummary(turn.messages, sessionId, "text-only"));
57030
- const willInjectNudge = opts.compress.injectNudge && !!turn.nudge?.shouldInject;
57214
+ const willInjectNudge = opts.compress.injectNudge && !!turn.nudge && (turn.nudge.shouldInject || emergencyNudge(turn.nudge));
57031
57215
  log2("info", diagNudge(turn, sessionId, tokenCount, config.modelContextLimit, parsed.model, willInjectNudge));
57032
57216
  processedMessages = stripKernelSummaries(turn.messages, turn.state);
57217
+ applyCompactionArchive(session, activeBefore, new Set(msgs.map((m2) => m2.id)), log2);
57033
57218
  reapOrphanBlocks(session, msgs, deactivateBlock);
57034
57219
  rebuiltMessages = coreToAnthropic(processedMessages, cacheControls);
57035
57220
  systemOut = injectSystem(parsed, opts, prompts);
@@ -57055,7 +57240,36 @@ function prepareAnthropic(parsed, req, opts, core, config, prompts, log2, sessio
57055
57240
  delete rebuilt.prompt_cache_key;
57056
57241
  return { body: JSON.stringify(rebuilt), session, processedMessages, originalMessages, anthropicSystem: parsed.system, protocol: "anthropic", stream: stream2, compressInjected: injectTools, pluginMode, nudge, prompts, renderTags: "text-only" };
57057
57242
  }
57058
- function prepareOpenai(parsed, req, opts, core, config, prompts, log2, session, pluginMode) {
57243
+ var OUTPUT_CLAMP_MARGIN_PCT = 0.05;
57244
+ var OUTPUT_CLAMP_MIN_MARGIN = 2048;
57245
+ var OUTPUT_CLAMP_FLOOR = 1024;
57246
+ var EMERGENCY_NUDGE_ESCALATION_PCT = 0.7;
57247
+ function estimateInputTokens(processedMessages, systemText, tools, lastInputTokens) {
57248
+ const est = estimateCoreMessages(processedMessages) + defaultCountTokens(systemText ?? "") + defaultCountTokens(JSON.stringify(tools ?? []));
57249
+ return Math.max(lastInputTokens > 0 ? lastInputTokens : 0, est);
57250
+ }
57251
+ function clampOutputBudget(requested, inputEstimate, nativeWindow) {
57252
+ const margin = Math.max(OUTPUT_CLAMP_MIN_MARGIN, Math.ceil(inputEstimate * OUTPUT_CLAMP_MARGIN_PCT));
57253
+ const cap = nativeWindow - inputEstimate - margin;
57254
+ if (cap < OUTPUT_CLAMP_FLOOR || cap >= requested) return void 0;
57255
+ return cap;
57256
+ }
57257
+ function emergencyNudge(nudge, escalationPct = EMERGENCY_NUDGE_ESCALATION_PCT) {
57258
+ if (!nudge || nudge.shouldInject) return false;
57259
+ if (nudge.compressibleRanges.length === 0) return false;
57260
+ return nudge.contextUsage >= escalationPct;
57261
+ }
57262
+ function clampOutgoingOutput(rebuilt, field, ctx, sessionId, log2) {
57263
+ const raw = rebuilt[field];
57264
+ if (typeof raw !== "number") return;
57265
+ const inputEstimate = estimateInputTokens(ctx.processedMessages, ctx.systemText, ctx.tools, ctx.lastInputTokens);
57266
+ const capped = clampOutputBudget(raw, inputEstimate, ctx.nativeWindow);
57267
+ if (capped !== void 0) {
57268
+ rebuilt[field] = capped;
57269
+ log2("info", `[${sessionId}] output budget clamped ${raw} -> ${capped} (input~${inputEstimate}, window=${ctx.nativeWindow}); prevents input+output overflow (#453)`);
57270
+ }
57271
+ }
57272
+ function prepareOpenai(parsed, req, opts, core, config, prompts, log2, session, pluginMode, nativeWindow) {
57059
57273
  const sessionId = session.id;
57060
57274
  const stream2 = parsed.stream === true;
57061
57275
  ++session.stats.requests;
@@ -57074,6 +57288,7 @@ function prepareOpenai(parsed, req, opts, core, config, prompts, log2, session,
57074
57288
  openaiSystemText = systemText;
57075
57289
  originalMessages = msgs;
57076
57290
  const tokenCount = session.stats.lastInputTokens;
57291
+ const activeBefore = new Set(session.state.blocks.filter((b2) => b2.active).map((b2) => b2.blockId));
57077
57292
  const turn = core.processTurn({ messages: msgs, state: session.state, config, tokenCount, renderTags: "text-only" });
57078
57293
  session.state = turn.state;
57079
57294
  session.stats.compressCreditTokens = 0;
@@ -57085,9 +57300,10 @@ function prepareOpenai(parsed, req, opts, core, config, prompts, log2, session,
57085
57300
  if (t) session.meta.title = t;
57086
57301
  }
57087
57302
  log2("info", diagTagSummary(turn.messages, sessionId, "text-only"));
57088
- const willInjectNudge = opts.compress.injectNudge && !!turn.nudge?.shouldInject && shouldInject;
57303
+ const willInjectNudge = opts.compress.injectNudge && !!turn.nudge && shouldInject && (turn.nudge.shouldInject || emergencyNudge(turn.nudge));
57089
57304
  log2("info", diagNudge(turn, sessionId, tokenCount, config.modelContextLimit, parsed.model, willInjectNudge));
57090
57305
  processedMessages = stripKernelSummaries(turn.messages, turn.state);
57306
+ applyCompactionArchive(session, activeBefore, new Set(msgs.map((m2) => m2.id)), log2);
57091
57307
  reapOrphanBlocks(session, msgs, deactivateBlock);
57092
57308
  rebuiltMessages = systemToUser(coreToOpenai(processedMessages));
57093
57309
  const sysParts = [];
@@ -57111,6 +57327,7 @@ function prepareOpenai(parsed, req, opts, core, config, prompts, log2, session,
57111
57327
  processedMessages = [];
57112
57328
  }
57113
57329
  const rebuilt = { ...parsed, messages: rebuiltMessages, tools: toolsOut };
57330
+ clampOutgoingOutput(rebuilt, typeof parsed.max_completion_tokens === "number" ? "max_completion_tokens" : "max_tokens", { systemText: openaiSystemText, tools: toolsOut, processedMessages, lastInputTokens: session.stats.lastInputTokens, nativeWindow }, sessionId, log2);
57114
57331
  delete rebuilt.prompt_cache_retention;
57115
57332
  if (stream2 && rebuilt.stream_options === void 0) {
57116
57333
  rebuilt.stream_options = { include_usage: true };
@@ -57119,7 +57336,7 @@ function prepareOpenai(parsed, req, opts, core, config, prompts, log2, session,
57119
57336
  markDirty(session);
57120
57337
  return { body: JSON.stringify(rebuilt), session, processedMessages, originalMessages, protocol: "openai", stream: stream2, compressInjected: injectTools, pluginMode, nudge, prompts, openaiSystemText, renderTags: "text-only" };
57121
57338
  }
57122
- function prepareResponses(parsed, req, opts, core, config, prompts, log2, session, identity, pluginMode, upstreamOrigin) {
