llm-switcher 1.1.11 → 1.2.2

package/docs/diagrams/system-topology.mmd

@@ -0,0 +1,30 @@
+flowchart LR
+    subgraph Clients[" Dev Clients & Coding CLIs "]
+        CC["Claude Code CLI\n(/v1/messages)"]
+        CDX["OpenAI Codex CLI\n(/v1/responses)"]
+    end
+
+    subgraph Optimizers[" Optional Middle-Layer "]
+        OPT["Token Optimizers\n(Headroom / RTK)"]
+    end
+
+    subgraph Gateway[" LLM Switcher Edge Gateway (:3456) "]
+        ROUTER["Edge Router\n(Zero-Mutation)"]
+        HEALER["Healer Engine\n(Auto-Fix Schemas)"]
+        IR["Bi-Directional IR\n(Event Stream Synth)"]
+        ROUTER --> HEALER --> IR
+    end
+
+    subgraph Upstreams[" Upstream Providers "]
+        INTACT["intact Gateway\n(Recommended Pooler)"]
+        OTHER["9Router / Vertex / Other"]
+    end
+
+    CC -->|direct /v1/messages| ROUTER
+    CDX -->|direct /v1/responses| ROUTER
+    CC -.->|prune prompt| OPT
+    CDX -.->|prune prompt| OPT
+    OPT -->|forward :3456| ROUTER
+
+    IR -->|contract & pool| INTACT
+    IR -->|standard call| OTHER

