feat(headroom): add proxy lifecycle management + dashboard UI

Build on the optional Headroom Token Saver from Carmelo Campos
(PR: feat: add optional Headroom token saver). Add managed start/stop
of the local headroom proxy from the dashboard, install detection,
status probing, and a simplified Token Saver UI.

- detect headroom CLI + python>=3.10, probe proxy /health
- spawn/stop proxy as a detached, pid-tracked process
- /api/headroom/{status,start,stop} routes, gated local-only in dashboardGuard
- one-click Start/Stop Headroom modal, no manual config needed
- claude<->openai shape conversion for /v1/compress via 9router translators

Thanks to Carmelo Campos (@carmelogunsroses) for the original Headroom integration.

Co-authored-by: Cursor <cursoragent@cursor.com>

open-sse/AGENTS.md

@@ -4,17 +4,20 @@ Provider-agnostic SSE engine: one OpenAI-style request → any provider (LLM cha
 
 ## Request lifecycle (chat)
 
-`handlers/chatCore.js` → `services/model.js` `parseModel` (resolve `provider/model`) → `executors/index.js` `getExecutor(provider)` → `translator/index.js` `translateRequest` (client format → provider format) → `executor.execute()` (streams upstream) → `translateResponse` (provider chunks → client format) → SSE out.
+`handlers/chatCore.js` → `services/model.js` `parseModel` (resolve `provider/model`) → **pre-translate hooks** (`rtk/` tool_result compress, `rtk/headroom.js` proxy compress, `rtk/caveman.js` system inject — all fail-open) → `executors/index.js` `getExecutor(provider)` → `translator/index.js` `translateRequest` (client format → provider format) → `executor.execute()` (streams upstream) → `translateResponse` (provider chunks → client format) → SSE out.
 
 ## Directory map
 
 - `config/` — ALL constants/config (no hardcode elsewhere). `providers.js`/`registry/` (provider defs), `providerModels.js` (alias→models matrix), `runtimeConfig.js` (timeouts, token limits), `*Constants.js`.
-- `translator/` — format conversion. `request/<from>-to-<to>.js`, `response/<from>-to-<to>.js`, `schema/` (enums: ROLE, CLAUDE_BLOCK…), `concerns/` (shared logic), `formats/` (per-format). See `tests/translator/AGENTS.md`.
+- `translator/` — format conversion. `request/<from>-to-<to>.js`, `response/<from>-to-<to>.js`, `schema/` (enums: ROLE, CLAUDE_BLOCK…), `concerns/` (shared logic), `formats.js`+`formats/` (per-format). `index.js` is the registry/entry.
 - `executors/` — per-provider upstream call. `base.js` (BaseExecutor), one file per special provider, `index.js` map.
 - `providers/` — registry build + `capabilities.js` + `pricing.js`. Entry: `index.js` (PROVIDERS).
-- `handlers/` — per-modality cores (chat/image/embedding/tts/stt/search) + sub-provider folders.
-- `services/` — `tokenRefresh/`, `usage/`, `combo.js`, `accountFallback.js`, `model.js`.
-- `utils/` — streamHandler, error, sessionManager, claudeCloaking.
+- `handlers/` — per-modality cores (chat/image/embedding/tts/stt/search) + sub-provider folders. `chatCore/` has the streaming/non-streaming/sse-to-json handlers.
+- `rtk/` — request token-killer. `index.js` compresses `tool_result` content in-place (OpenAI/Claude/Kiro shapes); `filters/` per-tool compressors + `autodetect.js`; `headroom.js` external compress proxy; `caveman.js` system-prompt injector.
+- `transformer/` — `responsesTransformer.js` (Chat Completions SSE → Codex Responses API SSE), `streamToJsonConverter.js`.
+- `shared/` — cross-provider auth/identity: `clineAuth.js`, `machineId.js`, `qoder/`.
+- `services/` — `model.js`, `provider.js`, `accountFallback.js`, `combo.js`, `compact.js`, `tokenRefresh/`+`tokenRefresh.js`, `oauthCredentialManager.js`, `usage/`, `projectId.js`, `kiroModels.js`/`qoderModels.js`.
+- `utils/` — streamHandler, stream, sse, error, sessionManager, claudeCloaking, clientDetector, proxyFetch (patches global fetch), cursorProtobuf/cursorChecksum, ollamaTransform.
 