57339
+ function prepareResponses(parsed, req, opts, core, config, prompts, log2, session, identity, pluginMode, upstreamOrigin, nativeWindow) {
57123
57340
  const sessionId = session.id;
57124
57341
  const stream2 = parsed.stream === true;
57125
57342
  ++session.stats.requests;
@@ -57176,7 +57393,7 @@ function prepareResponses(parsed, req, opts, core, config, prompts, log2, sessio
57176
57393
  if (t) session.meta.title = t;
57177
57394
  }
57178
57395
  log2("info", diagTagSummary(turn.messages, sessionId, "text-only"));
57179
- const willInjectNudge = opts.compress.injectNudge && !!turn.nudge?.shouldInject && shouldInject && !isCompactionTrigger;
57396
+ const willInjectNudge = opts.compress.injectNudge && !!turn.nudge && shouldInject && !isCompactionTrigger && (turn.nudge.shouldInject || emergencyNudge(turn.nudge));
57180
57397
  log2("info", diagNudge(turn, sessionId, tokenCount, config.modelContextLimit, parsed.model, willInjectNudge));
57181
57398
  processedMessages = stripKernelSummaries(turn.messages, turn.state);
57182
57399
  reapOrphanBlocks(session, msgs, deactivateBlock);
@@ -57227,6 +57444,9 @@ function prepareResponses(parsed, req, opts, core, config, prompts, log2, sessio
57227
57444
  log2("info", `[${sessionId}] codex compact intercepted (trigger); forged SSE with ${summaries.length} block summary(s), upstream not contacted`);
57228
57445
  }
57229
57446
  const rebuilt = { ...parsed, input: rebuiltInput, tools: toolsOut };
57447
+ if (!isCompactionTrigger) {
57448
+ clampOutgoingOutput(rebuilt, "max_output_tokens", { systemText: (responsesProjection?.systemParts ?? []).join("\n"), tools: toolsOut, processedMessages, lastInputTokens: session.stats.lastInputTokens, nativeWindow }, sessionId, log2);
57449
+ }
57230
57450
  const promptCacheKey = resolvePromptCacheKey(
57231
57451
  rebuilt.prompt_cache_key,
57232
57452
  identity,
@@ -57317,7 +57537,7 @@ function prepareResponsesCompact(body, parsed, session, req, core, config, log2)
57317
57537
  compressInjected: false,
57318
57538
  resetAfterSuccess: true
57319
57539
  };
57320
- if (codexCompactMode() !== "intercept" || !isCodexClient(req.headers) || !Array.isArray(parsed.input)) {
57540
+ if (codexCompactMode() !== "intercept" || !isCodexClient(req.headers) || !Array.isArray(parsed.input) || !codexCompactGatePre(session, config.modelContextLimit)) {
57321
57541
  return base;
57322
57542
  }
57323
57543
  const prevState = session.state;
@@ -57709,187 +57929,205 @@ ${hdrText}
57709
57929
  return;
57710
57930
  }
57711
57931
  if (prepared?.pluginMode) {
57712
- if (prepared.stream) {
57713
- if (prepared.protocol === "responses") {
57714
- await pipePluginResponsesWithStrip(upstream.body, res, prepared.session, (msg) => log2("info", `[${prepared.session.id}] ${msg}`));
57715
- } else {
57716
- await pipePluginChatWithStrip(upstream.body, res, prepared.protocol, prepared.session, (msg) => log2("info", `[${prepared.session.id}] ${msg}`));
57717
- }
57718
- } else {
57719
- await pipePluginJson(upstream.body, res, prepared.session, prepared.protocol);
57720
- }
57721
- clearUpstreamTimer();
57722
- return;
57723
- }
57724
- let responseBody = upstream.body;
57725
- if (prepared !== null && prepared.stream && maxFakeCompletionRetries() > 0) {
57726
- const resolvedBuf = await resolveFakeCompletion(upstream.body, {
57727
- protocol: prepared.protocol,
57728
- body,
57729
- upstreamUrl,
57730
- reqHeaders: buildForwardHeaders(headers),
57731
- proxyUrl,
57732
- signal: clientAbort.signal,
57733
- session: prepared.session,
57734
- log: log2
57735
- });
57736
- responseBody = bufferToStream(resolvedBuf);
57737
- }
57738
- const useRewriter = prepared !== null && prepared.compressInjected && prepared.processedMessages.length > 0;
57739
- if (!useRewriter || prepared === null) {
57740
- if (prepared && prepared.resetAfterSuccess) {
57741
- const [toClient, toObserve] = responseBody.tee();
57742
- const observed = observeResponsesTerminalState(toObserve, prepared.stream);
57743
- const tagLog = (msg) => log2("info", `[${prepared.session.id}] ${msg}`);
57744
- if ((upstream.headers.get("content-type") ?? "").includes("text/event-stream")) {
57745
- await pipePluginResponsesWithStrip(toClient, res, void 0, tagLog);
57746
- } else {
57747
- await pipeThrough(toClient, res);
57748
- }
57749
- clearUpstreamTimer();
57750
- const terminal = await observed;
57751
- if (terminal === "completed") {
57752
- markNativeCompactionBoundary(prepared.session);
57753
- log2("info", `[${prepared.session.id}] native Responses compact completed; rebase scheduled for next Responses turn`);
57754
- } else {
57755
- log2("warn", `[${prepared.session.id}] native compact response terminal=${terminal}; rebase NOT scheduled`);
57932
+ try {
57933
+ let pluginBody = upstream.body;
57934
+ if (prepared.stream && maxFakeCompletionRetries() > 0) {
57935
+ const resolvedBuf = await resolveFakeCompletion(pluginBody, {
57936
+ protocol: prepared.protocol,
57937
+ body,
57938
+ upstreamUrl,
57939
+ reqHeaders: buildForwardHeaders(headers),
57940
+ proxyUrl,
57941
+ signal: clientAbort.signal,
57942
+ session: prepared.session,
57943
+ log: log2
57944
+ });
57945
+ pluginBody = bufferToStream(resolvedBuf);
57756
57946
  }
57757
- } else if (prepared && prepared.stream && (upstream.headers.get("content-type") ?? "").includes("text/event-stream")) {
57758
- const p2 = prepared;
57759
- const tagLog = (msg) => log2("info", `[${p2.session.id}] ${msg}`);
57760
- if (p2.protocol === "responses") {
57761
- await pipePluginResponsesWithStrip(responseBody, res, void 0, tagLog);
57947
+ if (prepared.stream) {
57948
+ if (prepared.protocol === "responses") {
57949
+ await pipePluginResponsesWithStrip(pluginBody, res, prepared.session, (msg) => log2("info", `[${prepared.session.id}] ${msg}`));
57950
+ } else {
57951
+ await pipePluginChatWithStrip(pluginBody, res, prepared.protocol, prepared.session, (msg) => log2("info", `[${prepared.session.id}] ${msg}`));
57952
+ }
57762
57953
  } else {
57763
