feat(codex): self-contained codex app-server transport (executor + provider + sign-in) (#11205)

Merged after conflict resolution: the 5 conflicting test files were the base-red drains that #11201 already landed on the tip — kept the tip versions; the feature content is untouched. Validated on the combined batch board + this branch: codex-app-server + codex-gpt56-catalog 25/25, typecheck:core clean, docs-counts green (351 providers), provider-consistency 268/351/0. The opt-in codex-app-server transport (JSON-RPC-over-WS, turn/completed-awaited close, Responses SSE bridge) leaves the default codex path untouched. Thank you @arminanton — a 3.4k-line transport with the docs wave and tests to match!

.env.example

@@ -1303,6 +1303,30 @@ CURSOR_USER_AGENT="Cursor/3.4"
 # set to true/1/yes to enable. Used by: open-sse/executors/codex.ts.
 # OMNIROUTE_CODEX_DROP_NONSTANDARD_EVENTS=true
 
+# Codex app-server WebSocket transport (opt-in). When a WebSocket URL and a
+# capability token are both provided, Codex requests are routed through a local
+# `codex app-server` sidecar over JSON-RPC instead of the HTTP Responses API.
+# Each var is also settable per-connection via providerSpecificData; the env var
+# is the process-wide fallback. Used by:
+# open-sse/executors/codex/appServerConfig.ts.
+#
+# WebSocket endpoint of the codex app-server (ws:// or wss://). Required to
+# enable the transport; leaving it unset keeps Codex on its HTTP transports.
+# OMNIROUTE_CODEX_APPSERVER_WS=ws://127.0.0.1:8081
+# Inline capability/bearer token presented to the app-server.
+# OMNIROUTE_CODEX_APPSERVER_WS_TOKEN=deadbeef...
+# Path to a file holding the capability token (produced by
+# `codex app-server --ws-token-file <path>`). Used when the inline token above
+# is not set.
+# OMNIROUTE_CODEX_APPSERVER_WS_TOKEN_FILE=/run/codex-ws-token
+# Working directory the app-server turn runs in (defaults to /tmp).
+# OMNIROUTE_CODEX_APPSERVER_CWD=/tmp
+# Approval policy passed to the app-server turn (e.g. never, on-request).
+# OMNIROUTE_CODEX_APPSERVER_APPROVAL=never
+# Sandbox policy passed to the app-server turn (e.g. read-only,
+# workspace-write, danger-full-access).
+# OMNIROUTE_CODEX_APPSERVER_SANDBOX=read-only
+
 # ═══════════════════════════════════════════════════════════════════════════════
 # 13. CLI FINGERPRINT COMPATIBILITY (Anti-Detection)
 # ═══════════════════════════════════════════════════════════════════════════════

.gitleaks.toml

@@ -92,5 +92,9 @@
       # - x-api-key PUBLICO do Firefly web (documentado em open-sse/utils/publicCreds.ts:207);
       #   as duas ocorrencias sinalizadas estao em COMENTARIOS JSDoc, o runtime le de resolvePublicCred().
       '''omniroute-kimi-sponsor-banner-dismissed-v\d+''',
+      # CheaperInference sponsor banner localStorage key (upstream #11196 /
+      # eb5797370). Same UI-identifier pattern as the kimi banner above, not a
+      # credential; the generic-api-key rule flags the long hyphenated string.
+      '''omniroute-cheaperinference-sponsor-banner-dismissed-v\d+''',
       '''SunbreakWebUI1''',
     ]

AGENTS.md

@@ -46,7 +46,7 @@ Repository map and Reference Documentation sections below.
 
 ## Project at a Glance
 
-**OmniRoute** — unified AI proxy/router. One endpoint, 350 LLM providers, auto-fallback.
+**OmniRoute** — unified AI proxy/router. One endpoint, 351 LLM providers, auto-fallback.
 
 | Layer         | Location                | Purpose                                                                                                                                                                   |
 | ------------- | ----------------------- | ------------------------------------------------------------------------------------------------------------------------------------------------------------------------- |

README.md

@@ -7,7 +7,7 @@
 
 # 🚀 OmniRoute — The Free AI Gateway
 
-<img src="./docs/diagrams/readme-hero.svg" width="100%" alt="OmniRoute — Never stop coding. Every AI tool → 349 providers — 90+ free — through one endpoint. Claude Code, Codex, Cursor, Cline, Copilot & Antigravity into FREE Claude / GPT / Gemini with auto-fallback. RTK + Caveman stacked compression saves 15–95% tokens (~89% avg) — never hit limits. 350 AI providers · 90+ free tiers · ~1.51B free tokens/mo · 19 routing strategies · $0 to start."/>
+<img src="./docs/diagrams/readme-hero.svg" width="100%" alt="OmniRoute — Never stop coding. Every AI tool → 351 providers — 90+ free — through one endpoint. Claude Code, Codex, Cursor, Cline, Copilot & Antigravity into FREE Claude / GPT / Gemini with auto-fallback. RTK + Caveman stacked compression saves 15–95% tokens (~89% avg) — never hit limits. 351 AI providers · 90+ free tiers · ~1.51B free tokens/mo · 19 routing strategies · $0 to start."/>
 
 </div>
 
@@ -210,7 +210,7 @@ curl http://localhost:20128/v1/chat/completions \
 
 </div>
 
-<img src="./docs/diagrams/promise-pillars.svg" width="100%" alt="The Promise — One endpoint. 349 providers. Never stop building — OmniRoute picks the cheapest one that works. Six pillars: Never hit limits (auto-fallback across 349 providers in milliseconds, zero downtime) · Save up to 95% tokens (RTK + Caveman stacked compression cuts 15–95%, ~89% avg on tool-heavy sessions) · $0 to start (90+ free tiers, 56 free forever — no card needed) · Every tool works (33 coding agents through one config) · One endpoint (OpenAI ↔ Claude ↔ Gemini ↔ Responses API at /v1) · Production-grade (circuit breakers, TLS stealth, MCP 110 tools, A2A, memory, guardrails, evals — 25,000+ tests)."/>
+<img src="./docs/diagrams/promise-pillars.svg" width="100%" alt="The Promise — One endpoint. 351 providers. Never stop building — OmniRoute picks the cheapest one that works. Six pillars: Never hit limits (auto-fallback across 351 providers in milliseconds, zero downtime) · Save up to 95% tokens (RTK + Caveman stacked compression cuts 15–95%, ~89% avg on tool-heavy sessions) · $0 to start (90+ free tiers, 56 free forever — no card needed) · Every tool works (33 coding agents through one config) · One endpoint (OpenAI ↔ Claude ↔ Gemini ↔ Responses API at /v1) · Production-grade (circuit breakers, TLS stealth, MCP 110 tools, A2A, memory, guardrails, evals — 25,000+ tests)."/>
 
 <br/>
 <br/>
@@ -461,7 +461,7 @@ All **19** strategies — mix & match per combo step:
 