 ## Conventions
 
@@ -33,3 +36,4 @@ Provider-agnostic SSE engine: one OpenAI-style request → any provider (LLM cha
 - OpenAI bridge is lossy (thinking, non-base64 images, tool ids, is_error) — prefer a direct route for fragile pairs.
 - `registry/index.js` is an auto-generated static import list; regenerate it (don't hand-edit) after adding a `registry/{id}.js`. REGISTRY_TEMPLATE is excluded by design.
 - Special binary/protobuf formats (kiro EventStream, cursor protobuf, commandcode NDJSON) don't round-trip through OpenAI — handle in their executor.
+- `rtk/` + `headroom.js` mutate the request body in-place and are **fail-open**: any error returns null and leaves the body untouched — never throw out of them. RTK skips `is_error`/`status:"error"` tool results to preserve traces.

open-sse/executors/codex.js

@@ -22,7 +22,8 @@ const SERVER_ID_PATTERN = /^(rs|fc|resp|msg)_/;
 // Hosted tool types that Codex/OpenAI Responses executes server-side
 const CODEX_HOSTED_TOOL_TYPES = new Set([
   "image_generation", "web_search", "web_search_preview", "file_search",
-  "computer", "computer_use_preview", "code_interpreter", "mcp", "local_shell"
+  "computer", "computer_use_preview", "code_interpreter", "mcp", "local_shell",
+  "tool_search"
 ]);
 
 // Allowlist of fields accepted by Codex Responses API — anything else is stripped

open-sse/executors/default.js

@@ -116,6 +116,11 @@ export class DefaultExecutor extends BaseExecutor {
   }
 
   buildUrl(model, stream, urlIndex = 0, credentials = null) {
+    // Runtime transport (multi-endpoint providers): use the sourceFormat-matched endpoint
+    const rt = credentials?.runtimeTransport;
+    if (rt?.baseUrl) {
+      return rt.urlSuffix ? `${rt.baseUrl}${rt.urlSuffix}` : rt.baseUrl;
+    }
     if (this.provider?.startsWith?.("openai-compatible-")) {
       const baseUrl = credentials?.providerSpecificData?.baseUrl || OPENAI_COMPAT_BASE;
       const normalized = baseUrl.replace(/\/$/, "");
@@ -156,8 +161,9 @@ export class DefaultExecutor extends BaseExecutor {
   }
 
   buildHeaders(credentials, stream = true) {
-    const headers = { "Content-Type": "application/json", ...this.config.headers };
-    const desc = AUTH_DESCRIPTORS[this.provider] || this.resolveAuthDescriptor();
+    const rt = credentials?.runtimeTransport;
+    const headers = { "Content-Type": "application/json", ...(rt ? rt.headers : this.config.headers) };
+    const desc = rt?.auth || AUTH_DESCRIPTORS[this.provider] || this.resolveAuthDescriptor();
     // Hooks run BEFORE auth so dynamic overlays (claude cached headers) can't clobber the token.
     for (const hook of desc.hooks || []) HEADER_HOOKS[hook]?.(headers, credentials);
     applyAuth(headers, desc, credentials);

open-sse/executors/xiaomi-tokenplan.js

@@ -8,13 +8,13 @@ export class XiaomiTokenplanExecutor extends DefaultExecutor {
     super("xiaomi-tokenplan");
   }
 
-  // Token Plan keys are region-specific — always OpenAI-compatible /chat/completions
+  // Token Plan keys are region-specific. Route per sourceFormat-matched transport:
+  // claude → Anthropic /anthropic/v1/messages, openai → /chat/completions.
   buildUrl(model, stream, urlIndex = 0, credentials = null) {
     const baseUrl = resolveXiaomiTokenplanBaseUrl(credentials);
-    // Claude-native aliases route to the Anthropic-compatible messages endpoint
-    // if (getModelTargetFormat(this.provider, model) === FORMATS.CLAUDE) {
-    //   return `${baseUrl.replace(/\/v1\/?$/, "/anthropic/v1")}/messages`;
-    // }
+    if (credentials?.runtimeTransport?.format === "claude") {
+      return `${baseUrl.replace(/\/v1\/?$/, "")}/anthropic/v1/messages`;
+    }
     return `${baseUrl}/chat/completions`;
   }
 }

open-sse/handlers/chatCore.js

@@ -1,4 +1,4 @@
-import { detectFormat, getTargetFormat } from "../services/provider.js";
+import { detectFormat, getTargetFormat, resolveTransport } from "../services/provider.js";
 import { translateRequest } from "../translator/index.js";
 import { FORMATS } from "../translator/formats.js";
 import { normalizeClaudePassthrough } from "../translator/formats/claude.js";
@@ -20,7 +20,9 @@ import { handleStreamingResponse, buildOnStreamComplete } from "./chatCore/strea
 import { detectClientTool, isNativePassthrough } from "../utils/clientDetector.js";
 import { dedupeTools } from "../utils/toolDeduper.js";
 import { injectCaveman } from "../rtk/caveman.js";
+import { injectPonytail } from "../rtk/ponytail.js";
 import { compressMessages, formatRtkLog } from "../rtk/index.js";
+import { compressWithHeadroom, formatHeadroomLog } from "../rtk/headroom.js";
 import { getCapabilitiesForModel } from "../providers/capabilities.js";
 import { stripUnsupportedModalities } from "../translator/concerns/modality.js";
 import { prefetchRemoteImages } from "../translator/concerns/prefetch.js";
@@ -32,7 +34,7 @@ import { prefetchRemoteImages } from "../translator/concerns/prefetch.js";
  * @param {object} options.credentials - Provider credentials
  * @param {string} options.sourceFormatOverride - Override detected source format (e.g. "openai-responses")
  */
-export async function handleChatCore({ body, modelInfo, credentials, log, onCredentialsRefreshed, onRequestSuccess, onDisconnect, clientRawRequest, connectionId, userAgent, apiKey, ccFilterNaming, rtkEnabled, cavemanEnabled, cavemanLevel, sourceFormatOverride, providerThinking }) {
+export async function handleChatCore({ body, modelInfo, credentials, log, onCredentialsRefreshed, onRequestSuccess, onDisconnect, clientRawRequest, connectionId, userAgent, apiKey, ccFilterNaming, rtkEnabled, headroomEnabled, headroomUrl, headroomCompressUserMessages, cavemanEnabled, cavemanLevel, ponytailEnabled, ponytailLevel, sourceFormatOverride, providerThinking }) {
   const { provider, model } = modelInfo;
   const requestStartTime = Date.now();
 