- await pipePluginChatWithStrip(responseBody, res, p2.protocol, void 0, tagLog);
57954
+ await pipePluginJson(pluginBody, res, prepared.session, prepared.protocol);
57764
57955
  }
57765
- clearUpstreamTimer();
57766
- } else if (prepared && (upstream.headers.get("content-type") ?? "").includes("application/json")) {
57767
- await pipePluginJson(responseBody, res, void 0, prepared.protocol);
57768
- clearUpstreamTimer();
57769
- } else {
57770
- await pipeThrough(responseBody, res);
57956
+ } finally {
57771
57957
  clearUpstreamTimer();
57772
57958
  }
57773
57959
  return;
57774
57960
  }
57775
- const ctx = {
57776
- core,
57777
- config,
57778
- messages: prepared.originalMessages,
57779
- session: prepared.session,
57780
- log: (msg) => log2("info", `[${prepared.session.id}] ${msg}`),
57781
- debug: opts.debug
57782
- };
57783
- if (prepared.stream) {
57784
- let streamToRead = responseBody;
57785
- let dumpRaw;
57786
- if (opts.dumpSse) {
57787
- const [a, b2] = responseBody.tee();
57788
- streamToRead = a;
57789
- dumpRaw = dumpStreamToFile(b2, opts.dumpSse, `${Date.now()}-${safeSessionId(prepared.session.id)}-raw.sse`);
57790
- }
57791
- try {
57792
- const parsedReq = JSON.parse(typeof body === "string" ? body : body.toString("utf8"));
57793
- const reqHeaders = buildForwardHeaders(headers);
57794
- const textProtocol = prepared.protocol === "responses" && !!prepared.responsesTextProtocol;
57795
- const systemPrompt = textProtocol ? buildCompressHybridSystemPrompt(prepared.prompts ?? defaultPrompts) : buildCompressSystemPrompt(prepared.prompts ?? defaultPrompts);
57796
- const adapter = pickAdapter(prepared.protocol, parsedReq, textProtocol, prepared.responsesProjection, prepared.anthropicSystem, prepared.openaiSystemText);
57797
- const refreshFolded = (current) => {
57798
- const turn = core.processTurn({
57799
- messages: prepared.originalMessages,
57800
- state: prepared.session.state,
57801
- config,
57802
- tokenCount: prepared.session.stats.lastInputTokens,
57803
- renderTags: prepared.renderTags ?? "text-only"
57804
- });
57805
- prepared.session.state = turn.state;
57806
- const records = current.filter((m2) => typeof m2.id === "string" && m2.id.startsWith("acp_loop_"));
57807
- return stripKernelSummaries([...turn.messages, ...records], turn.state);
57808
- };
57809
- const loop = runCompressLoop(
57810
- streamToRead,
57811
- { core, config, messages: prepared.processedMessages.length > 0 ? prepared.processedMessages : prepared.originalMessages, compressMessages: prepared.originalMessages, session: prepared.session, log: ctx.log, proxyUrl, protocol: prepared.protocol, textProtocol, debug: opts.debug, nudge: prepared.nudge, refreshFolded },
57812
- parsedReq,
57813
- { url: upstreamUrl, headers: reqHeaders },
57814
- adapter,
57815
- systemPrompt,
57816
- clientAbort.signal
57817
- );
57818
- let protocolFragmentWarned = false;
57819
- for await (const chunk of loop) {
57820
- if (res.destroyed || res.writableEnded) break;
57821
- {
57822
- const s3 = chunk.toString("utf8");
57823
- if (s3.includes("<acp ") || s3.includes("</acp")) {
57824
- log2("warn", `[${prepared.session.id}] tag echo: ${prepared.protocol} response stream contains <acp tag`);
57825
- } else if (!protocolFragmentWarned && containsToolCallXmlFragment(s3)) {
57826
- protocolFragmentWarned = true;
57827
- log2("warn", `[${prepared.session.id}] tag echo: ${prepared.protocol} response stream contains tool-call XML fragment (possible tag echo; not stripped)`);
57828
- }
57829
- }
57830
- res.write(chunk);
57831
- if (res.writableNeedDrain) {
57832
- await Promise.race([
57833
- new Promise((r) => res.once("drain", () => r())),
57834
- new Promise((r) => res.once("close", () => r()))
57835
- ]);
57961
+ let responseBody = upstream.body;
57962
+ try {
57963
+ if (prepared !== null && prepared.stream && maxFakeCompletionRetries() > 0) {
57964
+ const resolvedBuf = await resolveFakeCompletion(upstream.body, {
57965
+ protocol: prepared.protocol,
57966
+ body,
57967
+ upstreamUrl,
57968
+ reqHeaders: buildForwardHeaders(headers),
57969
+ proxyUrl,
57970
+ signal: clientAbort.signal,
57971
+ session: prepared.session,
57972
+ log: log2
57973
+ });
57974
+ responseBody = bufferToStream(resolvedBuf);
57975
+ }
57976
+ const useRewriter = prepared !== null && prepared.compressInjected && prepared.processedMessages.length > 0;
57977
+ if (!useRewriter || prepared === null) {
57978
+ if (prepared && prepared.resetAfterSuccess) {
57979
+ const [toClient, toObserve] = responseBody.tee();
57980
+ const observed = observeResponsesTerminalState(toObserve, prepared.stream);
57981
+ const tagLog = (msg) => log2("info", `[${prepared.session.id}] ${msg}`);
57982
+ if ((upstream.headers.get("content-type") ?? "").includes("text/event-stream")) {
57983
+ await pipePluginResponsesWithStrip(toClient, res, void 0, tagLog);
57984
+ } else {
57985
+ await pipeThrough(toClient, res);
57986
+ }
57987
+ const terminal = await observed;
57988
+ if (terminal === "completed") {
57989
+ markNativeCompactionBoundary(prepared.session);
57990
+ log2("info", `[${prepared.session.id}] native Responses compact completed; rebase scheduled for next Responses turn`);
57991
+ } else {
57992
+ log2("warn", `[${prepared.session.id}] native compact response terminal=${terminal}; rebase NOT scheduled`);
57993
+ }
57994
+ } else if (prepared && prepared.stream && (upstream.headers.get("content-type") ?? "").includes("text/event-stream")) {
57995
+ const p2 = prepared;
57996
+ const tagLog = (msg) => log2("info", `[${p2.session.id}] ${msg}`);
57997
+ if (p2.protocol === "responses") {
57998
+ await pipePluginResponsesWithStrip(responseBody, res, void 0, tagLog);
57999
+ } else {
58000
+ await pipePluginChatWithStrip(responseBody, res, p2.protocol, void 0, tagLog);
57836
58001
  }
57837
- if (res.destroyed || res.writableEnded) break;
58002
+ } else if (prepared && (upstream.headers.get("content-type") ?? "").includes("application/json")) {
58003
+ await pipePluginJson(responseBody, res, void 0, prepared.protocol);
58004
+ } else {