 </div>
 
-<img src="./docs/diagrams/comparison-table.svg" width="100%" alt="What sets OmniRoute apart — comparison table vs 9router, OpenRouter, CLIProxyAPI and LiteLLM across 13 capabilities. OmniRoute: 349 providers, 90+ free providers built-in, 19 routing strategies, 12-engine token compression, built-in MCP server with 110 tools, A2A agent protocol, persistent memory, guardrails, cloud agents, TLS fingerprint stealth, Desktop/Termux/PWA, 43 i18n UI locales, 100% MIT self-hosted. OmniRoute is the only one with the full set; competitors show a mix of checks, partials and crosses. Verified from each project&apos;s docs."/>
+<img src="./docs/diagrams/comparison-table.svg" width="100%" alt="What sets OmniRoute apart — comparison table vs 9router, OpenRouter, CLIProxyAPI and LiteLLM across 13 capabilities. OmniRoute: 351 providers, 90+ free providers built-in, 19 routing strategies, 12-engine token compression, built-in MCP server with 110 tools, A2A agent protocol, persistent memory, guardrails, cloud agents, TLS fingerprint stealth, Desktop/Termux/PWA, 43 i18n UI locales, 100% MIT self-hosted. OmniRoute is the only one with the full set; competitors show a mix of checks, partials and crosses. Verified from each project&apos;s docs."/>
 
 <sub>📊 Full methodology &amp; per-feature detail vs 9router, OpenRouter, CLIProxyAPI &amp; LiteLLM → [`docs/comparison/OMNIROUTE_VS_ALTERNATIVES.md`](docs/comparison/OMNIROUTE_VS_ALTERNATIVES.md)</sub>
 
@@ -646,7 +646,7 @@ of your shell history. → [CLI Integrations](docs/guides/CLI-INTEGRATIONS.md)
 
 </div>
 
-> The most complete catalog of any open-source router: **349 providers**, **90+ with a free tier**, **56 free forever**.
+> The most complete catalog of any open-source router: **351 providers**, **90+ with a free tier**, **56 free forever**.
 
 <div align="center">
 

docker-compose.yml

@@ -43,8 +43,21 @@ x-common: &common
     - LIVE_WS_ALLOWED_ORIGINS=${LIVE_WS_ALLOWED_ORIGINS:-http://localhost:20128,http://127.0.0.1:20128}
     - REDIS_URL=${REDIS_URL:-redis://redis:6379}
     - NODE_OPTIONS=--max-old-space-size=2048
+    # Codex App-Server transport (provider: codex-app-server). Inert unless the
+    # `codex-app-server` compose profile is up (the sidecar below). Points the app
+    # at the internal sidecar; the capability token is shared via the mounted file.
+    - OMNIROUTE_CODEX_APPSERVER_WS=${OMNIROUTE_CODEX_APPSERVER_WS:-ws://codex-app-server:1456}
+    - OMNIROUTE_CODEX_APPSERVER_WS_TOKEN_FILE=${OMNIROUTE_CODEX_APPSERVER_WS_TOKEN_FILE:-/run/codex-appserver/token}
   volumes:
     - ./data:/app/data
+    # Shared capability token + codex auth for the app-server WS. Only meaningful
+    # when the codex-app-server profile is active. The token dir carries the WS
+    # capability token; the codex home is where the dashboard "Apply auth" writes
+    # ~/.codex/auth.json (getCliConfigPaths("codex") = <home>/.codex; the base
+    # image runs as `node`, so /home/node/.codex) and the SAME volume is mounted
+    # into the sidecar so its `codex app-server` reads the same auth.
+    - codex-appserver-token:/run/codex-appserver
+    - codex-appserver-home:/home/node/.codex
   healthcheck:
     test: ["CMD", "node", "healthcheck.mjs"]
     interval: 30s
@@ -290,6 +303,59 @@ services:
     profiles:
       - cliproxyapi
 
+  # ── Profile: codex-app-server (Codex CLI app-server sidecar) ──────────
+  # A PLAIN Codex app-server for the `codex-app-server` provider: OmniRoute drives
+  # the Codex CLI's own `codex app-server` over JSON-RPC/WebSocket instead of
+  # replaying a session token to the API. It listens ONLY on the internal compose
+  # network (ws://codex-app-server:1456), guarded by a capability token — it is
+  # NEVER published to the host / internet. The Codex CLI (baked into
+  # omniroute:base) self-manages its OpenAI OAuth via the shared ~/.codex volume,
+  # which the dashboard "Apply auth" (device-OAuth) writes and this sidecar reads.
+  #
+  # NOTE: this is the GENERIC public sidecar. An operator wanting residential /
+  # UDP egress (via a TUN sidecar) runs that separately as an override; it is
+  # intentionally not shipped here.
+  codex-app-server:
+    image: omniroute:base
+    container_name: omniroute-codex-app-server
+    restart: unless-stopped
+    # Generate the WS capability token on first boot if absent, then run the
+    # app-server. entrypoint is overridden because the base image's default is the
+    # Next.js server.
+    entrypoint: ["/bin/sh", "-c"]
+    command:
+      - |
+        set -e
+        TOKEN_FILE=/run/codex-appserver/token
+        mkdir -p /run/codex-appserver
+        if [ ! -s "$$TOKEN_FILE" ]; then
+          # 32-byte hex capability token; shared with the app via the token volume.
+          TF="$$TOKEN_FILE" node -e 'require("fs").writeFileSync(process.env.TF, require("crypto").randomBytes(32).toString("hex"))' 2>/dev/null || \
+            { head -c 32 /dev/urandom | od -An -tx1 | tr -d ' \n' > "$$TOKEN_FILE"; }
+          chmod 600 "$$TOKEN_FILE"
+        fi
+        exec codex app-server \
+          --listen ws://0.0.0.0:1456 \
+          --ws-auth capability-token \
+          --ws-token-file "$$TOKEN_FILE"
+    environment:
+      - CODEX_HOME=/home/node/.codex
+      - RUST_LOG=${CODEX_APPSERVER_RUST_LOG:-warn}
+    volumes:
+      - codex-appserver-token:/run/codex-appserver
+      - codex-appserver-home:/home/node/.codex
+    # No `ports:` — internal-only. Reached at ws://codex-app-server:1456 over the
+    # compose network by the omniroute app.
+    healthcheck:
+      test:
+        ["CMD", "node", "-e", "require('http').get('http://127.0.0.1:1456/readyz',r=>process.exit(r.statusCode===200?0:1)).on('error',()=>process.exit(1))"]
+      interval: 30s
+      timeout: 5s
+      retries: 3
+      start_period: 10s
+    profiles:
+      - codex-app-server
+
 volumes:
   chatgpt-web-codex-browser-data:
     name: omniroute-chatgpt-web-codex-browser-data
@@ -301,3 +367,7 @@ volumes:
     name: omniroute-qdrant-data
   bifrost-data:
     name: omniroute-bifrost-data
+  codex-appserver-token:
+    name: omniroute-codex-appserver-token
+  codex-appserver-home:
+    name: omniroute-codex-appserver-home

docs/architecture/ARCHITECTURE.md

@@ -17,7 +17,7 @@ It provides a single OpenAI-compatible endpoint (`/v1/*`) and routes traffic acr
 
