billion-context 0.1.84 → 0.1.85

package/dist/index.js

@@ -44164,6 +44164,9 @@ function validateConfig(config) {
       "nudge.maxContextLimitPct must not exceed nudge.emergencyThresholdPct"
     );
   }
+  if (config.nudge.minPressureBenefitTokens !== void 0 && (!Number.isFinite(config.nudge.minPressureBenefitTokens) || config.nudge.minPressureBenefitTokens < 0)) {
+    errors.push("nudge.minPressureBenefitTokens must be finite and >= 0");
+  }
   if (config.promotionThreshold < 1) {
     errors.push("promotionThreshold must be >= 1");
   }
@@ -46021,6 +46024,9 @@ function resolveAdaptiveGrowth(modelContextLimit, nudge) {
     )
   );
 }
+function resolveMinPressureBenefit(modelContextLimit, nudge) {
+  return nudge.minPressureBenefitTokens ?? Math.max(5e3, Math.round(modelContextLimit * 0.01));
+}
 function pendingByTier(state, recommendation, countTokens, minCompressRange) {
   const out = {};
   const merged = recommendation?.recommendedRanges ?? [];
@@ -46047,6 +46053,7 @@ function decideNudge(input) {
   const limit = config.modelContextLimit;
   const usage = limit > 0 ? tokenCount / limit : 0;
   const nudgeGrowthTokens = resolveAdaptiveGrowth(limit, config.nudge);
+  const minPressureBenefit = resolveMinPressureBenefit(limit, config.nudge);
   const overLimit = usage >= config.nudge.maxContextLimitPct;
   const emergencyOverride = usage >= config.nudge.emergencyThresholdPct;
   const pressure = overLimit || emergencyOverride;
@@ -46072,6 +46079,7 @@ function decideNudge(input) {
   );
   let injectedTier = null;
   let injectedReason = "";
+  let bestPending = 0;
   const t1Eff = tiers[1]?.pending ?? 0;
   const t2Pen = tiers[2]?.pending ?? 0;
   const t3Pen = tiers[3]?.pending ?? 0;
@@ -46085,7 +46093,6 @@ function decideNudge(input) {
       candidates.push(2, 3);
     }
     let best = null;
-    let bestPending = 0;
     for (const t of candidates) {
       const p2 = tiers[t]?.pending ?? 0;
       if (p2 > bestPending) {
@@ -46093,7 +46100,7 @@ function decideNudge(input) {
         best = t;
       }
     }
-    if (best !== null && bestPending > 0) {
+    if (best !== null && bestPending >= minPressureBenefit) {
       injectedTier = best;
       const label = emergencyOverride ? "EMERGENCY" : "OVER-LIMIT";
       injectedReason = best === 1 ? `${label} T1: max effective pending ${bestPending}, usage ${Math.round(usage * 100)}%` : `${label} T${best} distill: max pending ${bestPending} (T1 effective ${t1Eff}, T2 ${t2Pen}, T3 ${t3Pen}), usage ${Math.round(usage * 100)}%`;
@@ -46127,7 +46134,7 @@ function decideNudge(input) {
     reason = injectedReason;
   } else if (pressure) {
     const label = emergencyOverride ? "EMERGENCY" : "OVER-LIMIT";
-    reason = `${label}: usage ${Math.round(usage * 100)}% but no tier has effective compressible content (T1 effective ${t1Eff}, T2 ${t2Pen}, T3 ${t3Pen}) \u2014 nudge suppressed to avoid offering ranges below minCompressRange`;
+    reason = bestPending === 0 ? `${label}: usage ${Math.round(usage * 100)}% but no tier has effective compressible content (T1 effective ${t1Eff}, T2 ${t2Pen}, T3 ${t3Pen}) \u2014 nudge suppressed to avoid offering ranges below minCompressRange` : `${label}: usage ${Math.round(usage * 100)}% but max pending ${bestPending} < min benefit ${minPressureBenefit} tokens (T1 effective ${t1Eff}, T2 ${t2Pen}, T3 ${t3Pen}) \u2014 suppressed: rewriting below the benefit floor reclaims almost nothing while usage stays high; truncate.threshold remains the safety valve`;
   } else {
     const tiersList = [1, 2, 3];
     const eligible = tiersList.filter((t) => config.tiers.enabled || t === 1);
@@ -46183,6 +46190,7 @@ function decideNudge(input) {
       hasPendingNudge: hasPendingNudge ? 1 : 0,
       overLimit: overLimit ? 1 : 0,
       emergencyOverride: emergencyOverride ? 1 : 0,
+      minPressureBenefit,
       pendingT1: tiers[1].pending,
       pendingT2: tiers[2].pending,
       pendingT3: tiers[3].pending
@@ -51157,7 +51165,7 @@ function generateRootCA() {
   cert.setSubject(attrs);
   cert.setIssuer(attrs);
   cert.setExtensions([
-    { name: "basicConstraints", cA: true },
+    { name: "basicConstraints", cA: true, critical: true },
     { name: "keyUsage", keyCertSign: true, cRLSign: true, digitalSignature: true },
     { name: "subjectKeyIdentifier" }
   ]);
@@ -51193,16 +51201,10 @@ function ensureRootCA() {
   rootKey = import_node_forge.default.pki.privateKeyFromPem(key);
   writeCombinedBundle();
 }
-function getSecureContext(host) {
+function mintHostCert(host) {
   if (!rootCertPem || !rootKeyPem || !rootCert || !rootKey) {
     throw new Error("CA not initialized \u2014 call ensureRootCA() first");
   }
-  const cached = secureContextCache.get(host);