package/docs/diagrams/system-topology.svg

@@ -0,0 +1,125 @@
+<svg xmlns="http://www.w3.org/2000/svg" viewBox="0 0 1712.757 298.51666666666665" width="1712.757" height="298.51666666666665" style="--bg:#1a1b26;--fg:#a9b1d6;--line:#3d59a1;--accent:#7aa2f7;--muted:#565f89;background:var(--bg)">
+<style>
+  @import url('https://fonts.googleapis.com/css2?family=Inter:wght@400;500;600;700&amp;display=swap');
+  text { font-family: 'Inter', system-ui, sans-serif; }
+  svg {
+    /* Derived from --bg and --fg (overridable via --line, --accent, etc.) */
+    --_text:          var(--fg);
+    --_text-sec:      var(--muted, color-mix(in srgb, var(--fg) 60%, var(--bg)));
+    --_text-muted:    var(--muted, color-mix(in srgb, var(--fg) 40%, var(--bg)));
+    --_text-faint:    color-mix(in srgb, var(--fg) 25%, var(--bg));
+    --_line:          var(--line, color-mix(in srgb, var(--fg) 50%, var(--bg)));
+    --_arrow:         var(--accent, color-mix(in srgb, var(--fg) 85%, var(--bg)));
+    --_node-fill:     var(--surface, color-mix(in srgb, var(--fg) 3%, var(--bg)));
+    --_node-stroke:   var(--border, color-mix(in srgb, var(--fg) 20%, var(--bg)));
+    --_group-fill:    var(--bg);
+    --_group-hdr:     color-mix(in srgb, var(--fg) 5%, var(--bg));
+    --_inner-stroke:  color-mix(in srgb, var(--fg) 12%, var(--bg));
+    --_key-badge:     color-mix(in srgb, var(--fg) 10%, var(--bg));
+  }
+</style>
+<defs>
+  <marker id="arrowhead" markerWidth="8" markerHeight="5" refX="7" refY="2.5" orient="auto">
+    <polygon points="0 0, 8 2.5, 0 5" fill="var(--_arrow)" stroke="var(--_arrow)" stroke-width="0.75" stroke-linejoin="round" />
+  </marker>
+  <marker id="arrowhead-start" markerWidth="8" markerHeight="5" refX="1" refY="2.5" orient="auto-start-reverse">
+    <polygon points="8 0, 0 2.5, 8 5" fill="var(--_arrow)" stroke="var(--_arrow)" stroke-width="0.75" stroke-linejoin="round" />
+  </marker>
+</defs>
+<g class="subgraph" data-id="Clients" data-label=" Dev Clients &amp; Coding CLIs ">
+  <rect x="40" y="40" width="180.654" height="191.60000000000002" rx="0" ry="0" fill="var(--_group-fill)" stroke="var(--_node-stroke)" stroke-width="1" />
+  <rect x="40" y="40" width="180.654" height="28" rx="0" ry="0" fill="var(--_group-hdr)" stroke="var(--_node-stroke)" stroke-width="1" />
+  <text x="52" y="54" font-size="12" font-weight="600" fill="var(--_text-sec)" dy="4.199999999999999"> Dev Clients &amp; Coding CLIs </text>
+</g>
+<g class="subgraph" data-id="Optimizers" data-label=" Optional Middle-Layer ">
+  <rect x="384.674" y="90.56666666666666" width="176.949" height="113.80000000000001" rx="0" ry="0" fill="var(--_group-fill)" stroke="var(--_node-stroke)" stroke-width="1" />
+  <rect x="384.674" y="90.56666666666666" width="176.949" height="28" rx="0" ry="0" fill="var(--_group-hdr)" stroke="var(--_node-stroke)" stroke-width="1" />
+  <text x="396.674" y="104.56666666666666" font-size="12" font-weight="600" fill="var(--_text-sec)" dy="4.199999999999999"> Optional Middle-Layer </text>
+</g>
+<g class="subgraph" data-id="Gateway" data-label=" LLM Switcher Edge Gateway (:3456) ">
+  <rect x="728.0190000000001" y="92.55000000000001" width="563.89" height="113.80000000000001" rx="0" ry="0" fill="var(--_group-fill)" stroke="var(--_node-stroke)" stroke-width="1" />
+  <rect x="728.0190000000001" y="92.55000000000001" width="563.89" height="28" rx="0" ry="0" fill="var(--_group-hdr)" stroke="var(--_node-stroke)" stroke-width="1" />
+  <text x="740.0190000000001" y="106.55000000000001" font-size="12" font-weight="600" fill="var(--_text-sec)" dy="4.199999999999999"> LLM Switcher Edge Gateway (:3456) </text>
+</g>
+<g class="subgraph" data-id="Upstreams" data-label=" Upstream Providers ">
+  <rect x="1462.4630000000002" y="48.45" width="210.29399999999998" height="174.70000000000002" rx="0" ry="0" fill="var(--_group-fill)" stroke="var(--_node-stroke)" stroke-width="1" />
+  <rect x="1462.4630000000002" y="48.45" width="210.29399999999998" height="28" rx="0" ry="0" fill="var(--_group-hdr)" stroke="var(--_node-stroke)" stroke-width="1" />
+  <text x="1474.4630000000002" y="62.45" font-size="12" font-weight="600" fill="var(--_text-sec)" dy="4.199999999999999"> Upstream Providers </text>
+</g>
+<polyline class="edge" data-from="CC" data-to="ROUTER" data-style="solid" data-arrow-start="false" data-arrow-end="true" data-label="direct /v1/messages" points="200.208,101.93333333333334 366.674,101.93333333333334 366.674,56.26666666666667 710.0190000000001,56.26666666666667 710.0190000000001,150 744.0190000000001,150" fill="none" stroke="var(--_line)" stroke-width="1" marker-end="url(#arrowhead)" />
+<polyline class="edge" data-from="CDX" data-to="ROUTER" data-style="solid" data-arrow-start="false" data-arrow-end="true" data-label="direct /v1/responses" points="209.1,197.66666666666669 366.674,197.66666666666669 366.674,239.36666666666667 710.0190000000001,239.36666666666667 710.0190000000001,176.9 744.0190000000001,176.9" fill="none" stroke="var(--_line)" stroke-width="1" marker-end="url(#arrowhead)" />
+<polyline class="edge" data-from="CC" data-to="OPT" data-style="dotted" data-arrow-start="false" data-arrow-end="true" data-label="prune prompt" points="200.208,119.86666666666667 254.654,119.86666666666667 254.654,144.81666666666666 366.674,144.81666666666666 366.674,152.5 400.674,152.5" fill="none" stroke="var(--_line)" stroke-width="1" stroke-dasharray="4 4" marker-end="url(#arrowhead)" />
+<polyline class="edge" data-from="CDX" data-to="OPT" data-style="dotted" data-arrow-start="false" data-arrow-end="true" data-label="prune prompt" points="209.1,179.73333333333335 254.654,179.73333333333335 254.654,178.11666666666667 366.674,178.11666666666667 366.674,170.43333333333334 400.674,170.43333333333334" fill="none" stroke="var(--_line)" stroke-width="1" stroke-dasharray="4 4" marker-end="url(#arrowhead)" />
+<polyline class="edge" data-from="OPT" data-to="ROUTER" data-style="solid" data-arrow-start="false" data-arrow-end="true" data-label="forward :3456" points="545.623,161.46666666666667 710.019,161.4666666666667 710.019,163.45000000000002 744.0190000000001,163.45000000000002" fill="none" stroke="var(--_line)" stroke-width="1" marker-end="url(#arrowhead)" />
+<polyline class="edge" data-from="IR" data-to="INTACT" data-style="solid" data-arrow-start="false" data-arrow-end="true" data-label="contract &amp; pool" points="1275.909,172.41666666666669 1325.909,172.41666666666669 1325.909,180.10000000000002 1444.4630000000002,180.10000000000002 1444.4630000000002,180.25 1478.4630000000002,180.25000000000003" fill="none" stroke="var(--_line)" stroke-width="1" marker-end="url(#arrowhead)" />
+<polyline class="edge" data-from="IR" data-to="OTHER" data-style="solid" data-arrow-start="false" data-arrow-end="true" data-label="standard call" points="1275.909,154.48333333333335 1333.631,154.48333333333335 1333.631,128.925 1444.4630000000002,128.925 1444.4630000000002,110.9 1478.4630000000002,110.90000000000002" fill="none" stroke="var(--_line)" stroke-width="1" marker-end="url(#arrowhead)" />
+<polyline class="edge" data-from="ROUTER" data-to="HEALER" data-style="solid" data-arrow-start="false" data-arrow-end="true" points="876.3710000000001,163.45000000000002 916.3710000000001,163.45000000000002" fill="none" stroke="var(--_line)" stroke-width="1" marker-end="url(#arrowhead)" />
+<polyline class="edge" data-from="HEALER" data-to="IR" data-style="solid" data-arrow-start="false" data-arrow-end="true" points="1072.4350000000002,163.45000000000002 1112.4350000000002,163.45000000000002" fill="none" stroke="var(--_line)" stroke-width="1" marker-end="url(#arrowhead)" />