 Core capabilities:
 
-- OpenAI-compatible API surface for CLI/tools (338 providers, 100 executors)
+- OpenAI-compatible API surface for CLI/tools (349 providers, 107 executors)
 - Request/response translation across provider formats
 - Model combo fallback (multi-model sequence)
 - Structured combo steps (`provider + model + connection`) with runtime ordering by `compositeTiers`

docs/architecture/CODEBASE_DOCUMENTATION.md

@@ -451,7 +451,7 @@ open-sse/
 ├── types.d.ts
 ├── config/                 Provider registries, header profiles, identity, …
 ├── handlers/               Request handlers (chat, embeddings, audio, image, …)
-├── executors/              101 provider-specific HTTP executors
+├── executors/              107 provider-specific HTTP executors
 ├── translator/             Format conversion (OpenAI ↔ Claude ↔ Gemini ↔ Cursor ↔ Kiro)
 ├── transformer/            Responses API ↔ Chat Completions stream transformer
 ├── services/               80+ service modules (combos, fallback, quotas, identity, …)

docs/diagrams/cli-terminal.svg

@@ -1,4 +1,4 @@
-<svg viewBox="0 0 1200 350" xmlns="http://www.w3.org/2000/svg" role="img" aria-label="Animated terminal demoing the OmniRoute CLI: omniroute providers list (350 providers registered, anthropic, codex, glm, kimi shown active), omniroute combo list (always-on priority, cost-saver, fusion-panel, context-relay) and omniroute health (healthy, 18412 requests in 24h, p95 412ms, circuit breakers 24 closed, 1 half-open, 0 open), cycling over the 80+ command surface: providers, oauth, keys, combo, nodes, models, cache, compression, cost, usage, quota, health, resilience, telemetry, logs, audit, mcp, a2a, cloud, memory, skills, eval, doctor, repl, tunnel, backup, sync, webhooks, policy, pricing, translator, simulate and more.">
+<svg viewBox="0 0 1200 350" xmlns="http://www.w3.org/2000/svg" role="img" aria-label="Animated terminal demoing the OmniRoute CLI: omniroute providers list (351 providers registered, anthropic, codex, glm, kimi shown active), omniroute combo list (always-on priority, cost-saver, fusion-panel, context-relay) and omniroute health (healthy, 18412 requests in 24h, p95 412ms, circuit breakers 24 closed, 1 half-open, 0 open), cycling over the 80+ command surface: providers, oauth, keys, combo, nodes, models, cache, compression, cost, usage, quota, health, resilience, telemetry, logs, audit, mcp, a2a, cloud, memory, skills, eval, doctor, repl, tunnel, backup, sync, webhooks, policy, pricing, translator, simulate and more.">
 <desc>Compact animated terminal cycling three real OmniRoute CLI commands with a typewriter effect and a scrolling subcommand ticker; the first frame shows the completed providers-list screen.</desc>
 <defs><clipPath id="tickerClip"><rect x="12" y="304" width="1176" height="40"/></clipPath><clipPath id="tw0"><rect x="64" y="46" height="26" width="0"><animate attributeName="width" calcMode="discrete" values="0;31;61;92;122;153;184;214;245;245" keyTimes="0;0.012;0.018;0.024;0.030;0.036;0.042;0.048;0.054;1" dur="18s" repeatCount="indefinite"/></rect></clipPath><clipPath id="tw1"><rect x="64" y="46" height="26" width="0"><animate attributeName="width" calcMode="discrete" values="0;26;51;76;102;128;153;178;204;204" keyTimes="0;0.348;0.351;0.357;0.363;0.369;0.375;0.381;0.387;1" dur="18s" repeatCount="indefinite"/></rect></clipPath><clipPath id="tw2"><rect x="64" y="46" height="26" width="0"><animate attributeName="width" calcMode="discrete" values="0;20;41;61;82;102;122;143;163;163" keyTimes="0;0.678;0.684;0.690;0.696;0.702;0.708;0.714;0.720;1" dur="18s" repeatCount="indefinite"/></rect></clipPath></defs>
 <rect width="1200" height="350" fill="#0d1117"/>

docs/diagrams/comparison-table.svg

@@ -1,4 +1,4 @@
-<svg viewBox="0 0 1200 780" xmlns="http://www.w3.org/2000/svg" role="img" aria-label="Comparison table: OmniRoute versus 9router, OpenRouter, CLIProxyAPI and LiteLLM across 13 capabilities. OmniRoute is the only one with the full set: 350 providers, 90+ free providers built-in, 19 routing strategies, 12-engine token compression, a built-in MCP server with 110 tools, A2A protocol, persistent memory, guardrails, cloud agents, TLS fingerprint stealth, desktop/Termux/PWA, 43 UI locales and 100% MIT self-hosted. 9router has free providers, RTK compression and translation but no MCP, A2A, memory, guardrails, cloud agents or stealth. OpenRouter is a hosted SaaS with 400+ models, guardrails and a hosted MCP but is not self-hosted and lacks A2A, memory, cloud agents and stealth. CLIProxyAPI is a light OAuth proxy with two routing strategies. LiteLLM has 100+ providers, A2A and extensive guardrails but no memory, compression, free tier, stealth or cloud agents.">
+<svg viewBox="0 0 1200 780" xmlns="http://www.w3.org/2000/svg" role="img" aria-label="Comparison table: OmniRoute versus 9router, OpenRouter, CLIProxyAPI and LiteLLM across 13 capabilities. OmniRoute is the only one with the full set: 351 providers, 90+ free providers built-in, 19 routing strategies, 12-engine token compression, a built-in MCP server with 110 tools, A2A protocol, persistent memory, guardrails, cloud agents, TLS fingerprint stealth, desktop/Termux/PWA, 43 UI locales and 100% MIT self-hosted. 9router has free providers, RTK compression and translation but no MCP, A2A, memory, guardrails, cloud agents or stealth. OpenRouter is a hosted SaaS with 400+ models, guardrails and a hosted MCP but is not self-hosted and lacks A2A, memory, cloud agents and stealth. CLIProxyAPI is a light OAuth proxy with two routing strategies. LiteLLM has 100+ providers, A2A and extensive guardrails but no memory, compression, free tier, stealth or cloud agents.">
   <desc>Static-header comparison table where each capability row fades in top to bottom; the OmniRoute column is highlighted and shows a check or a leading value in every row, while competitors show a mix of checks, partials and crosses.</desc>
   <defs>
     <pattern id="gC" width="32" height="32" patternUnits="userSpaceOnUse"><path d="M 32 0 L 0 0 0 32" fill="none" stroke="#ffffff" stroke-opacity="0.05" stroke-width="1"/></pattern>