-  if (cached) {
-    secureContextCache.delete(host);
-    secureContextCache.set(host, cached);
-    return cached;
-  }
   const keys = import_node_forge.default.pki.rsa.generateKeyPair({ bits: 2048 });
   const cert = import_node_forge.default.pki.createCertificate();
   cert.publicKey = keys.publicKey;
@@ -51212,16 +51214,35 @@ function getSecureContext(host) {
   cert.validity.notAfter.setFullYear(cert.validity.notBefore.getFullYear() + 2);
   cert.setSubject([{ name: "commonName", value: host }]);
   cert.setIssuer(rootCert.subject.attributes);
+  const rootSki = rootCert.generateSubjectKeyIdentifier().getBytes();
   cert.setExtensions([
     { name: "basicConstraints", cA: false },
     { name: "keyUsage", digitalSignature: true, keyEncipherment: true },
     { name: "extKeyUsage", serverAuth: true },
-    { name: "subjectAltName", altNames: [{ type: 2, value: host }] }
+    { name: "subjectAltName", altNames: [{ type: 2, value: host }] },
+    { name: "subjectKeyIdentifier" },
+    { name: "authorityKeyIdentifier", keyIdentifier: rootSki }
   ]);
   cert.sign(rootKey, import_node_forge.default.md.sha256.create());
+  return {
+    certPem: import_node_forge.default.pki.certificateToPem(cert),
+    keyPem: import_node_forge.default.pki.privateKeyToPem(keys.privateKey)
+  };
+}
+function getSecureContext(host) {
+  if (!rootCertPem || !rootKeyPem || !rootCert || !rootKey) {
+    throw new Error("CA not initialized \u2014 call ensureRootCA() first");
+  }
+  const cached = secureContextCache.get(host);
+  if (cached) {
+    secureContextCache.delete(host);
+    secureContextCache.set(host, cached);
+    return cached;
+  }
+  const { certPem, keyPem } = mintHostCert(host);
   const ctx = tls2.createSecureContext({
-    cert: import_node_forge.default.pki.certificateToPem(cert),
-    key: import_node_forge.default.pki.privateKeyToPem(keys.privateKey),
+    cert: certPem,
+    key: keyPem,
     ca: rootCertPem
   });
   secureContextCache.set(host, ctx);
@@ -52001,10 +52022,11 @@ async function* runCompressLoop(upstream, ctx, requestBody, requestOptions, adap
         if (realCalls > 0) ctx.log(`[acp-loop] round ${round}: ${realCalls} real tool call(s) forwarded to client`);
       }
       if (proxyResults.length > 0) {
+        const requiresThinkingSignature = ctx.protocol === "anthropic";
         if (reasoningSegments.length > 0) {
           for (let i = 0; i < reasoningSegments.length; i++) {
             const seg = reasoningSegments[i];
-            if (seg.text.length === 0 || seg.signature.length === 0) continue;
+            if (seg.text.length === 0 || requiresThinkingSignature && seg.signature.length === 0) continue;
             const reasoningMsg = {
               id: i === 0 ? `acp_loop_r${round}_reasoning` : `acp_loop_r${round}_reasoning_${i + 1}`,
               role: "assistant",
@@ -52127,7 +52149,7 @@ async function* runCompressLoop(upstream, ctx, requestBody, requestOptions, adap
         try {
           respResult = await fetchUpstream(newBody);
         } catch (e) {
-          if (!(e instanceof UpstreamHttpError) || e.status < 400 || e.status >= 500 || degradedRetried || !coreMessages.some(isLoopThinking)) {
+          if (!(e instanceof UpstreamHttpError) || e.status < 400 || e.status >= 500 || degradedRetried || ctx.protocol !== "anthropic" || !coreMessages.some(isLoopThinking)) {
             throw e;
           }
           degradedRetried = true;
@@ -52349,6 +52371,7 @@ data: ${JSON.stringify(obj)}
 }
 async function* iterSseEvents(stream2) {
   const reader = stream2.getReader();
+  const decoder = new TextDecoder();
   let buf = "";
   try {
     while (true) {
@@ -52360,7 +52383,7 @@ async function* iterSseEvents(stream2) {
         break;
       }
       if (done) break;
-      buf += new TextDecoder().decode(value, { stream: true });
+      buf += decoder.decode(value, { stream: true });
       buf = buf.replace(/\r\n|\r/g, "\n");
       let idx;
       while ((idx = buf.indexOf("\n\n")) >= 0) {
@@ -52838,6 +52861,7 @@ function filterRealToolFragments(parsed, realIndexes) {
 }
 async function* iterSseChunks(stream2) {
   const reader = stream2.getReader();
+  const decoder = new TextDecoder();
   let buf = "";
   try {
     while (true) {
@@ -52849,7 +52873,7 @@ async function* iterSseChunks(stream2) {
         break;
       }
       if (done) break;
-      buf += new TextDecoder().decode(value, { stream: true });
+      buf += decoder.decode(value, { stream: true });
       buf = buf.replace(/\r\n|\r/g, "\n");
       let idx;
       while ((idx = buf.indexOf("\n\n")) >= 0) {
@@ -53175,6 +53199,7 @@ function createOpenaiAdapter(requestBody, clientSystem) {