@@ -44,7 +46,10 @@ export async function handleChatCore({ body, modelInfo, credentials, log, onCred
 
   const alias = PROVIDER_ID_TO_ALIAS[provider] || provider;
   const modelTargetFormat = getModelTargetFormat(alias, model);
-  const targetFormat = modelTargetFormat || getTargetFormat(provider);
+  // Multi-endpoint providers: pick transport matching sourceFormat → zero translation
+  const runtimeTransport = resolveTransport(provider, sourceFormat);
+  const targetFormat = modelTargetFormat || runtimeTransport?.format || getTargetFormat(provider);
+  if (runtimeTransport && credentials) credentials.runtimeTransport = runtimeTransport;
   const stripList = getModelStrip(alias, model);
   const upstreamModel = getModelUpstreamId(alias, model);
 
@@ -149,12 +154,23 @@ export async function handleChatCore({ body, modelInfo, credentials, log, onCred
   const rtkLine = formatRtkLog(rtkStats);
   if (rtkLine) console.log(rtkLine);
 
+  // Headroom: optional external proxy compression; fail open if proxy is absent.
+  const headroomStats = await compressWithHeadroom(translatedBody, { enabled: headroomEnabled, url: headroomUrl, model: upstreamModel, format: finalFormat, compressUserMessages: headroomCompressUserMessages });
+  const headroomLine = formatHeadroomLog(headroomStats);
+  if (headroomLine) log?.info?.("HEADROOM", headroomLine);
+
   // Caveman: inject terse-style system prompt
   if (cavemanEnabled && cavemanLevel) {
     injectCaveman(translatedBody, finalFormat, cavemanLevel);
     log?.debug?.("CAVEMAN", `${cavemanLevel} | ${finalFormat}`);
   }
 
+  // Ponytail: inject lazy-senior-dev system prompt
+  if (ponytailEnabled && ponytailLevel) {
+    injectPonytail(translatedBody, finalFormat, ponytailLevel);
+    log?.debug?.("PONYTAIL", `${ponytailLevel} | ${finalFormat}`);
+  }
+
   const executor = getExecutor(provider);
   trackPendingRequest(model, provider, connectionId, true);
   appendRequestLog({ model, provider, connectionId, status: "PENDING" }).catch(() => { });

open-sse/providers/index.js

@@ -32,7 +32,10 @@ export const PROVIDER_MODELS = {};
 export const PROVIDER_OAUTH = {};
 export const PROVIDER_MEDIA = {};
 for (const entry of REGISTRY) {
-  if (entry.transport) PROVIDERS[entry.id] = buildTransport(entry.transport, entry.oauth);
+  if (entry.transport) {
+    PROVIDERS[entry.id] = buildTransport(entry.transport, entry.oauth);
+    if (entry.transports) PROVIDERS[entry.id].transports = entry.transports;
+  }
   if (entry.models !== undefined) PROVIDER_MODELS[entry.alias || entry.id] = entry.models.map(normalizeModel);
   if (entry.oauth) PROVIDER_OAUTH[entry.id] = entry.oauth;
   // Build PROVIDER_MEDIA from top-level fields (post-migration) + legacy entry.media

open-sse/providers/registry/deepseek.js

@@ -1,3 +1,5 @@
+import { CLAUDE_API_HEADERS } from "../shared.js";
+
 export default {
   id: "deepseek",
   priority: 110,
@@ -24,6 +26,20 @@ export default {
       scope: "all",
     },
   },
+  // Multi-endpoint: pick the transport matching client sourceFormat to skip translation.
+  transports: [
+    {
+      format: "openai",
+      baseUrl: "https://api.deepseek.com/chat/completions",
+      auth: { combined: true, header: "Authorization", scheme: "bearer" },
+    },
+    {
+      format: "claude",
+      baseUrl: "https://api.deepseek.com/anthropic/v1/messages",
+      headers: { ...CLAUDE_API_HEADERS },
+      auth: { combined: true, header: "x-api-key", scheme: "raw" },
+    },
+  ],
   models: [
     { id: "deepseek-v4-pro", name: "DeepSeek V4 Pro" },
     { id: "deepseek-v4-pro-max", name: "DeepSeek V4 Pro Max", upstreamModelId: "deepseek-v4-pro" },