+<g class="edge-label" data-from="CC" data-to="ROUTER" data-label="direct /v1/messages">
+  <rect x="592.2990000000001" y="40.26666666666667" width="109.72000000000001" height="30.3" rx="2" ry="2" fill="var(--bg)" stroke="var(--_inner-stroke)" stroke-width="1" />
+  <text x="647.1590000000001" y="55.41666666666667" text-anchor="middle" font-size="11" font-weight="400" fill="var(--_text-sec)" dy="3.8499999999999996">direct /v1/messages</text>
+</g>
+<g class="edge-label" data-from="CDX" data-to="ROUTER" data-label="direct /v1/responses">
+  <rect x="588.7350000000001" y="223.36666666666667" width="113.28400000000002" height="30.3" rx="2" ry="2" fill="var(--bg)" stroke="var(--_inner-stroke)" stroke-width="1" />
+  <text x="645.3770000000002" y="238.51666666666668" text-anchor="middle" font-size="11" font-weight="400" fill="var(--_text-sec)" dy="3.8499999999999996">direct /v1/responses</text>
+</g>
+<g class="edge-label" data-from="CC" data-to="OPT" data-label="prune prompt">
+  <rect x="262.654" y="128.81666666666666" width="80.02000000000001" height="30.3" rx="2" ry="2" fill="var(--bg)" stroke="var(--_inner-stroke)" stroke-width="1" />
+  <text x="302.664" y="143.96666666666667" text-anchor="middle" font-size="11" font-weight="400" fill="var(--_text-sec)" dy="3.8499999999999996">prune prompt</text>
+</g>
+<g class="edge-label" data-from="CDX" data-to="OPT" data-label="prune prompt">
+  <rect x="262.654" y="162.11666666666667" width="80.02000000000001" height="30.3" rx="2" ry="2" fill="var(--bg)" stroke="var(--_inner-stroke)" stroke-width="1" />
+  <text x="302.664" y="177.26666666666668" text-anchor="middle" font-size="11" font-weight="400" fill="var(--_text-sec)" dy="3.8499999999999996">prune prompt</text>
+</g>
+<g class="edge-label" data-from="OPT" data-to="ROUTER" data-label="forward :3456">
+  <rect x="603.623" y="145.46666666666667" width="82.39600000000002" height="30.3" rx="2" ry="2" fill="var(--bg)" stroke="var(--_inner-stroke)" stroke-width="1" />
+  <text x="644.821" y="160.61666666666667" text-anchor="middle" font-size="11" font-weight="400" fill="var(--_text-sec)" dy="3.8499999999999996">forward :3456</text>
+</g>
+<g class="edge-label" data-from="IR" data-to="INTACT" data-label="contract &amp; pool">
+  <rect x="1333.909" y="164.10000000000002" width="86.55400000000002" height="30.3" rx="2" ry="2" fill="var(--bg)" stroke="var(--_inner-stroke)" stroke-width="1" />
+  <text x="1377.1860000000001" y="179.25000000000003" text-anchor="middle" font-size="11" font-weight="400" fill="var(--_text-sec)" dy="3.8499999999999996">contract &amp; pool</text>
+</g>
+<g class="edge-label" data-from="IR" data-to="OTHER" data-label="standard call">
+  <rect x="1341.631" y="112.92500000000001" width="78.83200000000001" height="30.3" rx="2" ry="2" fill="var(--bg)" stroke="var(--_inner-stroke)" stroke-width="1" />
+  <text x="1381.047" y="128.07500000000002" text-anchor="middle" font-size="11" font-weight="400" fill="var(--_text-sec)" dy="3.8499999999999996">standard call</text>
+</g>
+<g class="node" data-id="CC" data-label="Claude Code CLI
+(/v1/messages)" data-shape="rectangle">
+  <rect x="60.446" y="84" width="139.762" height="53.800000000000004" rx="0" ry="0" fill="var(--_node-fill)" stroke="var(--_node-stroke)" stroke-width="0.75" />
+  <text x="130.327" y="110.9" text-anchor="middle" font-size="13" font-weight="500" fill="var(--_text)"><tspan x="130.327" dy="-3.9000000000000012">Claude Code CLI</tspan><tspan x="130.327" dy="16.900000000000002">(/v1/messages)</tspan></text>
+</g>
+<g class="node" data-id="CDX" data-label="OpenAI Codex CLI
+(/v1/responses)" data-shape="rectangle">
+  <rect x="60.446" y="161.8" width="148.654" height="53.800000000000004" rx="0" ry="0" fill="var(--_node-fill)" stroke="var(--_node-stroke)" stroke-width="0.75" />
+  <text x="134.773" y="188.70000000000002" text-anchor="middle" font-size="13" font-weight="500" fill="var(--_text)"><tspan x="134.773" dy="-3.9000000000000012">OpenAI Codex CLI</tspan><tspan x="134.773" dy="16.900000000000002">(/v1/responses)</tspan></text>
+</g>
+<g class="node" data-id="OPT" data-label="Token Optimizers
+(Headroom / RTK)" data-shape="rectangle">
+  <rect x="400.674" y="134.56666666666666" width="144.949" height="53.800000000000004" rx="0" ry="0" fill="var(--_node-fill)" stroke="var(--_node-stroke)" stroke-width="0.75" />
+  <text x="473.1485" y="161.46666666666667" text-anchor="middle" font-size="13" font-weight="500" fill="var(--_text)"><tspan x="473.1485" dy="-3.9000000000000012">Token Optimizers</tspan><tspan x="473.1485" dy="16.900000000000002">(Headroom / RTK)</tspan></text>
+</g>
+<g class="node" data-id="ROUTER" data-label="Edge Router
+(Zero-Mutation)" data-shape="rectangle">
+  <rect x="744.0190000000001" y="136.55" width="132.352" height="53.800000000000004" rx="0" ry="0" fill="var(--_node-fill)" stroke="var(--_node-stroke)" stroke-width="0.75" />
+  <text x="810.1950000000002" y="163.45000000000002" text-anchor="middle" font-size="13" font-weight="500" fill="var(--_text)"><tspan x="810.1950000000002" dy="-3.9000000000000012">Edge Router</tspan><tspan x="810.1950000000002" dy="16.900000000000002">(Zero-Mutation)</tspan></text>
+</g>
+<g class="node" data-id="HEALER" data-label="Healer Engine
+(Auto-Fix Schemas)" data-shape="rectangle">
+  <rect x="916.3710000000001" y="136.55" width="156.064" height="53.800000000000004" rx="0" ry="0" fill="var(--_node-fill)" stroke="var(--_node-stroke)" stroke-width="0.75" />
+  <text x="994.4030000000001" y="163.45000000000002" text-anchor="middle" font-size="13" font-weight="500" fill="var(--_text)"><tspan x="994.4030000000001" dy="-3.9000000000000012">Healer Engine</tspan><tspan x="994.4030000000001" dy="16.900000000000002">(Auto-Fix Schemas)</tspan></text>
+</g>
+<g class="node" data-id="IR" data-label="Bi-Directional IR
+(Event Stream Synth)" data-shape="rectangle">
+  <rect x="1112.4350000000002" y="136.55" width="163.474" height="53.800000000000004" rx="0" ry="0" fill="var(--_node-fill)" stroke="var(--_node-stroke)" stroke-width="0.75" />
+  <text x="1194.1720000000003" y="163.45000000000002" text-anchor="middle" font-size="13" font-weight="500" fill="var(--_text)"><tspan x="1194.1720000000003" dy="-3.9000000000000012">Bi-Directional IR</tspan><tspan x="1194.1720000000003" dy="16.900000000000002">(Event Stream Synth)</tspan></text>
+</g>
+<g class="node" data-id="INTACT" data-label="intact Gateway
+(Recommended Pooler)" data-shape="rectangle">
+  <rect x="1478.4630000000002" y="153.35000000000002" width="177.55299999999997" height="53.800000000000004" rx="0" ry="0" fill="var(--_node-fill)" stroke="var(--_node-stroke)" stroke-width="0.75" />
+  <text x="1567.2395000000001" y="180.25000000000003" text-anchor="middle" font-size="13" font-weight="500" fill="var(--_text)"><tspan x="1567.2395000000001" dy="-3.9000000000000012">intact Gateway</tspan><tspan x="1567.2395000000001" dy="16.900000000000002">(Recommended Pooler)</tspan></text>
+</g>
+<g class="node" data-id="OTHER" data-label="9Router / Vertex / Other" data-shape="rectangle">
+  <rect x="1478.4630000000002" y="92.45" width="178.29399999999998" height="36.900000000000006" rx="0" ry="0" fill="var(--_node-fill)" stroke="var(--_node-stroke)" stroke-width="0.75" />
+  <text x="1567.6100000000001" y="110.9" text-anchor="middle" font-size="13" font-weight="500" fill="var(--_text)" dy="4.55">9Router / Vertex / Other</text>
+</g>
+</svg>