 // src/loop/adapter-anthropic.ts
 async function* iterSseEvents2(stream2) {
   const reader = stream2.getReader();
+  const decoder = new TextDecoder();
   let buf = "";
   try {
     while (true) {
@@ -53186,7 +53211,7 @@ async function* iterSseEvents2(stream2) {
         break;
       }
       if (done) break;
-      buf += new TextDecoder().decode(value, { stream: true });
+      buf += decoder.decode(value, { stream: true });
       buf = buf.replace(/\r\n|\r/g, "\n");
       let idx;
       while ((idx = buf.indexOf("\n\n")) >= 0) {
@@ -53833,6 +53858,76 @@ ${FAKE_COMPLETION_HINT}`;
   return JSON.stringify(obj);
 }
 
+// src/acp-panel.ts
+var PANEL_BOX_TOP = "\u256D";
+var PANEL_BOX_TITLE = "ACP Context Analysis";
+var PANEL_BOX_FOOTER = "Tag visibility: tags injected to LLM only (deep copy), not persisted in session, not shown in terminal.";
+var PANEL_FALLBACK_HEADER = "\u{1F4CA} ACP status";
+function isAcpPanelText(text) {
+  const t = text.trim();
+  if (t.length === 0) return false;
+  return isBoxPanel(t) || isFallbackPanel(t);
+}
+function isBoxPanel(t) {
+  return t.startsWith(PANEL_BOX_TOP) && t.includes(PANEL_BOX_TITLE) && t.endsWith(PANEL_BOX_FOOTER);
+}
+function isFallbackPanel(t) {
+  const lines = t.split("\n");
+  if (lines[0] !== PANEL_FALLBACK_HEADER) return false;
+  for (let i = 1; i < lines.length; i++) {
+    const line = lines[i];
+    if (line.length > 0 && !line.startsWith("  ")) return false;
+  }
+  return true;
+}
+function messageText(content) {
+  if (typeof content === "string") return content;
+  if (Array.isArray(content)) {
+    let joined = "";
+    for (const part of content) {
+      const p2 = part;
+      if (p2 === null || typeof p2 !== "object") return void 0;
+      const pt2 = p2.type;
+      if (pt2 !== void 0 && pt2 !== "text" && pt2 !== "input_text") return void 0;
+      if (typeof p2.text === "string") joined += p2.text;
+    }
+    return joined;
+  }
+  return void 0;
+}
+function stripAcpPanelMessages(messages) {
+  if (!Array.isArray(messages)) return 0;
+  let stripped = 0;
+  for (let i = messages.length - 1; i >= 0; i--) {
+    const rec = messages[i];
+    if (rec === null || typeof rec !== "object") continue;
+    if (rec.role !== "user") continue;
+    const text = messageText(rec.content);
+    if (text !== void 0 && isAcpPanelText(text)) {
+      messages.splice(i, 1);
+      stripped++;
+    }
+  }
+  return stripped;
+}
+function stripAcpPanelResponsesInput(input) {
+  if (!Array.isArray(input)) return 0;
+  let stripped = 0;
+  for (let i = input.length - 1; i >= 0; i--) {
+    const rec = input[i];
+    if (rec === null || typeof rec !== "object") continue;
+    const type = rec.type;
+    if (type !== "message" && type !== void 0) continue;
+    if (rec.role !== "user") continue;
+    const text = messageText(rec.content);
+    if (text !== void 0 && isAcpPanelText(text)) {
+      input.splice(i, 1);
+      stripped++;
+    }
+  }
+  return stripped;
+}
+
 // src/stream-openai.ts
 function rewriteOpenaiJsonResponse(body, ctx) {
   if (!body || typeof body !== "object") return body;
@@ -55559,6 +55654,7 @@ function readOmpConfig(ompHome) {
 }
 function parseHermesYaml(text) {
   const result = { providers: {} };
+  const candidates = /* @__PURE__ */ new Map();
   let mode = "none";
   let sectionIndent = -1;
   let entryIndent = -1;
@@ -55603,8 +55699,8 @@ function parseHermesYaml(text) {
         current = m2 ? m2[1] : null;
         if (current && !result.providers[current]) result.providers[current] = {};
       } else if (indent > entryIndent && current) {
-        const apiMatch = /^api:\s*(\S+)/.exec(trimmed);
-        if (apiMatch) result.providers[current].api = apiMatch[1];
+        const apiMatch = /^(base_url|url|api):\s*(\S+)/.exec(trimmed);
+        if (apiMatch) candidates.set(`${current}\0${apiMatch[1]}`, apiMatch[2]);
       }
     } else {
       const dashMatch = /^-\s+(.*)$/.exec(trimmed);
@@ -55621,6 +55717,10 @@ function parseHermesYaml(text) {
       }
     }
   }
+  for (const name of Object.keys(result.providers)) {
+    const pick = candidates.get(`${name}\0base_url`) ?? candidates.get(`${name}\0url`) ?? candidates.get(`${name}\0api`);
+    if (pick !== void 0) result.providers[name].api = pick;
+  }
   return result;
 }
 function readHermesConfig(hermesHome) {
@@ -56921,6 +57021,16 @@ function resolveUpstream(_opts, reqUrl, req) {
       }
     }
   }
+  if (reqUrl.startsWith("http://") || reqUrl.startsWith("https://")) {
+    try {
+      const u2 = new URL(reqUrl);
+      const ownHost = (req?.headers.host ?? "").toLowerCase();
+      if (u2.host.toLowerCase() !== ownHost) {
+        return { upstream: `${u2.protocol}//${u2.host}`, rewrittenUrl: reqUrl, tunnel: true };