open-sse/providers/registry/glm.js

@@ -19,10 +19,7 @@ export default {
     baseUrl: "https://api.z.ai/api/anthropic/v1/messages",
     format: "claude",
     urlSuffix: "?beta=true",
-    headers: {
-      "Anthropic-Version": "2023-06-01",
-      "Anthropic-Beta": "claude-code-20250219,interleaved-thinking-2025-05-14",
-    },
+    headers: { ...CLAUDE_API_HEADERS },
     auth: {
       combined: true,
       header: "x-api-key",
@@ -32,6 +29,21 @@ export default {
       url: "https://api.z.ai/api/monitor/usage/quota/limit",
     },
   },
+  // Multi-endpoint: pick the transport matching client sourceFormat to skip translation.
+  transports: [
+    {
+      format: "openai",
+      baseUrl: "https://api.z.ai/api/coding/paas/v4/chat/completions",
+      auth: { combined: true, header: "Authorization", scheme: "bearer" },
+    },
+    {
+      format: "claude",
+      baseUrl: "https://api.z.ai/api/anthropic/v1/messages",
+      urlSuffix: "?beta=true",
+      headers: { ...CLAUDE_API_HEADERS },
+      auth: { combined: true, header: "x-api-key", scheme: "raw" },
+    },
+  ],
   models: [
     { id: "glm-5.2", name: "GLM 5.2" },
     { id: "glm-5.1", name: "GLM 5.1" },

open-sse/providers/registry/kimi-coding.js

@@ -20,10 +20,7 @@ export default {
     baseUrl: "https://api.kimi.com/coding/v1/messages",
     format: "claude",
     urlSuffix: "?beta=true",
-    headers: {
-      "Anthropic-Version": "2023-06-01",
-      "Anthropic-Beta": "claude-code-20250219,interleaved-thinking-2025-05-14",
-    },
+    headers: { ...CLAUDE_API_HEADERS },
     clientId: "17e5f671-d194-4dfb-9706-5516cb48c098",
     tokenUrl: "https://auth.kimi.com/api/oauth/token",
     refreshUrl: "https://auth.kimi.com/api/oauth/token",
@@ -36,6 +33,21 @@ export default {
       ],
     },
   },
+  // Multi-endpoint: pick the transport matching client sourceFormat to skip translation.
+  transports: [
+    {
+      format: "openai",
+      baseUrl: "https://api.kimi.com/coding/v1/chat/completions",
+      auth: { combined: true, header: "Authorization", scheme: "bearer", hooks: ["kimiHeaders"] },
+    },
+    {
+      format: "claude",
+      baseUrl: "https://api.kimi.com/coding/v1/messages",
+      urlSuffix: "?beta=true",
+      headers: { ...CLAUDE_API_HEADERS },
+      auth: { combined: true, header: "x-api-key", scheme: "raw", hooks: ["kimiHeaders"] },
+    },
+  ],
   models: [
     { id: "kimi-k2.6", name: "Kimi K2.6" },
     { id: "kimi-k2.5", name: "Kimi K2.5" },

open-sse/providers/registry/kimi.js

@@ -19,16 +19,28 @@ export default {
     baseUrl: "https://api.kimi.com/coding/v1/messages",
     format: "claude",
     urlSuffix: "?beta=true",
-    headers: {
-      "Anthropic-Version": "2023-06-01",
-      "Anthropic-Beta": "claude-code-20250219,interleaved-thinking-2025-05-14",
-    },
+    headers: { ...CLAUDE_API_HEADERS },
     auth: {
       combined: true,
       header: "x-api-key",
       scheme: "raw",
     },
   },
+  // Multi-endpoint: pick the transport matching client sourceFormat to skip translation.
+  transports: [
+    {
+      format: "openai",
+      baseUrl: "https://api.kimi.com/coding/v1/chat/completions",
+      auth: { combined: true, header: "Authorization", scheme: "bearer" },
+    },
+    {
+      format: "claude",
+      baseUrl: "https://api.kimi.com/coding/v1/messages",
+      urlSuffix: "?beta=true",
+      headers: { ...CLAUDE_API_HEADERS },
+      auth: { combined: true, header: "x-api-key", scheme: "raw" },
+    },
+  ],
   models: [
     { id: "kimi-k2.6", name: "Kimi K2.6" },
     { id: "kimi-k2.5", name: "Kimi K2.5" },