docs/diagrams/promise-pillars.svg

@@ -1,4 +1,4 @@
-<svg viewBox="0 0 1200 540" xmlns="http://www.w3.org/2000/svg" role="img" aria-label="The OmniRoute promise: one endpoint, 350 providers — never stop building, OmniRoute picks the cheapest one that works. Six pillars. Never hit limits: auto-fallback across 350 providers in milliseconds, quota out means the next provider takes over with zero downtime. Save up to 95 percent of tokens: RTK plus Caveman stacked compression cuts 15 to 95 percent of eligible tokens, about 89 percent average on tool-heavy sessions. Zero dollars to start: 90+ providers with a free tier, 56 free forever — Qoder, Pollinations, Cloudflare, SiliconFlow — no card needed. Every tool works: 33 coding agents including Claude Code, Codex, Cursor, Cline, Copilot and Antigravity through one config. One endpoint: OpenAI, Claude, Gemini and Responses API translation — point any tool at /v1 and it just works. Production-grade: circuit breakers, TLS stealth, MCP with 110 tools, A2A, memory, guardrails, evals — 25,000+ tests.">
+<svg viewBox="0 0 1200 540" xmlns="http://www.w3.org/2000/svg" role="img" aria-label="The OmniRoute promise: one endpoint, 351 providers — never stop building, OmniRoute picks the cheapest one that works. Six pillars. Never hit limits: auto-fallback across 351 providers in milliseconds, quota out means the next provider takes over with zero downtime. Save up to 95 percent of tokens: RTK plus Caveman stacked compression cuts 15 to 95 percent of eligible tokens, about 89 percent average on tool-heavy sessions. Zero dollars to start: 90+ providers with a free tier, 56 free forever — Qoder, Pollinations, Cloudflare, SiliconFlow — no card needed. Every tool works: 33 coding agents including Claude Code, Codex, Cursor, Cline, Copilot and Antigravity through one config. One endpoint: OpenAI, Claude, Gemini and Responses API translation — point any tool at /v1 and it just works. Production-grade: circuit breakers, TLS stealth, MCP with 110 tools, A2A, memory, guardrails, evals — 25,000+ tests.">
   <desc>Animated promise card: six pillar tiles fade in in reading order, then a soft colored border highlight sweeps from tile to tile in a continuous cycle.</desc>
   <defs>
     <pattern id="gridPaperP" width="32" height="32" patternUnits="userSpaceOnUse">
@@ -21,7 +21,7 @@
     <line x1="150" y1="53" x2="1160" y2="53" stroke="#232b38" stroke-width="1.5"/>
   </g>
   <g>
-    <text x="40" y="100" font-family="Inter, 'Segoe UI', Arial, Helvetica, system-ui, sans-serif" font-size="23" font-weight="600" fill="#c9d1d9">One endpoint. <tspan fill="#a78bfa" font-weight="800">350 providers.</tspan> Never stop building — OmniRoute picks <tspan fill="#7ee787" font-weight="700">the cheapest one that works</tspan>.</text>
+    <text x="40" y="100" font-family="Inter, 'Segoe UI', Arial, Helvetica, system-ui, sans-serif" font-size="23" font-weight="600" fill="#c9d1d9">One endpoint. <tspan fill="#a78bfa" font-weight="800">351 providers.</tspan> Never stop building — OmniRoute picks <tspan fill="#7ee787" font-weight="700">the cheapest one that works</tspan>.</text>
   </g>
 
   <g font-family="Inter, 'Segoe UI', Arial, Helvetica, system-ui, sans-serif">
@@ -38,7 +38,7 @@
         <line x1="3.9" y1="3.9" x2="18.1" y2="18.1"/>
       </g>
       <text x="102" y="170" font-size="18" font-weight="800" fill="#74b9ff">Never hit limits</text>
-      <text x="66" y="204" font-size="13.5" fill="#a1a1aa">Auto-fallback across 350 providers in</text>
+      <text x="66" y="204" font-size="13.5" fill="#a1a1aa">Auto-fallback across 351 providers in</text>
       <text x="66" y="226" font-size="13.5" fill="#a1a1aa">milliseconds. Quota out? The next provider</text>
       <text x="66" y="248" font-size="13.5" fill="#a1a1aa">takes over — zero downtime.</text>
     </g>

docs/diagrams/readme-hero.svg

@@ -1,4 +1,4 @@
-<svg viewBox="0 0 1200 548" xmlns="http://www.w3.org/2000/svg" role="img" aria-label="OmniRoute hero: Never stop coding. Every AI tool to 350 providers — 90+ free — through one endpoint. Claude Code, Codex, Cursor, Cline, Copilot and Antigravity into free Claude, GPT and Gemini with auto-fallback. RTK + Caveman stacked compression saves 15 to 95 percent of tokens — about 89 percent average on tool-heavy sessions — so you never hit limits. Stats: 350 AI providers, 90+ free tiers, about 1.51B free tokens per month, 15 to 95 percent token savings, 19 routing strategies, zero dollars to start.">
+<svg viewBox="0 0 1200 548" xmlns="http://www.w3.org/2000/svg" role="img" aria-label="OmniRoute hero: Never stop coding. Every AI tool to 351 providers — 90+ free — through one endpoint. Claude Code, Codex, Cursor, Cline, Copilot and Antigravity into free Claude, GPT and Gemini with auto-fallback. RTK + Caveman stacked compression saves 15 to 95 percent of tokens — about 89 percent average on tool-heavy sessions — so you never hit limits. Stats: 351 AI providers, 90+ free tiers, about 1.51B free tokens per month, 15 to 95 percent token savings, 19 routing strategies, zero dollars to start.">
   <desc>Animated hero card: a pulse travels the divider line and a compression bar demo repeatedly shrinks a prompt by up to 95 percent; all headline content is static and readable on the first frame.</desc>
   <defs>
     <pattern id="gridPaperH" width="32" height="32" patternUnits="userSpaceOnUse">
@@ -28,7 +28,7 @@
   <text x="48" y="138" font-family="Inter, 'Segoe UI', Arial, Helvetica, system-ui, sans-serif" font-size="60" font-weight="800" fill="#e9edf3">Never stop coding<tspan fill="#a855f7">.</tspan></text>
 
   <!-- subheadline -->
-  <text x="48" y="184" font-family="Inter, 'Segoe UI', Arial, Helvetica, system-ui, sans-serif" font-size="25" font-weight="600" fill="#c9d1d9">Every AI tool → <tspan fill="#a78bfa" font-weight="800">350 providers</tspan> — <tspan fill="#7ee787" font-weight="800">90+ free</tspan> — through one endpoint.</text>
+  <text x="48" y="184" font-family="Inter, 'Segoe UI', Arial, Helvetica, system-ui, sans-serif" font-size="25" font-weight="600" fill="#c9d1d9">Every AI tool → <tspan fill="#a78bfa" font-weight="800">351 providers</tspan> — <tspan fill="#7ee787" font-weight="800">90+ free</tspan> — through one endpoint.</text>
 