58005
+ await pipeThrough(responseBody, res);
57838
58006
  }
57839
- res.end();
57840
- } catch (e) {
57841
- emitStreamError(res, prepared.protocol, e?.message ?? String(e), (m2) => log2("error", `[${prepared.session.id}] ${m2}`));
57842
- } finally {
57843
- clearUpstreamTimer();
57844
- if (dumpRaw) await dumpRaw;
58007
+ return;
57845
58008
  }
57846
- } else {
57847
- try {
57848
- const buf = await upstream.arrayBuffer();
57849
- const text = Buffer.from(buf).toString("utf8");
58009
+ const ctx = {
58010
+ core,
58011
+ config,
58012
+ messages: prepared.originalMessages,
58013
+ session: prepared.session,
58014
+ log: (msg) => log2("info", `[${prepared.session.id}] ${msg}`),
58015
+ debug: opts.debug
58016
+ };
58017
+ if (prepared.stream) {
58018
+ let streamToRead = responseBody;
58019
+ let dumpRaw;
58020
+ if (opts.dumpSse) {
58021
+ const [a, b2] = responseBody.tee();
58022
+ streamToRead = a;
58023
+ dumpRaw = dumpStreamToFile(b2, opts.dumpSse, `${Date.now()}-${safeSessionId(prepared.session.id)}-raw.sse`);
58024
+ }
57850
58025
  try {
57851
- let json = JSON.parse(text);
57852
- if (prepared.protocol === "responses" && prepared.responsesTextProtocol) {
57853
- const requestBody = JSON.parse(typeof body === "string" ? body : body.toString("utf8"));
57854
- const requestHeaders = buildForwardHeaders(headers);
57855
- json = await compressLoopResponsesJson(
57856
- json,
57857
- { core, config, messages: prepared.originalMessages, session: prepared.session, log: ctx.log, proxyUrl, textProtocol: true },
57858
- requestBody,
57859
- { url: upstreamUrl, headers: requestHeaders }
57860
- );
57861
- }
57862
- const u2 = json.usage ?? {};
57863
- const { total, cached } = usageTotals(prepared.protocol, u2);
57864
- if (typeof total === "number") {
57865
- prepared.session.stats.inputTokens += total;
57866
- prepared.session.stats.lastInputTokens = Math.max(
57867
- 0,
57868
- total - (prepared.session.stats.compressCreditTokens ?? 0)
57869
- );
57870
- if (typeof cached === "number") {
57871
- prepared.session.stats.cachedTokens += cached;
57872
- prepared.session.stats.cacheSamples += 1;
58026
+ const parsedReq = JSON.parse(typeof body === "string" ? body : body.toString("utf8"));
58027
+ const reqHeaders = buildForwardHeaders(headers);
58028
+ const textProtocol = prepared.protocol === "responses" && !!prepared.responsesTextProtocol;
58029
+ const systemPrompt = textProtocol ? buildCompressHybridSystemPrompt(prepared.prompts ?? defaultPrompts) : buildCompressSystemPrompt(prepared.prompts ?? defaultPrompts);
58030
+ const adapter = pickAdapter(prepared.protocol, parsedReq, textProtocol, prepared.responsesProjection, prepared.anthropicSystem, prepared.openaiSystemText);
58031
+ const refreshFolded = (current) => {
58032
+ const turn = core.processTurn({
58033
+ messages: prepared.originalMessages,
58034
+ state: prepared.session.state,
58035
+ config,
58036
+ tokenCount: prepared.session.stats.lastInputTokens,
58037
+ renderTags: prepared.renderTags ?? "text-only"
58038
+ });
58039
+ prepared.session.state = turn.state;
58040
+ const records = current.filter((m2) => typeof m2.id === "string" && m2.id.startsWith("acp_loop_"));
58041
+ return stripKernelSummaries([...turn.messages, ...records], turn.state);
58042
+ };
58043
+ const loop = runCompressLoop(
58044
+ streamToRead,
58045
+ { core, config, messages: prepared.processedMessages.length > 0 ? prepared.processedMessages : prepared.originalMessages, compressMessages: prepared.originalMessages, session: prepared.session, log: ctx.log, proxyUrl, protocol: prepared.protocol, textProtocol, debug: opts.debug, nudge: prepared.nudge, refreshFolded },
58046
+ parsedReq,
58047
+ { url: upstreamUrl, headers: reqHeaders },
58048
+ adapter,
58049
+ systemPrompt,
58050
+ clientAbort.signal
58051
+ );
58052
+ let protocolFragmentWarned = false;
58053
+ for await (const chunk of loop) {
58054
+ if (res.destroyed || res.writableEnded) break;
58055
+ {
58056
+ const s3 = chunk.toString("utf8");
58057
+ if (s3.includes("<acp ") || s3.includes("</acp")) {
58058
+ log2("warn", `[${prepared.session.id}] tag echo: ${prepared.protocol} response stream contains <acp tag`);
58059
+ } else if (!protocolFragmentWarned && containsToolCallXmlFragment(s3)) {
58060
+ protocolFragmentWarned = true;
58061
+ log2("warn", `[${prepared.session.id}] tag echo: ${prepared.protocol} response stream contains tool-call XML fragment (possible tag echo; not stripped)`);
58062
+ }
58063
+ }
58064
+ res.write(chunk);
58065
+ if (res.writableNeedDrain) {
58066
+ await Promise.race([
58067
+ new Promise((r) => res.once("drain", () => r())),
58068
+ new Promise((r) => res.once("close", () => r()))
58069
+ ]);
57873
58070
  }
57874
- const out = u2.completion_tokens ?? u2.output_tokens;
57875
- if (typeof out === "number") prepared.session.stats.outputTokens += out;
58071
+ if (res.destroyed || res.writableEnded) break;
57876
58072
  }
57877
- if (prepared.protocol === "openai") {
57878
- rewriteOpenaiJsonResponse(json, ctx);
57879
- } else if (prepared.protocol === "responses") {
57880
- rewriteResponsesJsonResponse(json, ctx);
57881
- } else {
57882
- rewriteJsonResponse(json, ctx);
58073
+ res.end();
58074
+ } catch (e) {
58075
+ emitStreamError(res, prepared.protocol, e?.message ?? String(e), (m2) => log2("error", `[${prepared.session.id}] ${m2}`));
58076
+ } finally {
58077
+ clearUpstreamTimer();
58078
+ if (dumpRaw) await dumpRaw;
58079
+ markDirty(prepared.session);
58080
+ }
58081
+ } else {
58082
+ try {
58083
+ const buf = await upstream.arrayBuffer();
58084
+ const text = Buffer.from(buf).toString("utf8");
58085
+ try {
58086
+ let json = JSON.parse(text);
58087
+ if (prepared.protocol === "responses" && prepared.responsesTextProtocol) {
58088
+ const requestBody = JSON.parse(typeof body === "string" ? body : body.toString("utf8"));
58089
+ const requestHeaders = buildForwardHeaders(headers);
58090
+ json = await compressLoopResponsesJson(
58091
+ json,
58092
+ { core, config, messages: prepared.originalMessages, session: prepared.session, log: ctx.log, proxyUrl, textProtocol: true },