open-sse/providers/registry/minimax-cn.js

@@ -19,10 +19,7 @@ export default {
     baseUrl: "https://api.minimaxi.com/anthropic/v1/messages",
     format: "claude",
     urlSuffix: "?beta=true",
-    headers: {
-      "Anthropic-Version": "2023-06-01",
-      "Anthropic-Beta": "claude-code-20250219,interleaved-thinking-2025-05-14",
-    },
+    headers: { ...CLAUDE_API_HEADERS },
     quirks: {
       dropOutputConfig: true,
     },
@@ -41,6 +38,21 @@ export default {
       ],
     },
   },
+  // Multi-endpoint: pick the transport matching client sourceFormat to skip translation.
+  transports: [
+    {
+      format: "openai",
+      baseUrl: "https://api.minimaxi.com/v1/chat/completions",
+      auth: { combined: true, header: "Authorization", scheme: "bearer" },
+    },
+    {
+      format: "claude",
+      baseUrl: "https://api.minimaxi.com/anthropic/v1/messages",
+      urlSuffix: "?beta=true",
+      headers: { ...CLAUDE_API_HEADERS },
+      auth: { combined: true, header: "x-api-key", scheme: "raw" },
+    },
+  ],
   models: [
     { id: "MiniMax-M3", name: "MiniMax M3", targetFormat: "claude" },
     { id: "MiniMax-M2.7", name: "MiniMax M2.7" },

open-sse/providers/registry/minimax.js

@@ -19,10 +19,7 @@ export default {
     baseUrl: "https://api.minimax.io/anthropic/v1/messages",
     format: "claude",
     urlSuffix: "?beta=true",
-    headers: {
-      "Anthropic-Version": "2023-06-01",
-      "Anthropic-Beta": "claude-code-20250219,interleaved-thinking-2025-05-14",
-    },
+    headers: { ...CLAUDE_API_HEADERS },
     quirks: {
       dropOutputConfig: true,
     },
@@ -41,6 +38,21 @@ export default {
       ],
     },
   },
+  // Multi-endpoint: pick the transport matching client sourceFormat to skip translation.
+  transports: [
+    {
+      format: "openai",
+      baseUrl: "https://api.minimax.io/v1/chat/completions",
+      auth: { combined: true, header: "Authorization", scheme: "bearer" },
+    },
+    {
+      format: "claude",
+      baseUrl: "https://api.minimax.io/anthropic/v1/messages",
+      urlSuffix: "?beta=true",
+      headers: { ...CLAUDE_API_HEADERS },
+      auth: { combined: true, header: "x-api-key", scheme: "raw" },
+    },
+  ],
   models: [
     { id: "MiniMax-M3", name: "MiniMax M3", targetFormat: "claude" },
     { id: "MiniMax-M2.7", name: "MiniMax M2.7" },

open-sse/providers/registry/xiaomi-mimo.js

@@ -1,3 +1,5 @@
+import { CLAUDE_API_HEADERS } from "../shared.js";
+
 export default {
   id: "xiaomi-mimo",
   priority: 290,
@@ -21,6 +23,20 @@ export default {
     baseUrl: "https://api.xiaomimimo.com/v1/chat/completions",
     validateUrl: "https://api.xiaomimimo.com/v1/models",
   },
+  // Multi-endpoint: pick the transport matching client sourceFormat to skip translation.
+  transports: [
+    {
+      format: "openai",
+      baseUrl: "https://api.xiaomimimo.com/v1/chat/completions",
+      auth: { combined: true, header: "Authorization", scheme: "bearer" },
+    },
+    {
+      format: "claude",
+      baseUrl: "https://api.xiaomimimo.com/anthropic/v1/messages",
+      headers: { ...CLAUDE_API_HEADERS },
+      auth: { combined: true, header: "x-api-key", scheme: "raw" },
+    },
+  ],
   models: [
     { id: "mimo-v2.5-pro", name: "MiMo V2.5 Pro" },
     { id: "mimo-v2.5", name: "MiMo V2.5" },

open-sse/providers/registry/xiaomi-tokenplan.js

@@ -1,3 +1,5 @@
+import { CLAUDE_API_HEADERS } from "../shared.js";
+
 export default {
   id: "xiaomi-tokenplan",
   priority: 300,
@@ -29,6 +31,19 @@ export default {
     },
     defaultRegion: "sgp",
   },
+  // Multi-endpoint: pick the transport matching client sourceFormat to skip translation.
+  // baseUrl omitted — region-dynamic, resolved in the executor's buildUrl.
+  transports: [
+    {
+      format: "openai",
+      auth: { combined: true, header: "Authorization", scheme: "bearer" },
+    },
+    {
+      format: "claude",
+      headers: { ...CLAUDE_API_HEADERS },
+      auth: { combined: true, header: "x-api-key", scheme: "raw" },
+    },
+  ],
   models: [
     { id: "mimo-v2.5-pro", name: "MiMo V2.5 Pro" },
     { id: "mimo-v2.5-pro-claude", name: "MiMo V2.5 Pro (Claude Native)", targetFormat: "claude", upstreamModelId: "mimo-v2.5-pro" },

open-sse/rtk/caveman.js

@@ -1,100 +1,9 @@
 // Caveman injector: appends a caveman-style instruction into the system message
 // of the final request body, just before it is dispatched to the provider executor.
-// Dispatches by format so it works for both translated and native-passthrough flows.
 