   <!-- plug line -->
   <text x="48" y="222" font-family="Inter, 'Segoe UI', Arial, Helvetica, system-ui, sans-serif" font-size="16.5" fill="#a1a1aa">Claude Code · Codex · Cursor · Cline · Copilot · Antigravity&#160;&#160;→&#160;&#160;<tspan fill="#7ee787" font-weight="700">FREE</tspan> Claude / GPT / Gemini · auto-fallback</text>

docs/guides/CODEX-APP-SERVER-PROVIDER.md

@@ -0,0 +1,89 @@
+---
+title: "OpenAI Codex (App-Server) provider"
+version: 3.8.50
+lastUpdated: 2026-08-22
+---
+
+# OpenAI Codex — App-Server provider (`codex-app-server`)
+
+OmniRoute exposes **two** ways to use OpenAI Codex:
+
+| Provider | How it talks to OpenAI | Usage caveat |
+|---|---|---|
+| **`codex`** | Replays your ChatGPT/OpenAI OAuth token directly to the Responses API | **Yes** — the official session is not authorized for proxy/router use |
+| **`codex-app-server`** | Drives the **Codex CLI's own `codex app-server`** over JSON-RPC/WebSocket; the CLI owns and self-refreshes its OAuth (`~/.codex/auth.json`) exactly like an interactive `codex` session | **No** — OmniRoute never replays a token to the API |
+
+Because `codex-app-server` never replays a token, it does not carry the
+session-replay usage caveat. It does require a **Codex CLI reachable at the
+configured app-server URL**, and that CLI must be **signed in**.
+
+---
+
+## 1. Architecture
+
+```
+┌─ OmniRoute app ─────────────────┐        ┌─ codex-app-server sidecar ─────────┐
+│  CodexAppServerExecutor          │  WS    │  codex app-server                   │
+│  ws://codex-app-server:1456 ─────┼───────▶│  --listen ws://0.0.0.0:1456         │
+│  (+ capability token)            │  JSON  │  --ws-auth capability-token         │
+│                                  │  RPC   │  self-manages OpenAI OAuth          │
+└──────────────────────────────────┘        │  (~/.codex/auth.json, auto-refresh) │
+        │ shares (compose volumes)           └─────────────────────────────────────┘
+        ▼
+  codex-appserver-token   → the WS capability token (both mount it)
+  codex-appserver-home    → ~/.codex (auth.json written by the dashboard,
+                             read by the sidecar's codex app-server)
+```
+
+- The sidecar listens **only** on the internal compose network
+  (`ws://codex-app-server:1456`) behind a capability token. It is **never**
+  published to the host or internet.
+- The Codex CLI is baked into `omniroute:base`, so no codex install is needed on
+  the host or the user's machine when you run the sidecar.
+
+## 2. Bring it up
+
+```bash
+# Start the stack WITH the codex app-server sidecar profile:
+docker compose --profile base --profile codex-app-server up -d
+# (podman: podman compose --profile base --profile codex-app-server up -d)
+```
+
+The sidecar mints its WS capability token on first boot (into the shared
+`codex-appserver-token` volume) and the app reads the same token via
+`OMNIROUTE_CODEX_APPSERVER_WS_TOKEN_FILE`. No manual token wiring needed.
+
+## 3. Connect + sign in
+
+1. In the dashboard, add a connection for **OpenAI Codex (App-Server)**. No API
+   key or token is required (it's a no-auth provider — the sidecar owns auth).
+2. If the sidecar's Codex CLI is **not yet signed in**, the connection health
+   check reports *"running but not signed in"* (not a red auth error). Use
+   **Sign in with ChatGPT**: this runs the standard Codex device-OAuth in your
+   browser and then writes `~/.codex/auth.json` into the shared volume via
+   **Apply auth** (the same one login serves both the `codex` and
+   `codex-app-server` providers).
+3. Once signed in, the health check goes green (it verifies both `/readyz` **and**
+   `account/read` — i.e. up *and* authenticated) and turns work.
+
+The dashboard never clobbers a healthy existing `~/.codex/auth.json` — it writes
+only when the file is absent or its token is stale (a backup is always taken).
+
+## 4. Deployment scenarios
+
+- **Operator with an already-authenticated Codex CLI** — mount your host
+  `~/.codex` into the sidecar (`codex-appserver-home`) and skip the sign-in step.
+- **Public user, no codex installed locally** — irrelevant: the sidecar has the
+  CLI. The user only authenticates through the dashboard.
+- **Bare-metal OmniRoute (no sidecar, host codex)** — point
+  `OMNIROUTE_CODEX_APPSERVER_WS` at your own `codex app-server` and ensure the
+  host codex is signed in; the "codex not installed" hint appears if the binary
+  is missing.
+
+## 5. Residential / UDP egress (operator extra, not shipped)
+
+The generic sidecar above egresses over the container's normal network. An
+operator who needs Codex traffic to egress via a **residential exit** (e.g. a TUN
+tailscale sidecar carrying TCP + UDP/QUIC) runs that as a separate compose
+override; it is intentionally **not** part of the shipped `codex-app-server`
+profile. See the internal operations runbook for that setup.

docs/i18n/ar/llm.txt

@@ -4,7 +4,7 @@
 
 ---
 
-> OmniRoute is a free, open-source AI Gateway that acts as a universal API proxy for multi-provider LLMs. It provides smart routing, automatic fallback, load balancing, and format translation across 350 AI providers — all through a single OpenAI-compatible endpoint. Includes a built-in MCP Server (110 tools), A2A v0.3 protocol, Memory/Skills systems, Cloud Agents (codex, cursor, devin, jules), Guardrails framework, and an Electron desktop app.
+> OmniRoute is a free, open-source AI Gateway that acts as a universal API proxy for multi-provider LLMs. It provides smart routing, automatic fallback, load balancing, and format translation across 351 AI providers — all through a single OpenAI-compatible endpoint. Includes a built-in MCP Server (110 tools), A2A v0.3 protocol, Memory/Skills systems, Cloud Agents (codex, cursor, devin, jules), Guardrails framework, and an Electron desktop app.
 
 ## Overview
 
@@ -281,7 +281,7 @@ OmniRoute solves the problem of managing multiple AI provider subscriptions, quo
 ## Key Features (v3.8.50)
 
 ### Core Proxy
-- **350 AI providers** with automatic format translation
+- **351 AI providers** with automatic format translation
 - **Provider categories**: Free (90+ free tiers), OAuth, API Key, Self-Hosted, Custom (OpenAI/Anthropic-compatible)
 - **18 routing strategies**: priority, weighted, round-robin, fill-first, p2c, random, least-used, cost-optimized, reset-aware, reset-window, headroom, strict-random, auto, lkgp, context-optimized, context-relay, fusion, pipeline
 - **4-tier fallback**: Subscription → API Key → Cheap → Free

docs/i18n/az/llm.txt

@@ -4,7 +4,7 @@
 
 ---
 
-> OmniRoute is a free, open-source AI Gateway that acts as a universal API proxy for multi-provider LLMs. It provides smart routing, automatic fallback, load balancing, and format translation across 350 AI providers — all through a single OpenAI-compatible endpoint. Includes a built-in MCP Server (110 tools), A2A v0.3 protocol, Memory/Skills systems, Cloud Agents (codex, cursor, devin, jules), Guardrails framework, and an Electron desktop app.
+> OmniRoute is a free, open-source AI Gateway that acts as a universal API proxy for multi-provider LLMs. It provides smart routing, automatic fallback, load balancing, and format translation across 351 AI providers — all through a single OpenAI-compatible endpoint. Includes a built-in MCP Server (110 tools), A2A v0.3 protocol, Memory/Skills systems, Cloud Agents (codex, cursor, devin, jules), Guardrails framework, and an Electron desktop app.
 