+      }
+    } catch {
+    }
+  }
   return void 0;
 }
 async function startServer(opts) {
@@ -57658,6 +57768,10 @@ function prepareAnthropic(parsed, req, opts, core, config, prompts, log2, sessio
   let rebuiltMessages = parsed.messages;
   let systemOut = parsed.system;
   let toolsOut = parsed.tools;
+  const strippedPanels = stripAcpPanelMessages(parsed.messages);
+  if (strippedPanels > 0) {
+    log2("info", `[${sessionId}] stripped ${strippedPanels} ACP panel message(s) before projection (UI-only, issue #359)`);
+  }
   try {
     const { msgs, cacheControls } = anthropicToCore(parsed);
     originalMessages = msgs;
@@ -57747,6 +57861,10 @@ function prepareOpenai(parsed, req, opts, core, config, prompts, log2, session,
   const isTitleGen = maxTokens <= 200;
   const shouldInject = opts.compress.injectTool && !isTitleGen;
   const injectTools = shouldInject && !pluginMode;
+  const strippedPanels = stripAcpPanelMessages(parsed.messages);
+  if (strippedPanels > 0) {
+    log2("info", `[${sessionId}] stripped ${strippedPanels} ACP panel message(s) before projection (UI-only, issue #359)`);
+  }
   try {
     const { msgs, systemText } = openaiToCore(parsed);
     openaiSystemText = systemText;
@@ -57832,6 +57950,10 @@ function prepareResponses(parsed, req, opts, core, config, prompts, log2, sessio
   if (droppedEmpty > 0) {
     log2("info", `[${sessionId}] dropped ${droppedEmpty} whitespace-only message item(s) before projection (flattened-turn artifact)`);
   }
+  const strippedPanels = stripAcpPanelResponsesInput(parsed.input);
+  if (strippedPanels > 0) {
+    log2("info", `[${sessionId}] stripped ${strippedPanels} ACP panel message(s) before projection (UI-only, issue #359)`);
+  }
   const shouldInject = opts.compress.injectTool;
   const injectTools = shouldInject && !pluginMode;
   const isCompactionTrigger = hasCompactionTrigger(parsed.input);
@@ -62862,8 +62984,16 @@ function discoverRoutes(client, config) {
         if (url.protocol !== "http:" && url.protocol !== "https:") continue;
         if (hermesSeen.has(name)) continue;
         hermesSeen.add(name);
-        rewriteKeys.add(name);
-        httpRewrites.push({ key: name, realUpstream: real });
+        if (url.protocol === "https:") {
+          const host = url.hostname.toLowerCase();
+          if (host && !httpsSeen.has(host)) {
+            httpsSeen.add(host);
+            httpsDomains.push(host);
+          }
+        } else if (!rewriteKeys.has(name)) {
+          rewriteKeys.add(name);
+          httpRewrites.push({ key: name, realUpstream: real });
+        }
       } catch {
       }
     }
@@ -62894,8 +63024,23 @@ function discoverRoutes(client, config) {
 function discoverDomains(client, config) {
   return discoverRoutes(client, config).httpsDomains;
 }
-function buildPiEnv(origin, caPath, baseEnv) {
-  return { ...baseEnv, HTTPS_PROXY: origin, NODE_EXTRA_CA_CERTS: caPath, BILLION_CONTEXT_PROXY: origin };
+function buildPiEnv(origin, caPath, baseEnv, httpRewrites = [], httpsRewrites = []) {
+  const manifest = {};
+  for (const r of httpRewrites) {
+    if (r.key.length === 0 || r.realUpstream.length === 0) continue;
+    manifest[r.key] = wrapUpstream(origin, r.realUpstream);
+  }
+  for (const r of httpsRewrites) {
+    if (r.key.length === 0 || r.realUpstream.length === 0) continue;
+    manifest[r.key] = r.realUpstream;
+  }
+  return {
+    ...baseEnv,
+    HTTPS_PROXY: origin,
+    NODE_EXTRA_CA_CERTS: caPath,
+    BILLION_CONTEXT_PROXY: origin,
+    ...Object.keys(manifest).length > 0 ? { BILI_PROVIDER_REWRITES: JSON.stringify(manifest) } : {}
+  };
 }
 function buildCodexEnv(origin, caPath, baseEnv) {
   return { ...baseEnv, HTTPS_PROXY: origin, SSL_CERT_FILE: caPath, BILLION_CONTEXT_PROXY: origin };
@@ -63426,363 +63571,6 @@ function writeOverlayFileAtomic(overlay, fileName, contents) {
     }
   }
 }
-function writePiCompactionDisabledSettings(piHome, overlay) {
-  let settings = {};
-  try {
-    const parsed = JSON.parse(fs12.readFileSync(path15.join(piHome, "settings.json"), "utf8"));
-    if (parsed && typeof parsed === "object" && !Array.isArray(parsed)) {
-      settings = parsed;
-    }
-  } catch {
-  }
-  const compactionRaw = settings.compaction;
-  const compaction = compactionRaw && typeof compactionRaw === "object" && !Array.isArray(compactionRaw) ? { ...compactionRaw } : {};
-  compaction.enabled = false;
-  settings.compaction = compaction;
-  writeOverlayFileAtomic(overlay, "settings.json", JSON.stringify(settings, null, 2));
-}
-function preparePiHttpRewrite(piHome, origin, httpRewrites, httpsRewrites) {
-  if (!fs12.existsSync(piHome)) return void 0;