-import { FORMATS } from "../translator/formats.js";
+import { injectSystemPrompt } from "./systemInject.js";
 import { CAVEMAN_PROMPTS } from "./cavemanPrompts.js";
 
-const SEP = "\n\n";
-
 export function injectCaveman(body, format, level) {
-  const prompt = CAVEMAN_PROMPTS[level];
-  if (!body || !prompt) return;
-
-  switch (format) {
-    case FORMATS.CLAUDE:
-      injectClaudeSystem(body, prompt);
-      return;
-    case FORMATS.GEMINI:
-    case FORMATS.GEMINI_CLI:
-    case FORMATS.VERTEX:
-    case FORMATS.ANTIGRAVITY:
-      // Antigravity wraps Gemini shape in body.request → injectGeminiSystem handles it
-      injectGeminiSystem(body, prompt);
-      return;
-    default:
-      // OpenAI and OpenAI-shaped formats (responses/codex/cursor/kiro/ollama)
-      injectMessagesSystem(body, prompt);
-  }
-}
-
-// OpenAI-shaped: messages[] (chat) or input[] (responses) or instructions (responses string)
-function injectMessagesSystem(body, prompt) {
-  // OpenAI Responses API: top-level string field
-  if (typeof body.instructions === "string") {
-    body.instructions = body.instructions
-      ? `${body.instructions}${SEP}${prompt}`
-      : prompt;
-    return;
-  }
-
-  const arr = Array.isArray(body.messages) ? body.messages
-    : Array.isArray(body.input) ? body.input
-    : null;
-  if (!arr) return;
-
-  const idx = arr.findIndex(m => m && (m.role === "system" || m.role === "developer"));
-  if (idx >= 0) {
-    appendToOpenAIMessage(arr[idx], prompt);
-  } else {
-    arr.unshift({ role: "system", content: prompt });
-  }
-}
-
-function appendToOpenAIMessage(msg, prompt) {
-  if (typeof msg.content === "string") {
-    msg.content = `${msg.content}${SEP}${prompt}`;
-  } else if (Array.isArray(msg.content)) {
-    // Responses-style array of parts {type:"input_text"|"text", text}
-    msg.content.push({ type: "input_text", text: prompt });
-  } else {
-    msg.content = prompt;
-  }
-}
-
-// Claude shape: body.system as string | array of {type:"text", text}
-// Insert before the last cache_control block to keep caveman inside the cached prefix.
-function injectClaudeSystem(body, prompt) {
-  if (typeof body.system === "string" && body.system.length > 0) {
-    body.system = `${body.system}${SEP}${prompt}`;
-    return;
-  }
-  if (Array.isArray(body.system)) {
-    const block = { type: "text", text: prompt };
-    let lastCacheIdx = -1;
-    for (let i = body.system.length - 1; i >= 0; i--) {
-      if (body.system[i]?.cache_control) { lastCacheIdx = i; break; }
-    }
-    if (lastCacheIdx >= 0) {
-      body.system.splice(lastCacheIdx, 0, block);
-    } else {
-      body.system.push(block);
-    }
-    return;
-  }
-  body.system = prompt;
-}
-
-// Gemini shape: body.system_instruction | body.systemInstruction | body.request.systemInstruction
-// Each shape: { parts: [{ text }] }
-function injectGeminiSystem(body, prompt) {
-  const target = body.request && typeof body.request === "object" ? body.request : body;
-  const useSnake = Object.prototype.hasOwnProperty.call(target, "system_instruction");
-  const key = useSnake ? "system_instruction" : "systemInstruction";
-  const sys = target[key];
-  if (sys && Array.isArray(sys.parts)) {
-    sys.parts.push({ text: prompt });
-    return;
-  }
-  target[key] = { parts: [{ text: prompt }] };
+  injectSystemPrompt(body, format, CAVEMAN_PROMPTS[level]);
 }