package/ensure-ca-bundle.mjs

@@ -0,0 +1,28 @@
+// ============================================================
+// ensure-ca-bundle.mjs — the shim's one call into Node (R2, R4)
+//
+// The shim must point NODE_EXTRA_CA_CERTS at a bundle holding BOTH the user's own CA and the
+// switcher's ca.pem, so trusting the interceptor never costs the user a certificate they
+// already had. Building that bundle means reading files, comparing content byte for byte and
+// renaming atomically — none of which belongs in a shell script, and none at all in cmd.exe.
+//
+// It is a FILE rather than `node -e "..."` on purpose: cmd.exe re-parses quotes and reads %2
+// inside a percent-encoded URL, so a space anywhere in the checkout path would corrupt the
+// program or its arguments.
+//
+// argv: <inherited NODE_EXTRA_CA_CERTS> <switcher ca.pem> <state dir>
+// stdout: the bundle path, and nothing else — the shim compares it against -f before exporting.
+// exit 1 on any failure: the reason is already on stderr, and the shim then keeps the inherited
+// value and exports no path to a file that does not exist.
+// ============================================================
+
+import { ensureClaudeCaBundle } from './state.mjs';
+
+const [, , inheritedCa, switcherCaPath, stateDir] = process.argv;
+const result = ensureClaudeCaBundle(inheritedCa, switcherCaPath, stateDir);
+
+if (result && result.ok && result.path) {
+  process.stdout.write(result.path);
+  process.exit(0);
+}
+process.exit(1);

package/formats.mjs

@@ -19,11 +19,11 @@
 //   params: { maxTokens, temperature, topP, topK, stop[],
 //             presencePenalty?, frequencyPenalty? },
 //   thinking: { type:'enabled'|'disabled'|'adaptive', budget?, effort? } | null,
+//   responseFormat: { type:'json_object' } | { type:'json_schema', name?, schema, strict? } | null,
 //   stream: bool
 // }
 // ============================================================
 
-export const IN_FORMATS = ['anthropic', 'openai-chat', 'responses', 'vertex'];
 export const OUT_FORMATS = ['openai-chat', 'anthropic', 'vertex'];
 