-  const overlay = `${piHome}-bili`;
-  const generated = ["settings.json"];
-  let modelsRoot;
-  try {
-    const parsed = JSON.parse(fs12.readFileSync(path15.join(piHome, "models.json"), "utf8"));
-    if (parsed && typeof parsed === "object" && !Array.isArray(parsed)) {
-      modelsRoot = parsed;
-      generated.push("models.json");
-    }
-  } catch {
-  }
-  if (!refreshOverlayHome(piHome, overlay, generated)) return void 0;
-  if (modelsRoot) {
-    const providersVal = modelsRoot.providers;
-    if (providersVal && typeof providersVal === "object" && !Array.isArray(providersVal)) {
-      const providers = providersVal;
-      for (const r of httpRewrites) {
-        const prov = providers[r.key];
-        if (prov && typeof prov === "object" && !Array.isArray(prov)) {
-          const p2 = prov;
-          const existing = typeof p2.baseUrl === "string" ? p2.baseUrl : r.realUpstream;
-          p2.baseUrl = wrapUpstream(origin, unwrapUpstream2(existing));
-        }
-      }
-      for (const r of httpsRewrites) {
-        const prov = providers[r.key];
-        if (prov && typeof prov === "object" && !Array.isArray(prov)) {
-          const p2 = prov;
-          p2.baseUrl = r.realUpstream;
-        }
-      }
-    }
-    writeOverlayFileAtomic(overlay, "models.json", JSON.stringify(modelsRoot));
-  }
-  writePiCompactionDisabledSettings(piHome, overlay);
-  return overlay;
-}
-function atomicWriteTextFile(filePath, contents) {
-  const draft = `${filePath}.${process.pid}.bili-tmp`;
-  try {
-    fs12.writeFileSync(draft, contents);
-    fs12.renameSync(draft, filePath);
-  } catch {
-    try {
-      fs12.rmSync(draft, { force: true });
-    } catch {
-    }
-    throw new Error(`bili: could not write ${filePath}`);
-  }
-}
-function liveProxyPorts() {
-  const ports = /* @__PURE__ */ new Set();
-  const inst = readProxyInstanceFile();
-  if (isProxyInstanceFile(inst)) {
-    if (isPidAlive(inst.pid)) ports.add(inst.port);
-  } else if (inst) {
-    try {
-      ports.add(new URL(inst.origin).port === "" ? 80 : Number(new URL(inst.origin).port));
-    } catch {
-    }
-  }
-  return ports;
-}
-function unpackDeadProxyUrlsInFile(filePath, livePorts) {
-  let txt;
-  try {
-    txt = fs12.readFileSync(filePath, "utf8");
-  } catch {
-    return 0;
-  }
-  const re2 = /https?:\/\/(?:127\.0\.0\.1|localhost|\[::1\]):(\d+)\/bili\/(https?:\/\/\S+)/g;
-  let changed = 0;
-  const out = txt.replace(re2, (full, portStr, raw) => {
-    if (livePorts.has(Number(portStr))) return full;
-    changed += 1;
-    return raw;
-  });
-  if (changed === 0) return 0;
-  try {
-    atomicWriteTextFile(filePath, out);
-  } catch {
-    return 0;
-  }
-  return changed;
-}
-function mergeOmpCompactionDisabledYaml(text) {
-  const lines = text.split(/\r?\n/);
-  let compactionLine = -1;
-  for (let i = 0; i < lines.length; i++) {
-    const rawLine = lines[i];
-    const trimmed = rawLine.trim();
-    if (!trimmed || trimmed.startsWith("#")) continue;
-    const indent = rawLine.length - rawLine.trimStart().length;
-    if (indent === 0 && /^compaction:\s*(#.*)?$/.test(trimmed)) {
-      compactionLine = i;
-      break;
-    }
-  }
-  if (compactionLine === -1) {
-    const base = text === "" || text.endsWith("\n") ? text : `${text}
-`;
-    return `${base}compaction:
-  enabled: false
-`;
-  }
-  for (let i = compactionLine + 1; i < lines.length; i++) {
-    const rawLine = lines[i];
-    const trimmed = rawLine.trim();
-    if (!trimmed || trimmed.startsWith("#")) continue;
-    const indent = rawLine.length - rawLine.trimStart().length;
-    if (indent === 0) break;
-    if (/^enabled:/.test(trimmed)) {
-      const leading = rawLine.slice(0, rawLine.length - rawLine.trimStart().length);
-      const commentMatch = /\s+#.*$/.exec(rawLine);
-      const comment = commentMatch ? commentMatch[0] : "";
-      lines[i] = `${leading}enabled: false${comment}`;
-      return lines.join("\n");
-    }
-  }
-  let childIndent = 2;
-  for (let i = compactionLine + 1; i < lines.length; i++) {
-    const rawLine = lines[i];
-    const trimmed = rawLine.trim();
-    if (!trimmed || trimmed.startsWith("#")) continue;
-    const indent = rawLine.length - rawLine.trimStart().length;
-    if (indent === 0) break;
-    childIndent = indent;
-    break;
-  }
-  lines.splice(compactionLine + 1, 0, `${" ".repeat(childIndent)}enabled: false`);
-  return lines.join("\n");
-}
-function writeOmpCompactionDisabledConfig(ompHome, overlay) {
-  let base;
-  let baseIsJson = false;
-  for (const name of ["config.yml", "config.yaml"]) {
-    try {
-      base = fs12.readFileSync(path15.join(ompHome, name), "utf8");
-      break;
-    } catch {
-    }
-  }