58093
+ requestBody,
58094
+ { url: upstreamUrl, headers: requestHeaders }
58095
+ );
58096
+ }
58097
+ const u2 = json.usage ?? {};
58098
+ const { total, cached } = usageTotals(prepared.protocol, u2);
58099
+ if (typeof total === "number") {
58100
+ prepared.session.stats.inputTokens += total;
58101
+ prepared.session.stats.lastInputTokens = Math.max(
58102
+ 0,
58103
+ total - (prepared.session.stats.compressCreditTokens ?? 0)
58104
+ );
58105
+ if (typeof cached === "number") {
58106
+ prepared.session.stats.cachedTokens += cached;
58107
+ prepared.session.stats.cacheSamples += 1;
58108
+ }
58109
+ const out = u2.completion_tokens ?? u2.output_tokens;
58110
+ if (typeof out === "number") prepared.session.stats.outputTokens += out;
58111
+ }
58112
+ if (prepared.protocol === "openai") {
58113
+ rewriteOpenaiJsonResponse(json, ctx);
58114
+ } else if (prepared.protocol === "responses") {
58115
+ rewriteResponsesJsonResponse(json, ctx);
58116
+ } else {
58117
+ rewriteJsonResponse(json, ctx);
58118
+ }
58119
+ res.end(JSON.stringify(json));
58120
+ } catch {
58121
+ res.end(text);
57883
58122
  }
57884
- res.end(JSON.stringify(json));
57885
- } catch {
57886
- res.end(text);
58123
+ } finally {
58124
+ clearUpstreamTimer();
58125
+ markDirty(prepared.session);
57887
58126
  }
57888
- } finally {
57889
- clearUpstreamTimer();
57890
58127
  }
58128
+ } finally {
58129
+ clearUpstreamTimer();
57891
58130
  }
57892
- markDirty(prepared.session);
57893
58131
  }
57894
58132
  async function readStreamToBuffer(stream2, maxBytes = 1 << 20) {
57895
58133
  const reader = stream2.getReader();
@@ -58034,7 +58272,8 @@ function sendStats(res) {
58034
58272
  outputTokens: s3.stats.outputTokens,
58035
58273
  cacheSamples: s3.stats.cacheSamples,
58036
58274
  cacheHitPct: s3.stats.cacheSamples > 0 && s3.stats.inputTokens > 0 ? Math.round(s3.stats.cachedTokens / s3.stats.inputTokens * 100) : null,
58037
- lastSeen: new Date(s3.lastSeen).toISOString()
58275
+ lastSeen: new Date(s3.lastSeen).toISOString(),
58276
+ restored: s3.restored === true
58038
58277
  }));
58039
58278
  res.writeHead(200, { "content-type": "application/json" });
58040
58279
  res.end(JSON.stringify({ sessions: sessions2 }, null, 2));
@@ -62333,19 +62572,20 @@ function livePidHoldsOverlay(overlay) {
62333
62572
  function linkOverlayEntry(realHome, overlay, entry) {
62334
62573
  const target = path14.join(realHome, entry);
62335
62574
  const link = path14.join(overlay, entry);
62336
- try {
62337
- fs11.symlinkSync(target, link);
62338
- return true;
62339
- } catch (err2) {
62340
- const code = err2.code;
62341
- if (code !== "EPERM" && code !== "EACCES" && code !== "EINVAL") return false;
62342
- }
62343
62575
  let st2;
62344
62576
  try {
62345
62577
  st2 = fs11.lstatSync(target);
62346
62578
  } catch {
62347
62579
  return false;
62348
62580
  }
62581
+ const kind = st2.isDirectory() ? process.platform === "win32" ? "junction" : "dir" : "file";
62582
+ try {
62583
+ fs11.symlinkSync(target, link, kind);
62584
+ return true;
62585
+ } catch (err2) {
62586
+ const code = err2.code;
62587
+ if (code !== "EPERM" && code !== "EACCES" && code !== "EINVAL") return false;
62588
+ }
62349
62589
  if (st2.isDirectory()) {
62350
62590
  try {
62351
62591
  fs11.symlinkSync(target, link, "junction");
@@ -62376,17 +62616,39 @@ function isWriteThroughHardlink(overlayPath, realPath, st2) {
62376
62616
  }
62377
62617
  }
62378
62618
  function sqliteSetMembers(base) {
62379
- return [base, `${base}-wal`, `${base}-shm`];
62619
+ return [base, `${base}-wal`, `${base}-shm`, `${base}-journal`];
62620
+ }
62621
+ function freeConflictName(dst) {
62622
+ let candidate = `${dst}.bili-conflict`;
62623
+ let n = 1;
62624
+ while (n < 1e5) {
62625
+ try {
62626
+ fs11.lstatSync(candidate);
62627
+ candidate = `${dst}.bili-conflict.${n}`;
62628
+ n += 1;
62629
+ } catch {
62630
+ return candidate;
62631
+ }
62632
+ }
62633
+ return candidate;
62380
62634
  }
62381
62635
  function mergeSqliteSet(overlay, realHome, base) {
62382
- const members = sqliteSetMembers(base).filter((m2) => {
62636
+ const members = sqliteSetMembers(base);
62637
+ const statFile = (dir, m2) => {
62383
62638
  try {
62384
- return fs11.lstatSync(path14.join(overlay, m2)).isFile();
62639
+ const st2 = fs11.lstatSync(path14.join(dir, m2));
62640
+ return st2.isFile() ? st2 : void 0;
62385
62641
  } catch {
62386
- return false;
62642
+ return void 0;
62387
62643
  }
62388
- });
62389
- if (members.length === 0) return true;
62644
+ };
62645
+ const oMain = statFile(overlay, base);
62646
+ const rMain = statFile(realHome, base);
62647
+ let winner;
62648
+ if (oMain && rMain) winner = rMain.mtimeMs >= oMain.mtimeMs ? "real" : "overlay";
62649
+ else if (oMain) winner = "overlay";
62650
+ else if (rMain) winner = "real";
62651
+ else winner = "orphan";
62390
62652
  const undo = [];
62391
62653
  const rollback = () => {
62392
62654
  for (const step of undo.reverse()) {
@@ -62397,46 +62659,39 @@ function mergeSqliteSet(overlay, realHome, base) {
62397
62659
  }
62398
62660
  undo.length = 0;
62399
62661
  };
62662
+ const movePreserving = (src, dst) => {
62663
+ let dstStat;
62664
+ try {
62665
+ dstStat = fs11.lstatSync(dst);
62666
+ } catch {
62667
+ }
62668
+ if (dstStat) {
62669
+ if (dstStat.isDirectory()) throw new Error("target is a directory");
62670
+ const conflict = freeConflictName(dst);
62671
+ fs11.renameSync(dst, conflict);
62672
+ undo.push(() => fs11.renameSync(conflict, dst));
62673
+ }
62674
+ fs11.renameSync(src, dst);
62675
+ undo.push(() => fs11.renameSync(dst, src));
62676
+ };
62677
+ const preserveAsConflict = (src, name) => {
62678
+ const conflict = freeConflictName(path14.join(realHome, name));
62679
+ fs11.renameSync(src, conflict);
62680
+ undo.push(() => fs11.renameSync(conflict, src));
62681
+ };
62400
62682
  try {
62401
62683
  for (const m2 of members) {
62402
- const src = path14.join(overlay, m2);
62403
- const dst = path14.join(realHome, m2);
62404
- let dstStat;