 // ---------------- stop reasons ----------------
@@ -272,7 +272,10 @@ function sanitizeSchemaNode(schema) {
     }
     clean[k] = sanitizeSchemaNode(v);
   }
-  if (!clean.type && clean.properties) {
+  // An empty schema carries no shape, and Vertex reads that as a missing `type`.
+  // Heal it the way a missing schema is healed — but only when nothing else is here,
+  // so an `anyOf`/`$ref` shape is never handed a type it did not ask for.
+  if (!clean.type && (clean.properties || Object.keys(clean).length === 0)) {
     clean.type = 'object';
   }
   if (clean.type === 'object' && !clean.properties) {
@@ -325,10 +328,23 @@ function baseIR() {
   return {
     model: '', system: '', messages: [], tools: [], toolChoice: null,
     params: { maxTokens: null, temperature: null, topP: null, topK: null, stop: [] },
-    thinking: null, stream: false
+    thinking: null, responseFormat: null, stream: false
   };
 }
 
+// Structured output (Anthropic output_config.format, Responses text.format) -> IR responseFormat.
+// `text` asks for nothing.
+function responseFormatIR(f) {
+  if (!f || typeof f !== 'object') return null;
+  if (f.type === 'json_object') return { type: 'json_object' };
+  if (f.type !== 'json_schema') return null;
+  if (!f.schema || typeof f.schema !== 'object') return { type: 'json_object' };
+  const out = { type: 'json_schema', schema: f.schema };
+  if (typeof f.name === 'string' && f.name) out.name = f.name;
+  if (typeof f.strict === 'boolean') out.strict = f.strict;
+  return out;
+}
+
 // Anthropic Messages API -> IR (logic ported from the old transformAnthropicToOpenAI).
 function anthropicToIR(payload) {
   const ir = baseIR();
@@ -421,6 +437,7 @@ function anthropicToIR(payload) {
   if (Array.isArray(payload.stop_sequences)) ir.params.stop = stopList(payload.stop_sequences);
 
   ir.thinking = thinkingFromAnthropicParam(payload.thinking, payload.output_config?.effort);
+  ir.responseFormat = responseFormatIR(payload.output_config?.format || payload.output_format);
   return ir;
 }
 
@@ -455,73 +472,6 @@ function applyAllowedTools(ir, mode, list, nameOf) {
   ir.toolChoice = ir.tools.length ? (mode === 'required' ? 'required' : 'auto') : null;
 }
 
-function chatToIR(payload) {
-  const ir = baseIR();
-  ir.model = payload.model || '';
-  ir.stream = payload.stream === true;
-
-  if (Array.isArray(payload.messages)) {
-    for (const m of payload.messages) {
-      const role = m.role;
-      if (role === 'system' || role === 'developer') {
-        const t = typeof m.content === 'string' ? m.content : asText(m.content);
-        if (t) ir.system = ir.system ? `${ir.system}\n\n${t}` : t;
-        continue;
-      }
-      if (role === 'tool' || role === 'function') {
-        ir.messages.push({
-          role: 'tool', toolCallId: m.tool_call_id || m.id,
-          name: m.name, content: typeof m.content === 'string' ? m.content : asText(m.content)
-        });
-        continue;
-      }
-      const out = { role: role === 'assistant' ? 'assistant' : 'user' };
-      if (typeof m.content === 'string' || Array.isArray(m.content)) out.content = m.content;
-      else if (m.content != null) out.content = asText(m.content);
-      if (Array.isArray(m.tool_calls)) {
-        out.toolCalls = m.tool_calls.map((tc, i) => ({
-          id: tc.id || null,
-          name: tc.function?.name || tc.name || '',
-          args: parseArgs(tc.function?.arguments ?? tc.function?.args ?? tc.args)
-        })).filter(t => t.name);
-      }
-      if (out.content !== undefined || out.toolCalls) ir.messages.push(out);
-    }
-  }
-
-  const toolDefs = payload.tools;
-  if (Array.isArray(toolDefs) && toolDefs.length) {
-    ir.tools = toolDefs.map(t => {
-      const fn = t.function || t;
-      return { name: fn.name, description: fn.description || '', parameters: sanitizeJsonSchema(fn.parameters || fn.input_schema || {}) };
-    }).filter(t => t.name);
-  }
-
-  const tc = payload.tool_choice;
-  if (typeof tc === 'string') ir.toolChoice = tc;
-  else if (tc?.type === 'function' && tc.function?.name) ir.toolChoice = { name: tc.function.name };
-  else if (tc?.type === 'allowed_tools') applyAllowedTools(ir, tc.allowed_tools?.mode, tc.allowed_tools?.tools, t => t?.function?.name);
-
-  ir.params.maxTokens = payload.max_tokens ?? payload.max_completion_tokens ?? null;
-  if (typeof payload.temperature === 'number') ir.params.temperature = payload.temperature;
-  if (typeof payload.top_p === 'number') ir.params.topP = payload.top_p;
-  if (typeof payload.presence_penalty === 'number') ir.params.presencePenalty = payload.presence_penalty;
-  if (typeof payload.frequency_penalty === 'number') ir.params.frequencyPenalty = payload.frequency_penalty;
-  if (payload.stop !== undefined && payload.stop !== null) ir.params.stop = stopList(payload.stop);
-  if (payload.parallel_tool_calls === false) ir.params.parallelToolCalls = false;
-
-  if (payload.thinking && typeof payload.thinking === 'object') {
-    ir.thinking = thinkingFromAnthropicParam(payload.thinking, payload.reasoning_effort);
-  } else if (typeof payload.reasoning_effort === 'string') {
-    ir.thinking = isNoReasoningEffort(payload.reasoning_effort)
-      ? { type: 'disabled' }
-      : { type: 'enabled', budget: effortToBudget(payload.reasoning_effort), effort: payload.reasoning_effort };
-  } else if (payload.reasoning && typeof payload.reasoning === 'object') {
-    ir.thinking = thinkingFromReasoningParam(payload.reasoning);
-  }
-  return ir;
-}
-
 // OpenAI/OpenRouter `reasoning` object -> IR thinking.
 function thinkingFromReasoningParam(r) {
   if (!r || typeof r !== 'object') return null;
@@ -709,98 +659,20 @@ function responsesToIR(payload) {
   if (payload.reasoning && typeof payload.reasoning === 'object') {
     ir.thinking = thinkingFromReasoningParam(payload.reasoning);
   }
+  ir.responseFormat = responseFormatIR(payload.text?.format);
   return ir;
 }
 