-  if (base === void 0) {
-    try {
-      base = fs12.readFileSync(path15.join(ompHome, "settings.json"), "utf8");
-      baseIsJson = true;
-    } catch {
-    }
-  }
-  let merged;
-  if (base === void 0) {
-    merged = "compaction:\n  enabled: false\n";
-  } else if (baseIsJson) {
-    let settings = {};
-    try {
-      const parsed = JSON.parse(base);
-      if (parsed && typeof parsed === "object" && !Array.isArray(parsed)) {
-        settings = parsed;
-      }
-    } catch {
-    }
-    const compactionRaw = settings.compaction;
-    const compaction = compactionRaw && typeof compactionRaw === "object" && !Array.isArray(compactionRaw) ? { ...compactionRaw } : {};
-    compaction.enabled = false;
-    settings.compaction = compaction;
-    merged = JSON.stringify(settings, null, 2);
-  } else {
-    merged = mergeOmpCompactionDisabledYaml(base);
-  }
-  writeOverlayFileAtomic(overlay, "config.yml", merged);
-}
-function prepareOmpHttpRewrite(ompHome, origin, httpRewrites, httpsRewrites) {
-  if (!fs12.existsSync(ompHome)) return void 0;
-  const overlay = `${ompHome}-bili`;
-  const generated = ["config.yml"];
-  const modelsPath = path15.join(ompHome, "models.yml");
-  const unpacked = unpackDeadProxyUrlsInFile(modelsPath, liveProxyPorts());
-  if (unpacked > 0) {
-    console.error(`bili: unpacked ${unpacked} dead proxy URL(s) from the real ${modelsPath}`);
-  }
-  let modelsText;
-  try {
-    modelsText = fs12.readFileSync(modelsPath, "utf8");
-    generated.push("models.yml");
-  } catch {
-  }
-  if (!refreshOverlayHome(ompHome, overlay, generated)) return void 0;
-  if (modelsText !== void 0) {
-    const httpMap = new Map(httpRewrites.map((r) => [r.key, wrapUpstream(origin, r.realUpstream)]));
-    const httpsMap = new Map(httpsRewrites.map((r) => [r.key, r.realUpstream]));
-    const lines = modelsText.split(/\r?\n/);
-    let providersIndent = -1;
-    let providerIndent = -1;
-    let currentProvider = null;
-    for (let i = 0; i < lines.length; i++) {
-      const rawLine = lines[i];
-      const trimmed = rawLine.trim();
-      if (!trimmed || trimmed.startsWith("#")) continue;
-      const indent = rawLine.length - rawLine.trimStart().length;
-      if (providersIndent === -1) {
-        if (/^providers:\s*(#.*)?$/.test(trimmed)) providersIndent = indent;
-        continue;
-      }
-      if (indent <= providersIndent) break;
-      if (providerIndent === -1) providerIndent = indent;
-      if (indent === providerIndent) {
-        const m2 = /^([A-Za-z0-9_.-]+):/.exec(trimmed);
-        currentProvider = m2 ? m2[1] : null;
-      } else if (indent > providerIndent && currentProvider) {
-        const baseMatch = /^(baseUrl:\s*)(\S+)/.exec(trimmed);
-        if (baseMatch) {
-          const target = httpMap.get(currentProvider) ?? httpsMap.get(currentProvider);
-          if (target) {
-            const leading = rawLine.slice(0, rawLine.length - rawLine.trimStart().length);
-            const commentMatch = /\s+#.*$/.exec(rawLine);
-            const comment = commentMatch ? commentMatch[0] : "";
-            lines[i] = `${leading}baseUrl: ${target}${comment}`;
-          }
-        }
-      }
-    }
-    writeOverlayFileAtomic(overlay, "models.yml", lines.join("\n"));
-  }
-  writeOmpCompactionDisabledConfig(ompHome, overlay);
-  return overlay;
-}
-function prepareHermesHome(hermesHome, origin, rewrites) {
-  if (rewrites.length === 0) return void 0;
-  const cfgPath = path15.join(hermesHome, "config.yaml");
-  let txt;
-  try {
-    txt = fs12.readFileSync(cfgPath, "utf8");
-  } catch {
-    return void 0;
-  }
-  const wrapSet = new Set(rewrites.map((r) => r.realUpstream));
-  const eol = txt.includes("\r\n") ? "\r\n" : "\n";
-  const lines = txt.split(/\r?\n/);
-  let changed = false;
-  for (let i = 0; i < lines.length; i++) {
-    const rawLine = lines[i];
-    const m2 = /^(\s*(?:api|base_url|url):\s*)(\S+)(\s+#.*)?$/.exec(rawLine);
-    if (!m2) continue;
-    const rawUrl = m2[2].replace(/^["']|["']$/g, "");
-    if (!/^https?:\/\//i.test(rawUrl)) continue;
-    const real = unwrapUpstream2(rawUrl);
-    if (!wrapSet.has(real)) continue;
-    lines[i] = `${m2[1]}${wrapUpstream(origin, real)}${m2[3] ?? ""}`;
-    changed = true;
-  }
-  if (!changed) return void 0;
-  const overlay = `${hermesHome}-bili`;
-  if (!refreshOverlayHome(hermesHome, overlay, "config.yaml")) return void 0;
-  writeOverlayFileAtomic(overlay, "config.yaml", lines.join(eol));
-  return overlay;
-}
-function writeHermesIdentityPlugin(hermesHome, overlay, rewrites) {
-  try {
-    if (fs12.existsSync(path15.join(hermesHome, "plugins"))) return false;
-  } catch {
-    return false;
-  }