 ## Overview
 
@@ -281,7 +281,7 @@ OmniRoute solves the problem of managing multiple AI provider subscriptions, quo
 ## Key Features (v3.8.50)
 
 ### Core Proxy
-- **350 AI providers** with automatic format translation
+- **351 AI providers** with automatic format translation
 - **Provider categories**: Free (90+ free tiers), OAuth, API Key, Self-Hosted, Custom (OpenAI/Anthropic-compatible)
 - **18 routing strategies**: priority, weighted, round-robin, fill-first, p2c, random, least-used, cost-optimized, reset-aware, reset-window, headroom, strict-random, auto, lkgp, context-optimized, context-relay, fusion, pipeline
 - **4-tier fallback**: Subscription → API Key → Cheap → Free

docs/i18n/bg/llm.txt

@@ -4,7 +4,7 @@
 
 ---
 
-> OmniRoute is a free, open-source AI Gateway that acts as a universal API proxy for multi-provider LLMs. It provides smart routing, automatic fallback, load balancing, and format translation across 350 AI providers — all through a single OpenAI-compatible endpoint. Includes a built-in MCP Server (110 tools), A2A v0.3 protocol, Memory/Skills systems, Cloud Agents (codex, cursor, devin, jules), Guardrails framework, and an Electron desktop app.
+> OmniRoute is a free, open-source AI Gateway that acts as a universal API proxy for multi-provider LLMs. It provides smart routing, automatic fallback, load balancing, and format translation across 351 AI providers — all through a single OpenAI-compatible endpoint. Includes a built-in MCP Server (110 tools), A2A v0.3 protocol, Memory/Skills systems, Cloud Agents (codex, cursor, devin, jules), Guardrails framework, and an Electron desktop app.
 
 ## Overview
 
@@ -281,7 +281,7 @@ OmniRoute solves the problem of managing multiple AI provider subscriptions, quo
 ## Key Features (v3.8.50)
 
 ### Core Proxy
-- **350 AI providers** with automatic format translation
+- **351 AI providers** with automatic format translation
 - **Provider categories**: Free (90+ free tiers), OAuth, API Key, Self-Hosted, Custom (OpenAI/Anthropic-compatible)
 - **18 routing strategies**: priority, weighted, round-robin, fill-first, p2c, random, least-used, cost-optimized, reset-aware, reset-window, headroom, strict-random, auto, lkgp, context-optimized, context-relay, fusion, pipeline
 - **4-tier fallback**: Subscription → API Key → Cheap → Free

docs/i18n/bn/llm.txt

@@ -4,7 +4,7 @@
 
 ---
 
-> OmniRoute is a free, open-source AI Gateway that acts as a universal API proxy for multi-provider LLMs. It provides smart routing, automatic fallback, load balancing, and format translation across 350 AI providers — all through a single OpenAI-compatible endpoint. Includes a built-in MCP Server (110 tools), A2A v0.3 protocol, Memory/Skills systems, Cloud Agents (codex, cursor, devin, jules), Guardrails framework, and an Electron desktop app.
+> OmniRoute is a free, open-source AI Gateway that acts as a universal API proxy for multi-provider LLMs. It provides smart routing, automatic fallback, load balancing, and format translation across 351 AI providers — all through a single OpenAI-compatible endpoint. Includes a built-in MCP Server (110 tools), A2A v0.3 protocol, Memory/Skills systems, Cloud Agents (codex, cursor, devin, jules), Guardrails framework, and an Electron desktop app.
 
 ## Overview
 
@@ -281,7 +281,7 @@ OmniRoute solves the problem of managing multiple AI provider subscriptions, quo
 ## Key Features (v3.8.50)
 
 ### Core Proxy
-- **350 AI providers** with automatic format translation
+- **351 AI providers** with automatic format translation
 - **Provider categories**: Free (90+ free tiers), OAuth, API Key, Self-Hosted, Custom (OpenAI/Anthropic-compatible)
 - **18 routing strategies**: priority, weighted, round-robin, fill-first, p2c, random, least-used, cost-optimized, reset-aware, reset-window, headroom, strict-random, auto, lkgp, context-optimized, context-relay, fusion, pipeline
 - **4-tier fallback**: Subscription → API Key → Cheap → Free

docs/i18n/cs/llm.txt

@@ -4,7 +4,7 @@
 
 ---
 
-> OmniRoute is a free, open-source AI Gateway that acts as a universal API proxy for multi-provider LLMs. It provides smart routing, automatic fallback, load balancing, and format translation across 350 AI providers — all through a single OpenAI-compatible endpoint. Includes a built-in MCP Server (110 tools), A2A v0.3 protocol, Memory/Skills systems, Cloud Agents (codex, cursor, devin, jules), Guardrails framework, and an Electron desktop app.
+> OmniRoute is a free, open-source AI Gateway that acts as a universal API proxy for multi-provider LLMs. It provides smart routing, automatic fallback, load balancing, and format translation across 351 AI providers — all through a single OpenAI-compatible endpoint. Includes a built-in MCP Server (110 tools), A2A v0.3 protocol, Memory/Skills systems, Cloud Agents (codex, cursor, devin, jules), Guardrails framework, and an Electron desktop app.
 
 ## Overview
 
@@ -281,7 +281,7 @@ OmniRoute solves the problem of managing multiple AI provider subscriptions, quo
 ## Key Features (v3.8.50)
 
 ### Core Proxy
-- **350 AI providers** with automatic format translation
+- **351 AI providers** with automatic format translation
 - **Provider categories**: Free (90+ free tiers), OAuth, API Key, Self-Hosted, Custom (OpenAI/Anthropic-compatible)
 - **18 routing strategies**: priority, weighted, round-robin, fill-first, p2c, random, least-used, cost-optimized, reset-aware, reset-window, headroom, strict-random, auto, lkgp, context-optimized, context-relay, fusion, pipeline
 - **4-tier fallback**: Subscription → API Key → Cheap → Free

docs/i18n/da/llm.txt

@@ -4,7 +4,7 @@
 
 ---
 
-> OmniRoute is a free, open-source AI Gateway that acts as a universal API proxy for multi-provider LLMs. It provides smart routing, automatic fallback, load balancing, and format translation across 350 AI providers — all through a single OpenAI-compatible endpoint. Includes a built-in MCP Server (110 tools), A2A v0.3 protocol, Memory/Skills systems, Cloud Agents (codex, cursor, devin, jules), Guardrails framework, and an Electron desktop app.
+> OmniRoute is a free, open-source AI Gateway that acts as a universal API proxy for multi-provider LLMs. It provides smart routing, automatic fallback, load balancing, and format translation across 351 AI providers — all through a single OpenAI-compatible endpoint. Includes a built-in MCP Server (110 tools), A2A v0.3 protocol, Memory/Skills systems, Cloud Agents (codex, cursor, devin, jules), Guardrails framework, and an Electron desktop app.
 