62405
- try {
62406
- dstStat = fs11.lstatSync(dst);
62407
- } catch {
62408
- }
62409
- if (dstStat && dstStat.isDirectory()) {
62410
- rollback();
62411
- return false;
62412
- }
62413
- const srcStat = fs11.lstatSync(src);
62414
- if (dstStat && dstStat.mtimeMs >= srcStat.mtimeMs) {
62415
- try {
62416
- fs11.renameSync(src, `${dst}.bili-conflict`);
62417
- } catch {
62418
- rollback();
62419
- return false;
62420
- }
62421
- undo.push(() => fs11.renameSync(`${dst}.bili-conflict`, src));
62422
- continue;
62423
- }
62424
- if (dstStat) {
62425
- try {
62426
- fs11.renameSync(dst, `${dst}.bili-conflict`);
62427
- } catch {
62428
- rollback();
62429
- return false;
62430
- }
62431
- undo.push(() => fs11.renameSync(`${dst}.bili-conflict`, dst));
62432
- }
62433
- try {
62434
- fs11.renameSync(src, dst);
62435
- } catch {
62436
- rollback();
62437
- return false;
62684
+ const o = statFile(overlay, m2);
62685
+ const r = statFile(realHome, m2);
62686
+ if (winner === "overlay") {
62687
+ if (o) movePreserving(path14.join(overlay, m2), path14.join(realHome, m2));
62688
+ else if (r) preserveAsConflict(path14.join(realHome, m2), m2);
62689
+ } else if (winner === "real") {
62690
+ if (o) preserveAsConflict(path14.join(overlay, m2), m2);
62691
+ } else {
62692
+ if (o) preserveAsConflict(path14.join(overlay, m2), m2);
62693
+ else if (r) preserveAsConflict(path14.join(realHome, m2), m2);
62438
62694
  }
62439
- undo.push(() => fs11.renameSync(dst, src));
62440
62695
  }
62441
62696
  return true;
62442
62697
  } catch {
@@ -62445,6 +62700,8 @@ function mergeSqliteSet(overlay, realHome, base) {
62445
62700
  }
62446
62701
  }
62447
62702
  function refreshOverlayHome(realHome, overlay, generatedFile) {
62703
+ const generatedFiles = new Set(Array.isArray(generatedFile) ? generatedFile : [generatedFile]);
62704
+ const isGeneratedDraft = (name) => [...generatedFiles].some((g2) => name.startsWith(`.${g2}.`) && name.endsWith(".tmp"));
62448
62705
  try {
62449
62706
  fs11.mkdirSync(overlay, { recursive: true });
62450
62707
  } catch {
@@ -62475,7 +62732,7 @@ function refreshOverlayHome(realHome, overlay, generatedFile) {
62475
62732
  }
62476
62733
  const dbSets = [];
62477
62734
  for (const entry of overlayEntries) {
62478
- if (!entry.endsWith(".db") || entry === generatedFile) continue;
62735
+ if (!entry.endsWith(".db") || generatedFiles.has(entry)) continue;
62479
62736
  const members = sqliteSetMembers(entry);
62480
62737
  if (!members.some((m2) => m2 !== entry && overlayEntries.includes(m2))) continue;
62481
62738
  let mainSt;
@@ -62493,9 +62750,9 @@ function refreshOverlayHome(realHome, overlay, generatedFile) {
62493
62750
  }
62494
62751
  }
62495
62752
  for (const entry of overlayEntries) {
62496
- if (entry === generatedFile) continue;
62753
+ if (generatedFiles.has(entry)) continue;
62497
62754
  const overlayPath = path14.join(overlay, entry);
62498
- if (entry.startsWith(`.${generatedFile}.`) && entry.endsWith(".tmp")) {
62755
+ if (isGeneratedDraft(entry)) {
62499
62756
  try {
62500
62757
  fs11.unlinkSync(overlayPath);
62501
62758
  } catch {
@@ -62529,7 +62786,7 @@ function refreshOverlayHome(realHome, overlay, generatedFile) {
62529
62786
  fs11.unlinkSync(overlayPath);
62530
62787
  } catch {
62531
62788
  }
62532
- } else if (mergeOverlayEntry(overlayPath, realPath, generatedFile)) {
62789
+ } else if (mergeOverlayEntry(overlayPath, realPath, generatedFiles)) {
62533
62790
  try {
62534
62791
  fs11.rmSync(overlayPath, { recursive: true, force: true });
62535
62792
  } catch {
@@ -62551,7 +62808,7 @@ function refreshOverlayHome(realHome, overlay, generatedFile) {
62551
62808
  let total = 0;
62552
62809
  const linkFailures = [];
62553
62810
  for (const entry of realEntries) {
62554
- if (entry === generatedFile) continue;
62811
+ if (generatedFiles.has(entry)) continue;
62555
62812
  total += 1;
62556
62813
  const overlayPath = path14.join(overlay, entry);
62557
62814
  let present = false;
@@ -62585,7 +62842,7 @@ function refreshOverlayHome(realHome, overlay, generatedFile) {
62585
62842
  }
62586
62843
  return true;
62587
62844
  }
62588
- function mergeOverlayEntry(src, dst, excludedName) {
62845
+ function mergeOverlayEntry(src, dst, excludedNames) {
62589
62846
  let st2;
62590
62847
  try {
62591
62848
  st2 = fs11.lstatSync(src);
@@ -62600,7 +62857,7 @@ function mergeOverlayEntry(src, dst, excludedName) {
62600
62857
  if (st2.isDirectory()) {
62601
62858
  if (dstStat && !dstStat.isDirectory()) {
62602
62859
  try {
62603
- fs11.renameSync(dst, `${dst}.bili-conflict`);
62860
+ fs11.renameSync(dst, freeConflictName(dst));
62604
62861
  dstStat = void 0;
62605
62862
  } catch {
62606
62863
  return false;
@@ -62619,14 +62876,14 @@ function mergeOverlayEntry(src, dst, excludedName) {
62619
62876
  }
62620
62877
  let ok = true;
62621
62878
  for (const entry of entries) {
62622
- if (excludedName !== void 0 && entry === excludedName) continue;
62623
- if (!mergeOverlayEntry(path14.join(src, entry), path14.join(dst, entry), excludedName)) ok = false;
62879
+ if (excludedNames?.has(entry)) continue;
62880
+ if (!mergeOverlayEntry(path14.join(src, entry), path14.join(dst, entry), excludedNames)) ok = false;
62624
62881
  }
62625
62882
  return ok;
62626
62883
  }
62627
62884
  if (dstStat && dstStat.isDirectory()) {
62628
62885
  try {
62629
- fs11.renameSync(src, `${dst}.bili-conflict`);
62886
+ fs11.renameSync(src, freeConflictName(dst));
62630
62887
  return true;
62631
62888
  } catch {
62632
62889
  return false;
@@ -62649,7 +62906,7 @@ function mergeOverlayEntry(src, dst, excludedName) {
62649
62906
  }
62650
62907
  if (dstStat && dstStat.mtimeMs >= st2.mtimeMs) {
62651
62908
  try {
62652
- fs11.renameSync(src, `${dst}.bili-conflict`);
62909
+ fs11.renameSync(src, freeConflictName(dst));
62653
62910
  return true;
62654
62911
  } catch {
62655
62912
  return false;
@@ -62657,7 +62914,7 @@ function mergeOverlayEntry(src, dst, excludedName) {
62657
62914
  }
62658
62915
  if (dstStat) {
62659
62916
  try {
62660