-// Vertex generateContent -> IR.
-function vertexToIR(payload) {
-  const ir = baseIR();
-  ir.model = payload.model || '';
-  ir.stream = false; // proxy overrides per the :streamGenerateContent endpoint
-
-  const sys = payload.systemInstruction?.parts;
-  if (Array.isArray(sys)) {
-    ir.system = sys.map(p => p.text || '').filter(Boolean).join('\n\n');
-  } else if (typeof payload.system_instruction === 'string') {
-    ir.system = payload.system_instruction;
-  }
-
-  // Vertex has no tool call ids: auto-generate ids for functionCalls then match functionResponses by name (FIFO),
-  // so upstream OpenAI/Anthropic gets the correct tool_call <-> tool result pairs instead of "healing" them into text.
-  const pendingByName = new Map();
-  let callSeq = 0;
-
-  if (Array.isArray(payload.contents)) {
-    for (const c of payload.contents) {
-      const parts = Array.isArray(c?.parts) ? c.parts : [];
-      const texts = [];
-      const toolCalls = [];
-      for (const p of parts) {
-        if (!p || typeof p !== 'object') continue;
-        if (typeof p.text === 'string' && p.text && p.thought !== true) texts.push(p.text);
-        if (p.functionCall) {
-          const name = p.functionCall.name || '';
-          const id = p.functionCall.id || `call_vtx_${callSeq++}_${name}`;
-          if (!pendingByName.has(name)) pendingByName.set(name, []);
-          pendingByName.get(name).push(id);
-          const sig = p.thoughtSignature || p.thought_signature || null;
-          rememberToolSignature(id, sig);
-          toolCalls.push({ id, name, args: p.functionCall.args ?? {}, sig });
-        }
-        if (p.functionResponse) {
-          const fr = p.functionResponse;
-          const name = fr.name || '';
-          const queue = pendingByName.get(name);
-          const id = fr.id || (queue && queue.length ? queue.shift() : `call_vtx_orphan_${name}`);
-          ir.messages.push({
-            role: 'tool', toolCallId: id, name: name || null,
-            content: typeof fr.response === 'string' ? fr.response : JSON.stringify(fr.response ?? '')
-          });
-        }
-      }
-      if (c.role === 'model') {
-        const out = { role: 'assistant' };
-        if (texts.length) out.content = texts.join('');
-        if (toolCalls.length) out.toolCalls = toolCalls;
-        if (out.content !== undefined || out.toolCalls) ir.messages.push(out);
-      } else if (c.role === 'function') {
-        // functionResponse already pushed above
-      } else {
-        if (texts.length) ir.messages.push({ role: 'user', content: texts.join('') });
-      }
-    }
-  }
-
-  const fns = payload.tools?.flatMap(t => t.functionDeclarations || t.function_declarations || []);
-  if (Array.isArray(fns) && fns.length) {
-    ir.tools = fns.filter(f => f.name).map(f => ({
-      name: f.name, description: f.description || '',
-      parameters: sanitizeJsonSchema(f.parameters || {})
-    }));
-  }
-
-  const gc = payload.generationConfig || payload.generation_config || {};
-  if (typeof gc.maxOutputTokens === 'number') ir.params.maxTokens = gc.maxOutputTokens;
-  if (typeof gc.temperature === 'number') ir.params.temperature = gc.temperature;
-  if (typeof gc.topP === 'number') ir.params.topP = gc.topP;
-  if (Array.isArray(gc.stopSequences)) ir.params.stop = stopList(gc.stopSequences);
-  const th = gc.thinkingConfig || gc.thinking_config;
-  if (th && typeof th === 'object') {
-    if (th.thinkingBudget === 0) ir.thinking = { type: 'disabled' };
-    else if (th.thinkingBudget > 0) ir.thinking = { type: 'enabled', budget: clampBudget(th.thinkingBudget) };
-    else if (th.thinkingLevel) ir.thinking = { type: 'enabled', budget: effortToBudget(th.thinkingLevel), effort: th.thinkingLevel };
-    else if (th.includeThoughts) ir.thinking = { type: 'enabled', budget: 2048 };
-  }
-  return ir;
-}
-
+// R5: two inputs, one per tool — `anthropic` for Claude Code, `responses` for Codex. The
+// openai-chat and vertex routes were removed with their parsers; a format that reaches here
+// anyway is an error, never a silent fallback to another protocol (a different API key!).
+// The UPSTREAM formats are a different axis and are untouched: emitUpstreamBody still speaks
+// openai-chat, anthropic and vertex to whatever provider the profile points at.
 function parseToIR(clientFormat, payload) {
   switch (clientFormat) {
     case 'anthropic': return anthropicToIR(payload);
-    case 'openai-chat': return chatToIR(payload);
     case 'responses': return responsesToIR(payload);
-    case 'vertex': return vertexToIR(payload);
-    default: throw new Error(`Unknown client format: ${clientFormat}`);
+    default: throw new Error(`Unsupported input format: ${clientFormat}. Only anthropic (Claude Code) and responses (Codex) accept requests.`);
   }
 }
 