 ## Overview
 
@@ -281,7 +281,7 @@ OmniRoute solves the problem of managing multiple AI provider subscriptions, quo
 ## Key Features (v3.8.50)
 
 ### Core Proxy
-- **350 AI providers** with automatic format translation
+- **351 AI providers** with automatic format translation
 - **Provider categories**: Free (90+ free tiers), OAuth, API Key, Self-Hosted, Custom (OpenAI/Anthropic-compatible)
 - **18 routing strategies**: priority, weighted, round-robin, fill-first, p2c, random, least-used, cost-optimized, reset-aware, reset-window, headroom, strict-random, auto, lkgp, context-optimized, context-relay, fusion, pipeline
 - **4-tier fallback**: Subscription → API Key → Cheap → Free

docs/i18n/de/llm.txt

@@ -4,7 +4,7 @@
 
 ---
 
-> OmniRoute is a free, open-source AI Gateway that acts as a universal API proxy for multi-provider LLMs. It provides smart routing, automatic fallback, load balancing, and format translation across 350 AI providers — all through a single OpenAI-compatible endpoint. Includes a built-in MCP Server (110 tools), A2A v0.3 protocol, Memory/Skills systems, Cloud Agents (codex, cursor, devin, jules), Guardrails framework, and an Electron desktop app.
+> OmniRoute is a free, open-source AI Gateway that acts as a universal API proxy for multi-provider LLMs. It provides smart routing, automatic fallback, load balancing, and format translation across 351 AI providers — all through a single OpenAI-compatible endpoint. Includes a built-in MCP Server (110 tools), A2A v0.3 protocol, Memory/Skills systems, Cloud Agents (codex, cursor, devin, jules), Guardrails framework, and an Electron desktop app.
 
 ## Overview
 
@@ -281,7 +281,7 @@ OmniRoute solves the problem of managing multiple AI provider subscriptions, quo
 ## Key Features (v3.8.50)
 
 ### Core Proxy
-- **350 AI providers** with automatic format translation
+- **351 AI providers** with automatic format translation
 - **Provider categories**: Free (90+ free tiers), OAuth, API Key, Self-Hosted, Custom (OpenAI/Anthropic-compatible)
 - **18 routing strategies**: priority, weighted, round-robin, fill-first, p2c, random, least-used, cost-optimized, reset-aware, reset-window, headroom, strict-random, auto, lkgp, context-optimized, context-relay, fusion, pipeline
 - **4-tier fallback**: Subscription → API Key → Cheap → Free

docs/i18n/es/llm.txt

@@ -4,7 +4,7 @@
 
 ---
 
-> OmniRoute is a free, open-source AI Gateway that acts as a universal API proxy for multi-provider LLMs. It provides smart routing, automatic fallback, load balancing, and format translation across 350 AI providers — all through a single OpenAI-compatible endpoint. Includes a built-in MCP Server (110 tools), A2A v0.3 protocol, Memory/Skills systems, Cloud Agents (codex, cursor, devin, jules), Guardrails framework, and an Electron desktop app.
+> OmniRoute is a free, open-source AI Gateway that acts as a universal API proxy for multi-provider LLMs. It provides smart routing, automatic fallback, load balancing, and format translation across 351 AI providers — all through a single OpenAI-compatible endpoint. Includes a built-in MCP Server (110 tools), A2A v0.3 protocol, Memory/Skills systems, Cloud Agents (codex, cursor, devin, jules), Guardrails framework, and an Electron desktop app.
 
 ## Overview
 
@@ -281,7 +281,7 @@ OmniRoute solves the problem of managing multiple AI provider subscriptions, quo
 ## Key Features (v3.8.50)
 
 ### Core Proxy
-- **350 AI providers** with automatic format translation
+- **351 AI providers** with automatic format translation
 - **Provider categories**: Free (90+ free tiers), OAuth, API Key, Self-Hosted, Custom (OpenAI/Anthropic-compatible)
 - **18 routing strategies**: priority, weighted, round-robin, fill-first, p2c, random, least-used, cost-optimized, reset-aware, reset-window, headroom, strict-random, auto, lkgp, context-optimized, context-relay, fusion, pipeline
 - **4-tier fallback**: Subscription → API Key → Cheap → Free

docs/i18n/fa/llm.txt

@@ -4,7 +4,7 @@
 
 ---
 
-> OmniRoute is a free, open-source AI Gateway that acts as a universal API proxy for multi-provider LLMs. It provides smart routing, automatic fallback, load balancing, and format translation across 350 AI providers — all through a single OpenAI-compatible endpoint. Includes a built-in MCP Server (110 tools), A2A v0.3 protocol, Memory/Skills systems, Cloud Agents (codex, cursor, devin, jules), Guardrails framework, and an Electron desktop app.
+> OmniRoute is a free, open-source AI Gateway that acts as a universal API proxy for multi-provider LLMs. It provides smart routing, automatic fallback, load balancing, and format translation across 351 AI providers — all through a single OpenAI-compatible endpoint. Includes a built-in MCP Server (110 tools), A2A v0.3 protocol, Memory/Skills systems, Cloud Agents (codex, cursor, devin, jules), Guardrails framework, and an Electron desktop app.
 
 ## Overview
 
@@ -281,7 +281,7 @@ OmniRoute solves the problem of managing multiple AI provider subscriptions, quo
 ## Key Features (v3.8.50)
 
 ### Core Proxy
-- **350 AI providers** with automatic format translation
+- **351 AI providers** with automatic format translation
 - **Provider categories**: Free (90+ free tiers), OAuth, API Key, Self-Hosted, Custom (OpenAI/Anthropic-compatible)
 - **18 routing strategies**: priority, weighted, round-robin, fill-first, p2c, random, least-used, cost-optimized, reset-aware, reset-window, headroom, strict-random, auto, lkgp, context-optimized, context-relay, fusion, pipeline
 - **4-tier fallback**: Subscription → API Key → Cheap → Free

docs/i18n/fi/llm.txt

@@ -4,7 +4,7 @@
 
 ---
 
-> OmniRoute is a free, open-source AI Gateway that acts as a universal API proxy for multi-provider LLMs. It provides smart routing, automatic fallback, load balancing, and format translation across 350 AI providers — all through a single OpenAI-compatible endpoint. Includes a built-in MCP Server (110 tools), A2A v0.3 protocol, Memory/Skills systems, Cloud Agents (codex, cursor, devin, jules), Guardrails framework, and an Electron desktop app.
+> OmniRoute is a free, open-source AI Gateway that acts as a universal API proxy for multi-provider LLMs. It provides smart routing, automatic fallback, load balancing, and format translation across 351 AI providers — all through a single OpenAI-compatible endpoint. Includes a built-in MCP Server (110 tools), A2A v0.3 protocol, Memory/Skills systems, Cloud Agents (codex, cursor, devin, jules), Guardrails framework, and an Electron desktop app.
 