- fs11.renameSync(dst, `${dst}.bili-conflict`);
62917
+ fs11.renameSync(dst, freeConflictName(dst));
62661
62918
  } catch {
62662
62919
  return false;
62663
62920
  }
@@ -62692,45 +62949,58 @@ function writeOverlayFileAtomic(overlay, fileName, contents) {
62692
62949
  }
62693
62950
  }
62694
62951
  }
62695
- function preparePiHttpRewrite(piHome, origin, httpRewrites, httpsRewrites) {
62696
- if (httpRewrites.length === 0 && httpsRewrites.length === 0) return void 0;
62697
- const modelsPath = path14.join(piHome, "models.json");
62698
- let txt;
62952
+ function writePiCompactionDisabledSettings(piHome, overlay) {
62953
+ let settings = {};
62699
62954
  try {
62700
- txt = fs11.readFileSync(modelsPath, "utf8");
62955
+ const parsed = JSON.parse(fs11.readFileSync(path14.join(piHome, "settings.json"), "utf8"));
62956
+ if (parsed && typeof parsed === "object" && !Array.isArray(parsed)) {
62957
+ settings = parsed;
62958
+ }
62701
62959
  } catch {
62702
- return void 0;
62703
62960
  }
62704
- let parsed;
62961
+ const compactionRaw = settings.compaction;
62962
+ const compaction = compactionRaw && typeof compactionRaw === "object" && !Array.isArray(compactionRaw) ? { ...compactionRaw } : {};
62963
+ compaction.enabled = false;
62964
+ settings.compaction = compaction;
62965
+ writeOverlayFileAtomic(overlay, "settings.json", JSON.stringify(settings, null, 2));
62966
+ }
62967
+ function preparePiHttpRewrite(piHome, origin, httpRewrites, httpsRewrites) {
62968
+ if (!fs11.existsSync(piHome)) return void 0;
62969
+ const overlay = `${piHome}-bili`;
62970
+ const generated = ["settings.json"];
62971
+ let modelsRoot;
62705
62972
  try {
62706
- parsed = JSON.parse(txt);
62973
+ const parsed = JSON.parse(fs11.readFileSync(path14.join(piHome, "models.json"), "utf8"));
62974
+ if (parsed && typeof parsed === "object" && !Array.isArray(parsed)) {
62975
+ modelsRoot = parsed;
62976
+ generated.push("models.json");
62977
+ }
62707
62978
  } catch {
62708
- return void 0;
62709
62979
  }
62710
- if (!parsed || typeof parsed !== "object" || Array.isArray(parsed)) return void 0;
62711
- const root = parsed;
62712
- const providersVal = root.providers;
62713
- if (providersVal && typeof providersVal === "object" && !Array.isArray(providersVal)) {
62714
- const providers = providersVal;
62715
- for (const r of httpRewrites) {
62716
- const prov = providers[r.key];
62717
- if (prov && typeof prov === "object" && !Array.isArray(prov)) {
62718
- const p2 = prov;
62719
- const existing = typeof p2.baseUrl === "string" ? p2.baseUrl : r.realUpstream;
62720
- p2.baseUrl = wrapUpstream(origin, unwrapUpstream2(existing));
62980
+ if (!refreshOverlayHome(piHome, overlay, generated)) return void 0;
62981
+ if (modelsRoot) {
62982
+ const providersVal = modelsRoot.providers;
62983
+ if (providersVal && typeof providersVal === "object" && !Array.isArray(providersVal)) {
62984
+ const providers = providersVal;
62985
+ for (const r of httpRewrites) {
62986
+ const prov = providers[r.key];
62987
+ if (prov && typeof prov === "object" && !Array.isArray(prov)) {
62988
+ const p2 = prov;
62989
+ const existing = typeof p2.baseUrl === "string" ? p2.baseUrl : r.realUpstream;
62990
+ p2.baseUrl = wrapUpstream(origin, unwrapUpstream2(existing));
62991
+ }
62721
62992
  }
62722
- }
62723
- for (const r of httpsRewrites) {
62724
- const prov = providers[r.key];
62725
- if (prov && typeof prov === "object" && !Array.isArray(prov)) {
62726
- const p2 = prov;
62727
- p2.baseUrl = r.realUpstream;
62993
+ for (const r of httpsRewrites) {
62994
+ const prov = providers[r.key];
62995
+ if (prov && typeof prov === "object" && !Array.isArray(prov)) {
62996
+ const p2 = prov;
62997
+ p2.baseUrl = r.realUpstream;
62998
+ }
62728
62999
  }
62729
63000
  }
63001
+ writeOverlayFileAtomic(overlay, "models.json", JSON.stringify(modelsRoot));
62730
63002
  }
62731
- const overlay = `${piHome}-bili`;
62732
- if (!refreshOverlayHome(piHome, overlay, "models.json")) return void 0;
62733
- writeOverlayFileAtomic(overlay, "models.json", JSON.stringify(root));
63003
+ writePiCompactionDisabledSettings(piHome, overlay);
62734
63004
  return overlay;
62735
63005
  }
62736
63006
  function atomicWriteTextFile(filePath, contents) {
@@ -62781,55 +63051,145 @@ function unpackDeadProxyUrlsInFile(filePath, livePorts) {
62781
63051
  }
62782
63052
  return changed;
62783
63053
  }
63054
+ function mergeOmpCompactionDisabledYaml(text) {
63055
+ const lines = text.split(/\r?\n/);
63056
+ let compactionLine = -1;
63057
+ for (let i = 0; i < lines.length; i++) {
63058
+ const rawLine = lines[i];
63059
+ const trimmed = rawLine.trim();
63060
+ if (!trimmed || trimmed.startsWith("#")) continue;
63061
+ const indent = rawLine.length - rawLine.trimStart().length;
63062
+ if (indent === 0 && /^compaction:\s*(#.*)?$/.test(trimmed)) {
63063
+ compactionLine = i;
63064
+ break;
63065
+ }
63066
+ }
63067
+ if (compactionLine === -1) {
63068
+ const base = text === "" || text.endsWith("\n") ? text : `${text}
63069
+ `;
63070
+ return `${base}compaction:
63071
+ enabled: false
63072
+ `;
63073
+ }
63074
+ for (let i = compactionLine + 1; i < lines.length; i++) {
63075
+ const rawLine = lines[i];
63076
+ const trimmed = rawLine.trim();
63077
+ if (!trimmed || trimmed.startsWith("#")) continue;
63078
+ const indent = rawLine.length - rawLine.trimStart().length;
63079
+ if (indent === 0) break;
63080
+ if (/^enabled:/.test(trimmed)) {
63081
+ const leading = rawLine.slice(0, rawLine.length - rawLine.trimStart().length);
63082
+ const commentMatch = /\s+#.*$/.exec(rawLine);
63083
+ const comment = commentMatch ? commentMatch[0] : "";
63084
+ lines[i] = `${leading}enabled: false${comment}`;
63085
+ return lines.join("\n");
63086
+ }
63087
+ }
63088
+ let childIndent = 2;
63089
+ for (let i = compactionLine + 1; i < lines.length; i++) {
63090
+ const rawLine = lines[i];
63091
+ const trimmed = rawLine.trim();
63092
+ if (!trimmed || trimmed.startsWith("#")) continue;
63093
+ const indent = rawLine.length - rawLine.trimStart().length;
63094
+ if (indent === 0) break;