open-sse/rtk/headroom.js

@@ -0,0 +1,63 @@
+import { claudeToOpenAIRequest } from "../translator/request/claude-to-openai.js";
+import { openaiToClaudeRequest } from "../translator/request/openai-to-claude.js";
+
+const DEFAULT_TIMEOUT_MS = 3000;
+
+// POST messages to Headroom /v1/compress; returns compressed messages + stats or null.
+async function callCompress(url, messages, model, timeoutMs, compressUserMessages) {
+  const endpoint = `${String(url).replace(/\/$/, "")}/v1/compress`;
+  const payload = { messages, model };
+  if (compressUserMessages) payload.config = { compress_user_messages: true };
+  const res = await fetch(endpoint, {
+    method: "POST",
+    headers: { "Content-Type": "application/json" },
+    body: JSON.stringify(payload),
+    signal: AbortSignal.timeout(timeoutMs),
+  });
+  if (!res.ok) return null;
+  const data = await res.json();
+  if (!Array.isArray(data?.messages)) return null;
+  return data;
+}
+
+// Compress request body via Headroom proxy. Fail-open: returns null on any error.
+// /v1/compress only understands OpenAI shape, so Claude bodies are translated
+// to OpenAI, compressed, then translated back using 9Router's own translators.
+export async function compressWithHeadroom(body, { enabled, url, model, format, compressUserMessages, timeoutMs = DEFAULT_TIMEOUT_MS } = {}) {
+  if (!enabled || !url || !body) return null;
+
+  try {
+    // Claude shape: translate → OpenAI → compress → translate back.
+    if (format === "claude") {
+      const oai = claudeToOpenAIRequest(model, body, false);
+      if (!Array.isArray(oai?.messages)) return null;
+      const data = await callCompress(url, oai.messages, model, timeoutMs, compressUserMessages);
+      if (!data) return null;
+      const claudeBody = openaiToClaudeRequest(model, { ...oai, messages: data.messages }, false);
+      if (Array.isArray(claudeBody?.messages)) body.messages = claudeBody.messages;
+      if (claudeBody?.system !== undefined) body.system = claudeBody.system;
+      return data;
+    }
+
+    // OpenAI shape: messages/input go straight to the proxy.
+    const key = Array.isArray(body.messages) ? "messages"
+      : Array.isArray(body.input) ? "input"
+      : null;
+    if (!key) return null;
+    const data = await callCompress(url, body[key], model, timeoutMs, compressUserMessages);
+    if (!data) return null;
+    body[key] = data.messages;
+    return data;
+  } catch {
+    return null;
+  }
+}
+
+export function formatHeadroomLog(stats) {
+  if (!stats) return null;
+  const before = stats.tokens_before || 0;
+  const after = stats.tokens_after || 0;
+  const saved = stats.tokens_saved || 0;
+  const pct = before > 0 ? ((saved / before) * 100).toFixed(1) : "0";
+  return `saved ${saved} tokens / ${before} (${pct}%) ${after ? `after=${after}` : ""}`.trim();
+}

open-sse/rtk/ponytail.js

@@ -0,0 +1,9 @@
+// Ponytail injector: appends the "lazy senior dev" instruction into the system
+// message of the final request body, just before dispatch to the provider executor.
+
+import { injectSystemPrompt } from "./systemInject.js";
+import { PONYTAIL_PROMPTS } from "./ponytailPrompt.js";
+
+export function injectPonytail(body, format, level) {
+  injectSystemPrompt(body, format, PONYTAIL_PROMPTS[level]);
+}

open-sse/rtk/ponytailPrompt.js

@@ -0,0 +1,52 @@
+// Ponytail intensity-level prompts injected into system message to bias toward minimal code.
+// Adapted from ponytail skill (https://github.com/DietrichGebert/ponytail).
+
+export const PONYTAIL_LEVELS = {
+  LITE: "lite",
+  FULL: "full",
+  ULTRA: "ultra",
+};
+
+const SHARED_PERSONA = "You are a lazy senior developer. Lazy means efficient, not careless. The best code is the code never written.";
+
+const SHARED_LADDER = "Before writing code, stop at the first rung that holds: 1) Does this need to exist at all? (YAGNI) 2) Stdlib does it? Use it. 3) Native platform feature covers it? Use it (CSS over JS, DB constraint over app code). 4) Already-installed dependency solves it? Use it; never add a new one for what a few lines can do. 5) Can it be one line? One line. 6) Only then: the minimum code that works.";
+
+const SHARED_RULES = "No unrequested abstractions (no interface with one implementation, no factory for one product, no config for a value that never changes). No boilerplate or scaffolding \"for later\". Deletion over addition. Boring over clever. Fewest files possible; shortest working diff wins. Two stdlib options the same size: take the edge-case-correct one. Mark deliberate simplifications with a `ponytail:` comment naming the ceiling and upgrade path.";
+
+const SHARED_OUTPUT = "Code first. Then at most three short lines: what was skipped, when to add it. No essays or design notes. Pattern: `[code] → skipped: [X], add when [Y].`";
+
+const SHARED_NOT_LAZY = "Never simplify away: input validation at trust boundaries, error handling that prevents data loss, security, accessibility, anything explicitly requested. Non-trivial logic leaves ONE runnable check behind (an assert-based self-check or one small test file; no frameworks). Trivial one-liners need no test.";
+
+const SHARED_PERSISTENCE = "ACTIVE EVERY RESPONSE. No drift back to over-building. Still active if unsure.";
+
+export const PONYTAIL_PROMPTS = {
+  [PONYTAIL_LEVELS.LITE]: [
+    SHARED_PERSONA,
+    "Lite: build what's asked, but name the lazier alternative in one line. User picks.",
+    SHARED_LADDER,
+    SHARED_RULES,
+    SHARED_OUTPUT,
+    SHARED_NOT_LAZY,
+    SHARED_PERSISTENCE,
+  ].join(" "),
+
+  [PONYTAIL_LEVELS.FULL]: [
+    SHARED_PERSONA,
+    "Full: the ladder enforced. Stdlib and native first. Shortest diff, shortest explanation.",
+    SHARED_LADDER,
+    SHARED_RULES,
+    SHARED_OUTPUT,
+    SHARED_NOT_LAZY,
+    SHARED_PERSISTENCE,
+  ].join(" "),
+
+  [PONYTAIL_LEVELS.ULTRA]: [
+    SHARED_PERSONA,
+    "Ultra: YAGNI extremist. Deletion before addition. Ship the one-liner and challenge the rest of the requirement in the same response.",
+    SHARED_LADDER,
+    SHARED_RULES,
+    SHARED_OUTPUT,
+    SHARED_NOT_LAZY,
+    SHARED_PERSISTENCE,
+  ].join(" "),
+};