-  const dir = path15.join("plugins", "model-providers", "bili-session-identity");
-  try {
-    fs12.mkdirSync(path15.join(overlay, dir), { recursive: true });
-  } catch {
-    return false;
-  }
-  const keys = rewrites.map((r) => r.key);
-  const pluginDir = path15.join(overlay, dir);
-  try {
-    fs12.writeFileSync(path15.join(pluginDir, "__init__.py"), hermesIdentityPluginSource(keys));
-    fs12.writeFileSync(path15.join(pluginDir, "plugin.yaml"), [
-      "name: bili-session-identity",
-      "kind: model-provider",
-      "version: 1.0.0",
-      "description: billion-context session identity (installed by `bili hermes` \u2014 safe to delete)",
-      "author: billion-context",
-      ""
-    ].join("\n"));
-  } catch {
-    return false;
-  }
-  try {
-    return fs12.existsSync(path15.join(overlay, dir, "__init__.py"));
-  } catch {
-    return false;
-  }
-}
-function hermesIdentityPluginSource(providerKeys) {
-  return [
-    '"""Installed by `bili hermes` (billion-context) \u2014 safe to delete.',
-    "",
-    "Re-registers hermes's `custom` provider profile (plus a `custom:<name>`",
-    "profile for every provider the proxy fronts) with one addition: the",
-    "REAL hermes session id is sent as `prompt_cache_key` on every request,",
-    "which the billion-context proxy uses as the stable conversation",
-    "identity for its compression sessions. Removing this file restores",
-    'stock behavior (no prompt_cache_key)."""',
-    "from providers import get_provider_profile, register_provider",
-    "from providers.base import ProviderProfile",
-    "",
-    `_PROVIDER_KEYS = ${JSON.stringify(providerKeys)}`,
-    "",
-    '_custom = get_provider_profile("custom")',
-    "_BASE = type(_custom) if _custom is not None else ProviderProfile",
-    "",
-    "",
-    "class _BiliSessionIdentity(_BASE):",
-    "    def build_api_kwargs_extras(self, *, session_id=None, **ctx):",
-    "        extra_body, top_level = _BASE.build_api_kwargs_extras(",
-    "            self, session_id=session_id, **ctx",
-    "        )",
-    '        sid = str(session_id or "").strip()',
-    "        if sid:",
-    '            top_level.setdefault("prompt_cache_key", sid[:64])',
-    "        return extra_body, top_level",
-    "",
-    "",
-    "def _register(name, template, with_aliases):",
-    "    fields = {}",
-    "    if template is not None:",
-    "        # Aliases are copied ONLY for the canonical custom",
-    "        # re-registration: copying them onto custom:<name> entries",
-    "        # would steal the aliases (register_provider re-points each",
-    "        # alias at the registering name).",
-    '        for attr in (("aliases",) if with_aliases else ()) + ("env_vars", "base_url", "default_max_tokens"):',
-    "            value = getattr(template, attr, None)",
-    "            if value:",
-    "                fields[attr] = value",
-    "    register_provider(_BiliSessionIdentity(name=name, **fields))",
-    "",
-    "",
-    "if _custom is not None:",
-    '    _register("custom", _custom, True)',
-    "for _key in _PROVIDER_KEYS:",
-    '    _register("custom:" + _key, _custom, False)',
-    ""
-  ].join("\n");
-}
 function prepareDshHome(dshHome, origin, rewrites) {
   const cfgPath = path15.join(dshHome, "settings.yaml");
   let txt;
@@ -63944,7 +63732,16 @@ function pickEphemeralPort(host = LAUNCHER_DEFAULT_HOST) {
     });
   });
 }
-var INHERITED_PROXY_VARS = ["http_proxy", "https_proxy", "all_proxy", "HTTP_PROXY", "HTTPS_PROXY", "ALL_PROXY"];
+var INHERITED_PROXY_VARS = [
+  "http_proxy",
+  "https_proxy",
+  "all_proxy",
+  "HTTP_PROXY",
+  "HTTPS_PROXY",
+  "ALL_PROXY",
+  "no_proxy",
+  "NO_PROXY"
+];
 function stripInheritedProxy(env) {
   const cleaned = { ...env };
   for (const key of INHERITED_PROXY_VARS) delete cleaned[key];
@@ -64098,9 +63895,30 @@ async function runLaunch(params, deps = {}) {
   const port = parsePort(params.overrides.ACP_PORT ?? process.env.ACP_PORT);
   const passthrough = params.overrides.ACP_PASSTHROUGH === "1";
   const debug = params.overrides.ACP_DEBUG === "1";
-  const config = loadClientConfig(process.env, process.cwd());
   const base = baseClientName(params.client);
+  const discoveryEnv = base === "pi" || base === "omp" ? { ...process.env, PI_CODING_AGENT_DIR: void 0 } : process.env;
+  const config = loadClientConfig(discoveryEnv, process.cwd());
   const routes = discoverRoutes(base, config);
+  const piRealHome = resolvePiHome({ ...process.env, PI_CODING_AGENT_DIR: void 0 });
+  const ompRealHome = resolveOmpHome({ ...process.env, PI_CODING_AGENT_DIR: void 0 });
+  if (base === "pi" && (routes.httpRewrites.length > 0 || routes.httpsRewrites.length > 0)) {
+    const piExt = selfDistFile("agent/pi.js");