63095
+ childIndent = indent;
63096
+ break;
63097
+ }
63098
+ lines.splice(compactionLine + 1, 0, `${" ".repeat(childIndent)}enabled: false`);
63099
+ return lines.join("\n");
63100
+ }
63101
+ function writeOmpCompactionDisabledConfig(ompHome, overlay) {
63102
+ let base;
63103
+ let baseIsJson = false;
63104
+ for (const name of ["config.yml", "config.yaml"]) {
63105
+ try {
63106
+ base = fs11.readFileSync(path14.join(ompHome, name), "utf8");
63107
+ break;
63108
+ } catch {
63109
+ }
63110
+ }
63111
+ if (base === void 0) {
63112
+ try {
63113
+ base = fs11.readFileSync(path14.join(ompHome, "settings.json"), "utf8");
63114
+ baseIsJson = true;
63115
+ } catch {
63116
+ }
63117
+ }
63118
+ let merged;
63119
+ if (base === void 0) {
63120
+ merged = "compaction:\n enabled: false\n";
63121
+ } else if (baseIsJson) {
63122
+ let settings = {};
63123
+ try {
63124
+ const parsed = JSON.parse(base);
63125
+ if (parsed && typeof parsed === "object" && !Array.isArray(parsed)) {
63126
+ settings = parsed;
63127
+ }
63128
+ } catch {
63129
+ }
63130
+ const compactionRaw = settings.compaction;
63131
+ const compaction = compactionRaw && typeof compactionRaw === "object" && !Array.isArray(compactionRaw) ? { ...compactionRaw } : {};
63132
+ compaction.enabled = false;
63133
+ settings.compaction = compaction;
63134
+ merged = JSON.stringify(settings, null, 2);
63135
+ } else {
63136
+ merged = mergeOmpCompactionDisabledYaml(base);
63137
+ }
63138
+ writeOverlayFileAtomic(overlay, "config.yml", merged);
63139
+ }
62784
63140
  function prepareOmpHttpRewrite(ompHome, origin, httpRewrites, httpsRewrites) {
63141
+ if (!fs11.existsSync(ompHome)) return void 0;
63142
+ const overlay = `${ompHome}-bili`;
63143
+ const generated = ["config.yml"];
62785
63144
  const modelsPath = path14.join(ompHome, "models.yml");
62786
63145
  const unpacked = unpackDeadProxyUrlsInFile(modelsPath, liveProxyPorts());
62787
63146
  if (unpacked > 0) {
62788
63147
  console.error(`bili: unpacked ${unpacked} dead proxy URL(s) from the real ${modelsPath}`);
62789
63148
  }
62790
- if (httpRewrites.length === 0 && httpsRewrites.length === 0) return void 0;
62791
- let txt;
63149
+ let modelsText;
62792
63150
  try {
62793
- txt = fs11.readFileSync(modelsPath, "utf8");
63151
+ modelsText = fs11.readFileSync(modelsPath, "utf8");
63152
+ generated.push("models.yml");
62794
63153
  } catch {
62795
- return void 0;
62796
63154
  }
62797
- const httpMap = new Map(httpRewrites.map((r) => [r.key, wrapUpstream(origin, r.realUpstream)]));
62798
- const httpsMap = new Map(httpsRewrites.map((r) => [r.key, r.realUpstream]));
62799
- const lines = txt.split(/\r?\n/);
62800
- let providersIndent = -1;
62801
- let providerIndent = -1;
62802
- let currentProvider = null;
62803
- for (let i = 0; i < lines.length; i++) {
62804
- const rawLine = lines[i];
62805
- const trimmed = rawLine.trim();
62806
- if (!trimmed || trimmed.startsWith("#")) continue;
62807
- const indent = rawLine.length - rawLine.trimStart().length;
62808
- if (providersIndent === -1) {
62809
- if (/^providers:\s*(#.*)?$/.test(trimmed)) providersIndent = indent;
62810
- continue;
62811
- }
62812
- if (indent <= providersIndent) break;
62813
- if (providerIndent === -1) providerIndent = indent;
62814
- if (indent === providerIndent) {
62815
- const m2 = /^([A-Za-z0-9_.-]+):/.exec(trimmed);
62816
- currentProvider = m2 ? m2[1] : null;
62817
- } else if (indent > providerIndent && currentProvider) {
62818
- const baseMatch = /^(baseUrl:\s*)(\S+)/.exec(trimmed);
62819
- if (baseMatch) {
62820
- const target = httpMap.get(currentProvider) ?? httpsMap.get(currentProvider);
62821
- if (target) {
62822
- const leading = rawLine.slice(0, rawLine.length - rawLine.trimStart().length);
62823
- const commentMatch = /\s+#.*$/.exec(rawLine);
62824
- const comment = commentMatch ? commentMatch[0] : "";
62825
- lines[i] = `${leading}baseUrl: ${target}${comment}`;
63155
+ if (!refreshOverlayHome(ompHome, overlay, generated)) return void 0;
63156
+ if (modelsText !== void 0) {
63157
+ const httpMap = new Map(httpRewrites.map((r) => [r.key, wrapUpstream(origin, r.realUpstream)]));
63158
+ const httpsMap = new Map(httpsRewrites.map((r) => [r.key, r.realUpstream]));
63159
+ const lines = modelsText.split(/\r?\n/);
63160
+ let providersIndent = -1;
63161
+ let providerIndent = -1;
63162
+ let currentProvider = null;
63163
+ for (let i = 0; i < lines.length; i++) {
63164
+ const rawLine = lines[i];
63165
+ const trimmed = rawLine.trim();
63166
+ if (!trimmed || trimmed.startsWith("#")) continue;
63167
+ const indent = rawLine.length - rawLine.trimStart().length;
63168
+ if (providersIndent === -1) {
63169
+ if (/^providers:\s*(#.*)?$/.test(trimmed)) providersIndent = indent;
63170
+ continue;
63171
+ }
63172
+ if (indent <= providersIndent) break;
63173
+ if (providerIndent === -1) providerIndent = indent;
63174
+ if (indent === providerIndent) {
63175
+ const m2 = /^([A-Za-z0-9_.-]+):/.exec(trimmed);
63176
+ currentProvider = m2 ? m2[1] : null;
63177
+ } else if (indent > providerIndent && currentProvider) {
63178
+ const baseMatch = /^(baseUrl:\s*)(\S+)/.exec(trimmed);
63179
+ if (baseMatch) {
63180
+ const target = httpMap.get(currentProvider) ?? httpsMap.get(currentProvider);
63181
+ if (target) {
63182
+ const leading = rawLine.slice(0, rawLine.length - rawLine.trimStart().length);
63183
+ const commentMatch = /\s+#.*$/.exec(rawLine);
63184
+ const comment = commentMatch ? commentMatch[0] : "";
63185
+ lines[i] = `${leading}baseUrl: ${target}${comment}`;
63186
+ }
62826
63187
  }
62827
63188
  }
62828
63189
  }
63190
+ writeOverlayFileAtomic(overlay, "models.yml", lines.join("\n"));
62829
63191
  }
62830
- const overlay = `${ompHome}-bili`;
62831
- if (!refreshOverlayHome(ompHome, overlay, "models.yml")) return void 0;
62832
- writeOverlayFileAtomic(overlay, "models.yml", lines.join("\n"));
63192
+ writeOmpCompactionDisabledConfig(ompHome, overlay);
62833
63193
  return overlay;
62834
63194
  }
62835
63195
  function prepareHermesHome(hermesHome, origin, rewrites) {