@@ -966,6 +838,13 @@ function irToChatBody(ir, model, opts = {}) {
   if (typeof ir.params.frequencyPenalty === 'number') body.frequency_penalty = ir.params.frequencyPenalty;
   if (ir.params.stop.length) body.stop = ir.params.stop;
   if (ir.params.parallelToolCalls === false && body.tools) body.parallel_tool_calls = false;
+  if (ir.responseFormat?.type === 'json_schema') {
+    const js = { name: ir.responseFormat.name || 'response', schema: ir.responseFormat.schema };
+    if (typeof ir.responseFormat.strict === 'boolean') js.strict = ir.responseFormat.strict;
+    body.response_format = { type: 'json_schema', json_schema: js };
+  } else if (ir.responseFormat?.type === 'json_object') {
+    body.response_format = { type: 'json_object' };
+  }
 
   if (wantsThinking && mode === 'native') {
     body.reasoning_effort = ir.thinking?.effort && !['max', 'xhigh'].includes(ir.thinking.effort)
@@ -1085,6 +964,8 @@ function irToAnthropicBody(ir, model) {
   if (typeof ir.params.topP === 'number') body.top_p = ir.params.topP;
   if (typeof ir.params.topK === 'number') body.top_k = ir.params.topK;
   if (ir.params.stop.length) body.stop_sequences = ir.params.stop;
+  // Anthropic has no JSON mode without a schema.
+  if (ir.responseFormat?.type === 'json_schema') body.output_config = { format: { type: 'json_schema', schema: ir.responseFormat.schema } };
   if (ir.thinking && ir.thinking.type === 'adaptive') {
     body.thinking = { type: 'adaptive' };
   } else if (ir.thinking && ir.thinking.type !== 'disabled' && body.max_tokens > 1024) {
@@ -1206,6 +1087,10 @@ function irToVertexBody(ir, model) {
       gc.thinkingConfig.thinkingBudget = clampBudget(ir.thinking.budget ?? (ir.thinking.effort ? effortToBudget(ir.thinking.effort) : 2048));
     }
   }
+  if (ir.responseFormat) {
+    gc.responseMimeType = 'application/json';
+    if (ir.responseFormat.type === 'json_schema') gc.responseSchema = toGeminiSchema(ir.responseFormat.schema);
+  }
   if (Object.keys(gc).length) body.generationConfig = gc;
   return body;
 }
@@ -1255,6 +1140,8 @@ function toGeminiSchema(schema, root, seen) {
     return schema === 'object' ? { type: 'object', properties: {} } : { type: schema };
   }
   if (!schema || typeof schema !== 'object' || Array.isArray(schema)) return schema;
+  // The same hole as the string shorthand above: an empty schema has no type Vertex can read.
+  if (Object.keys(schema).length === 0) return { type: 'object', properties: {} };
   root = root || schema;
   seen = seen || new Set();
   if (seen.has(schema)) return {};
@@ -2305,7 +2192,7 @@ export {
   smartReasoning, smartText, smartToolCalls, smartUsage, smartFinish,
   firstChoice, smartDelta, sanitizeJsonSchema, toGeminiSchema, splitParts,
   budgetToEffort, effortToBudget, clampBudget, parseArgs, stringifyArgs,
-  anthropicToIR, chatToIR, responsesToIR, vertexToIR, parseToIR,
+  anthropicToIR, responsesToIR, parseToIR,
   healToolPairs, healAnthropicPayload, rememberToolSignature, lookupToolSignature, estimateTokens, irToChatBody, irToAnthropicBody, irToVertexBody, emitUpstreamBody,
   normalizeUpstream, createUpstreamNormalizer, createCollector,
   createThinkTagSplitter, splitThinkTags,

package/icons/intact.svg

@@ -0,0 +1 @@
+<svg xmlns="http://www.w3.org/2000/svg" viewBox="0 0 32 32"><defs><linearGradient id="lg" x1="0" y1="0" x2="1" y2="1"><stop offset="0" stop-color="#7c6cf6"/><stop offset="1" stop-color="#2dd4bf"/></linearGradient></defs><rect width="32" height="32" rx="9" fill="url(#lg)"/><path d="M16 6.5 24 9.6v6.1c0 5-3.4 8.6-8 10-4.6-1.4-8-5-8-10V9.6Z" fill="none" stroke="#fff" stroke-width="2.2" stroke-linejoin="round"/><path d="m12.4 16.2 2.6 2.6 4.8-5" fill="none" stroke="#fff" stroke-width="2.4" stroke-linecap="round" stroke-linejoin="round"/></svg>

package/mcp.mjs

@@ -11,7 +11,7 @@
 // ============================================================
 
 import fs from 'node:fs';
-import { claudeSettingsPath, loadConfig as loadSharedConfig, resolvePort, readAdminToken, TARGETS } from './state.mjs';
+import { claudeSettingsPath, loadConfig as loadSharedConfig, resolvePort, readAdminToken, STATE_DIR } from './state.mjs';
 