+    const extAvailable = piExt !== void 0 && fs12.existsSync(piExt) || piPluginInstalled(piRealHome);
+    if (!extAvailable) {
+      throw new Error(
+        "bili: pi needs provider URL rewrites but the bili extension cannot load (dist/agent/pi.js missing and the plugin is not installed in ~/.pi/agent/settings.json) \u2014 reinstall billion-context or run `bili plugin install pi`"
+      );
+    }
+  }
+  if (base === "omp" && routes.httpRewrites.length > 0) {
+    const ompExt = selfDistFile("agent/omp.js");
+    const extAvailable = ompExt !== void 0 && fs12.existsSync(ompExt) || ompPluginLoadedFrom(ompRealHome);
+    if (!extAvailable) {
+      throw new Error(
+        "bili: omp needs provider URL rewrites but the bili extension cannot load (dist/agent/omp.js missing and the plugin is not installed in ~/.omp/agent/config.yml) \u2014 reinstall billion-context or run `bili plugin install omp`"
+      );
+    }
+  }
   const biliRoutes = loadRoutes(process.env);
   const domains = dedupeInOrder([...routes.httpsDomains, ...params.mitmDomains ?? []]);
   const handle2 = await ensureProxyRunning({ host, port, passthrough, debug, mitmDomains: domains, modelWindows: collectModelWindows(config, base) }, deps);
@@ -64113,10 +63931,7 @@ async function runLaunch(params, deps = {}) {
   const ca = resolveCaCertPath(process.env);
   let env;
   let clientArgs = params.clientArgs;
-  let piOverlayHome;
-  let ompOverlayHome;
   let opencodeTmpFile;
-  let hermesOverlayHome;
   let dshOverlayHome;
   const tmpFiles = [];
   const directUrl = launcherDirectUrl(process.env);
@@ -64144,19 +63959,17 @@ async function runLaunch(params, deps = {}) {
   }
   const origin = handle2.origin;
   if (base === "pi") {
-    env = buildPiEnv(origin, ca, process.env);
-    piOverlayHome = preparePiHttpRewrite(resolvePiHome(process.env), origin, routes.httpRewrites, routes.httpsRewrites);
-    if (piOverlayHome) env.PI_CODING_AGENT_DIR = piOverlayHome;
+    env = buildPiEnv(origin, ca, process.env, routes.httpRewrites, routes.httpsRewrites);
+    delete env.PI_CODING_AGENT_DIR;
     const piExt = selfDistFile("agent/pi.js");
-    if (piExt && fs12.existsSync(piExt) && !piPluginInstalled(resolvePiHome(process.env))) {
+    if (piExt && fs12.existsSync(piExt) && !piPluginInstalled(piRealHome)) {
       clientArgs = ["-e", piExt, ...clientArgs];
     }
   } else if (base === "omp") {
-    env = buildPiEnv(origin, ca, process.env);
-    ompOverlayHome = prepareOmpHttpRewrite(resolveOmpHome(process.env), origin, routes.httpRewrites, routes.httpsRewrites);
-    if (ompOverlayHome) env.PI_CODING_AGENT_DIR = ompOverlayHome;
+    env = buildPiEnv(origin, ca, process.env, routes.httpRewrites);
+    delete env.PI_CODING_AGENT_DIR;
     const ompExt = selfDistFile("agent/omp.js");
-    if (ompExt && fs12.existsSync(ompExt) && !ompPluginLoadedFrom(resolveOmpHome(process.env))) {
+    if (ompExt && fs12.existsSync(ompExt) && !ompPluginLoadedFrom(ompRealHome)) {
       clientArgs = ["-e", ompExt, ...clientArgs];
     }
   } else if (base === "opencode") {
@@ -64166,18 +63979,12 @@ async function runLaunch(params, deps = {}) {
     opencodeTmpFile = prepareOpencodeHttpRewrite(resolveOpencodeConfigFile(process.env), origin, routes.httpRewrites, routes.httpsRewrites, opencodePluginPath);
     if (opencodeTmpFile) env.OPENCODE_CONFIG = opencodeTmpFile;
   } else if (base === "hermes") {
-    env = { ...process.env };
-    hermesOverlayHome = prepareHermesHome(resolveHermesHome(process.env), origin, routes.httpRewrites);
-    if (hermesOverlayHome) {
-      env.HERMES_HOME = hermesOverlayHome;
-      if (!writeHermesIdentityPlugin(resolveHermesHome(process.env), hermesOverlayHome, routes.httpRewrites)) {
-        console.error(
-          "bili: ~/.hermes has its own plugins/ \u2014 skipping the session-identity plugin; hermes requests will be rejected by the proxy without an identity (see #286)."
-        );
-      }
-    } else {
+    env = stripInheritedProxy(process.env);
+    env.HTTPS_PROXY = origin;
+    env.HERMES_CA_BUNDLE = ca;
+    if (routes.httpRewrites.length === 0 && routes.httpsDomains.length === 0) {
       console.error(
-        routes.httpRewrites.length === 0 ? "bili: no hermes providers found in ~/.hermes/config.yaml \u2014 traffic will NOT go through the proxy (configure a provider first)." : "bili: hermes config.yaml could not be rewritten (unreadable or no matching provider endpoints) \u2014 traffic will NOT go through the proxy."
+        "bili: no hermes providers found in ~/.hermes/config.yaml \u2014 traffic will NOT go through the proxy (configure a provider first)."
       );
     }
   } else if (base === "dsh") {