open-sse/rtk/systemInject.js

@@ -0,0 +1,98 @@
+// Shared system-prompt injector: appends an instruction into the system message of
+// the final request body, dispatching by format so it works for translated and
+// native-passthrough flows. Used by caveman.js and ponytail.js.
+
+import { FORMATS } from "../translator/formats.js";
+
+const SEP = "\n\n";
+
+export function injectSystemPrompt(body, format, prompt) {
+  if (!body || !prompt) return;
+
+  switch (format) {
+    case FORMATS.CLAUDE:
+      injectClaudeSystem(body, prompt);
+      return;
+    case FORMATS.GEMINI:
+    case FORMATS.GEMINI_CLI:
+    case FORMATS.VERTEX:
+    case FORMATS.ANTIGRAVITY:
+      // Antigravity wraps Gemini shape in body.request → injectGeminiSystem handles it
+      injectGeminiSystem(body, prompt);
+      return;
+    default:
+      // OpenAI and OpenAI-shaped formats (responses/codex/cursor/kiro/ollama)
+      injectMessagesSystem(body, prompt);
+  }
+}
+
+// OpenAI-shaped: messages[] (chat) or input[] (responses) or instructions (responses string)
+function injectMessagesSystem(body, prompt) {
+  // OpenAI Responses API: top-level string field
+  if (typeof body.instructions === "string") {
+    body.instructions = body.instructions
+      ? `${body.instructions}${SEP}${prompt}`
+      : prompt;
+    return;
+  }
+
+  const arr = Array.isArray(body.messages) ? body.messages
+    : Array.isArray(body.input) ? body.input
+    : null;
+  if (!arr) return;
+
+  const idx = arr.findIndex(m => m && (m.role === "system" || m.role === "developer"));
+  if (idx >= 0) {
+    appendToOpenAIMessage(arr[idx], prompt);
+  } else {
+    arr.unshift({ role: "system", content: prompt });
+  }
+}
+
+function appendToOpenAIMessage(msg, prompt) {
+  if (typeof msg.content === "string") {
+    msg.content = `${msg.content}${SEP}${prompt}`;
+  } else if (Array.isArray(msg.content)) {
+    // Responses-style array of parts {type:"input_text"|"text", text}
+    msg.content.push({ type: "input_text", text: prompt });
+  } else {
+    msg.content = prompt;
+  }
+}
+
+// Claude shape: body.system as string | array of {type:"text", text}
+// Insert before the last cache_control block to keep injection inside the cached prefix.
+function injectClaudeSystem(body, prompt) {
+  if (typeof body.system === "string" && body.system.length > 0) {
+    body.system = `${body.system}${SEP}${prompt}`;
+    return;
+  }
+  if (Array.isArray(body.system)) {
+    const block = { type: "text", text: prompt };
+    let lastCacheIdx = -1;
+    for (let i = body.system.length - 1; i >= 0; i--) {
+      if (body.system[i]?.cache_control) { lastCacheIdx = i; break; }
+    }
+    if (lastCacheIdx >= 0) {
+      body.system.splice(lastCacheIdx, 0, block);
+    } else {
+      body.system.push(block);
+    }
+    return;
+  }
+  body.system = prompt;
+}
+
+// Gemini shape: body.system_instruction | body.systemInstruction | body.request.systemInstruction
+// Each shape: { parts: [{ text }] }
+function injectGeminiSystem(body, prompt) {
+  const target = body.request && typeof body.request === "object" ? body.request : body;
+  const useSnake = Object.prototype.hasOwnProperty.call(target, "system_instruction");
+  const key = useSnake ? "system_instruction" : "systemInstruction";
+  const sys = target[key];
+  if (sys && Array.isArray(sys.parts)) {
+    sys.parts.push({ text: prompt });
+    return;
+  }
+  target[key] = { parts: [{ text: prompt }] };
+}

open-sse/services/provider.js

@@ -136,6 +136,16 @@ export function getTargetFormat(provider) {
   return config.format || "openai";
 }
 
+// Resolve which transport to use for a provider given the client sourceFormat.
+// Multi-endpoint providers (transport.transports[]) pick the entry matching sourceFormat
+// to avoid lossy translation; falls back to the default transport when no match.
+export function resolveTransport(provider, sourceFormat) {
+  const config = PROVIDERS[provider];
+  const transports = config?.transports;
+  if (!Array.isArray(transports) || !transports.length) return null;
+  return transports.find(t => t.format === sourceFormat) || null;
+}
+
 // Check if last message is from user
 export function isLastMessageFromUser(body) {
   const messages = body.messages || body.contents;