 const VERSION = (() => {
   try { return JSON.parse(fs.readFileSync(new URL('./package.json', import.meta.url), 'utf8')).version; } catch { return '0.0.0'; }
@@ -80,7 +80,7 @@ async function postSwitch(port, target, profile) {
 const TOOLS = [
   {
     name: 'switcher_status',
-    description: 'Get live status of LLM Switcher gateway: port, active multi-CLI targets (Claude Code, Codex, OpenAI, Vertex), and 1M context flags.',
+    description: 'Get live status of LLM Switcher gateway: port, the two tool targets (Claude Code, Codex), and the active profile of each.',
     inputSchema: {
       type: 'object',
       properties: {}
@@ -98,14 +98,14 @@ const TOOLS = [
   },
   {
     name: 'switcher_switch_profile',
-    description: 'Programmatically change the active profile for a specific CLI target (e.g. Claude Code, Codex, OpenAI, Vertex) or globally.',
+    description: 'Programmatically change the active profile for one tool: Claude Code (claude) or Codex (codex).',
     inputSchema: {
       type: 'object',
       properties: {
         target: {
           type: 'string',
-          description: 'CLI target to switch: "anthropic" (Claude Code), "responses" (Codex), "openai-chat", or "vertex"',
-          enum: TARGETS
+          description: 'Tool to switch: "claude" (Claude Code) or "codex" (Codex)',
+          enum: ['claude', 'codex']
         },
         profile: {
           type: 'string',
@@ -124,6 +124,17 @@ const TOOLS = [
         limit: { type: 'number', description: 'Maximum number of recent logs to return (default: 5)' }
       }
     }
+  },
+  {
+    name: 'switcher_models',
+    description: 'Get the latest model catalog for Claude Code and Codex, including discovered upstream models and slot assignments.',
+    inputSchema: {
+      type: 'object',
+      properties: {
+        tool: { type: 'string', enum: ['claude', 'codex'], description: 'Optional: filter by tool ("claude" or "codex")' },
+        refresh: { type: 'boolean', description: 'Query official upstream endpoints for new models before returning' }
+      }
+    }
   }
 ];
 
@@ -131,17 +142,29 @@ async function handleToolCall(name, args) {
   const cfg = loadConfig();
   const port = getMcpPort();
 
+  if (name === 'switcher_models') {
+    const { loadCatalogCache, refreshCatalog } = await import('./catalog.mjs');
+    if (args?.refresh) {
+      await refreshCatalog(STATE_DIR);
+    }
+    const cache = loadCatalogCache(STATE_DIR);
+    const t = args?.tool?.toLowerCase();
+    const res = {};
+    if (!t || t === 'claude') res.claude = cache.claude;
+    if (!t || t === 'codex') res.codex = cache.codex;
+    res.updatedAt = cache.updatedAt;
+    return { content: [{ type: 'text', text: JSON.stringify(res, null, 2) }] };
+  }
+
   if (name === 'switcher_status') {
     const live = await fetchStatus(port);
     const activeMap = live?.activeProfiles || cfg.activeProfiles || {};
     const text = [
       '=== LLM Switcher Gateway Status ===',
       `Service Running: ${live ? `YES (http://127.0.0.1:${port})` : 'NO / UNREACHABLE'}`,
-      `Active Profiles by CLI:`,
-      `  - Claude Code (/v1/messages)       : [${activeMap.anthropic || 'OFF'}] ${live?.claude1MTiers?.length ? `• 1M Context ACTIVE (${live.claude1MTiers.join(', ')})` : ''}`,
-      `  - Codex CLI   (/v1/responses)      : [${activeMap.responses || 'OFF'}] ${live?.isCodex1MActive ? '• 1M Context ACTIVE' : ''}`,
-      `  - OpenAI Chat (/v1/chat/completions): [${activeMap['openai-chat'] || 'OFF'}]`,
-      `  - Vertex      (/v1beta/models/*)   : [${activeMap.vertex || 'OFF'}]`,
+      `Active Profiles by tool:`,
+      `  - Claude Code (/v1/messages)  : [${activeMap.claude || 'OFF'}]`,
+      `  - Codex CLI   (/v1/responses) : [${activeMap.codex || 'OFF'}]`,
       '',
       `Available Profiles in config: ${Object.keys(cfg.profiles || {}).join(', ')}`,
       `Dashboard Web UI: http://127.0.0.1:${port}/ui`
@@ -220,7 +243,12 @@ async function handleToolCall(name, args) {
       if (res.success) {
         return { content: [{ type: 'text', text: `Successfully set ${target || 'global'} active profile to "${profile || 'OFF'}".` }] };
       }
-      return { content: [{ type: 'text', text: `Switch failed: ${res.error || 'Unknown error'}` }], isError: true };
+      // A 409 answer carries the clashing keys. Without them the agent cannot tell which two
+      // pointers of config.json disagree, and the only advice left is "go read the file".
+      const detail = Array.isArray(res.clashingKeys) && res.clashingKeys.length
+        ? ` Clashing keys: ${res.clashingKeys.join(', ')}.`
+        : '';
+      return { content: [{ type: 'text', text: `Switch failed: ${res.error || 'Unknown error'}.${detail}` }], isError: true };
     } catch (err) {
       return { content: [{ type: 'text', text: `Gateway connection failed: ${err.message}` }], isError: true };
     }