 ## Overview
 
@@ -281,7 +281,7 @@ OmniRoute solves the problem of managing multiple AI provider subscriptions, quo
 ## Key Features (v3.8.50)
 
 ### Core Proxy
-- **350 AI providers** with automatic format translation
+- **351 AI providers** with automatic format translation
 - **Provider categories**: Free (90+ free tiers), OAuth, API Key, Self-Hosted, Custom (OpenAI/Anthropic-compatible)
 - **18 routing strategies**: priority, weighted, round-robin, fill-first, p2c, random, least-used, cost-optimized, reset-aware, reset-window, headroom, strict-random, auto, lkgp, context-optimized, context-relay, fusion, pipeline
 - **4-tier fallback**: Subscription → API Key → Cheap → Free

docs/i18n/fr/llm.txt

@@ -4,7 +4,7 @@
 
 ---
 
-> OmniRoute is a free, open-source AI Gateway that acts as a universal API proxy for multi-provider LLMs. It provides smart routing, automatic fallback, load balancing, and format translation across 350 AI providers — all through a single OpenAI-compatible endpoint. Includes a built-in MCP Server (110 tools), A2A v0.3 protocol, Memory/Skills systems, Cloud Agents (codex, cursor, devin, jules), Guardrails framework, and an Electron desktop app.
+> OmniRoute is a free, open-source AI Gateway that acts as a universal API proxy for multi-provider LLMs. It provides smart routing, automatic fallback, load balancing, and format translation across 351 AI providers — all through a single OpenAI-compatible endpoint. Includes a built-in MCP Server (110 tools), A2A v0.3 protocol, Memory/Skills systems, Cloud Agents (codex, cursor, devin, jules), Guardrails framework, and an Electron desktop app.
 
 ## Overview
 
@@ -281,7 +281,7 @@ OmniRoute solves the problem of managing multiple AI provider subscriptions, quo
 ## Key Features (v3.8.50)
 
 ### Core Proxy
-- **350 AI providers** with automatic format translation
+- **351 AI providers** with automatic format translation
 - **Provider categories**: Free (90+ free tiers), OAuth, API Key, Self-Hosted, Custom (OpenAI/Anthropic-compatible)
 - **18 routing strategies**: priority, weighted, round-robin, fill-first, p2c, random, least-used, cost-optimized, reset-aware, reset-window, headroom, strict-random, auto, lkgp, context-optimized, context-relay, fusion, pipeline
 - **4-tier fallback**: Subscription → API Key → Cheap → Free

docs/i18n/gu/llm.txt

@@ -4,7 +4,7 @@
 
 ---
 
-> OmniRoute is a free, open-source AI Gateway that acts as a universal API proxy for multi-provider LLMs. It provides smart routing, automatic fallback, load balancing, and format translation across 350 AI providers — all through a single OpenAI-compatible endpoint. Includes a built-in MCP Server (110 tools), A2A v0.3 protocol, Memory/Skills systems, Cloud Agents (codex, cursor, devin, jules), Guardrails framework, and an Electron desktop app.
+> OmniRoute is a free, open-source AI Gateway that acts as a universal API proxy for multi-provider LLMs. It provides smart routing, automatic fallback, load balancing, and format translation across 351 AI providers — all through a single OpenAI-compatible endpoint. Includes a built-in MCP Server (110 tools), A2A v0.3 protocol, Memory/Skills systems, Cloud Agents (codex, cursor, devin, jules), Guardrails framework, and an Electron desktop app.
 
 ## Overview
 
@@ -281,7 +281,7 @@ OmniRoute solves the problem of managing multiple AI provider subscriptions, quo
 ## Key Features (v3.8.50)
 
 ### Core Proxy
-- **350 AI providers** with automatic format translation
+- **351 AI providers** with automatic format translation
 - **Provider categories**: Free (90+ free tiers), OAuth, API Key, Self-Hosted, Custom (OpenAI/Anthropic-compatible)
 - **18 routing strategies**: priority, weighted, round-robin, fill-first, p2c, random, least-used, cost-optimized, reset-aware, reset-window, headroom, strict-random, auto, lkgp, context-optimized, context-relay, fusion, pipeline
 - **4-tier fallback**: Subscription → API Key → Cheap → Free

docs/i18n/he/llm.txt

@@ -4,7 +4,7 @@
 
 ---
 
-> OmniRoute is a free, open-source AI Gateway that acts as a universal API proxy for multi-provider LLMs. It provides smart routing, automatic fallback, load balancing, and format translation across 350 AI providers — all through a single OpenAI-compatible endpoint. Includes a built-in MCP Server (110 tools), A2A v0.3 protocol, Memory/Skills systems, Cloud Agents (codex, cursor, devin, jules), Guardrails framework, and an Electron desktop app.
+> OmniRoute is a free, open-source AI Gateway that acts as a universal API proxy for multi-provider LLMs. It provides smart routing, automatic fallback, load balancing, and format translation across 351 AI providers — all through a single OpenAI-compatible endpoint. Includes a built-in MCP Server (110 tools), A2A v0.3 protocol, Memory/Skills systems, Cloud Agents (codex, cursor, devin, jules), Guardrails framework, and an Electron desktop app.
 
 ## Overview
 
@@ -281,7 +281,7 @@ OmniRoute solves the problem of managing multiple AI provider subscriptions, quo
 ## Key Features (v3.8.50)
 
 ### Core Proxy
-- **350 AI providers** with automatic format translation
+- **351 AI providers** with automatic format translation
 - **Provider categories**: Free (90+ free tiers), OAuth, API Key, Self-Hosted, Custom (OpenAI/Anthropic-compatible)
 - **18 routing strategies**: priority, weighted, round-robin, fill-first, p2c, random, least-used, cost-optimized, reset-aware, reset-window, headroom, strict-random, auto, lkgp, context-optimized, context-relay, fusion, pipeline
 - **4-tier fallback**: Subscription → API Key → Cheap → Free

docs/i18n/hi/llm.txt

@@ -4,7 +4,7 @@
 
 ---
 
-> OmniRoute is a free, open-source AI Gateway that acts as a universal API proxy for multi-provider LLMs. It provides smart routing, automatic fallback, load balancing, and format translation across 350 AI providers — all through a single OpenAI-compatible endpoint. Includes a built-in MCP Server (110 tools), A2A v0.3 protocol, Memory/Skills systems, Cloud Agents (codex, cursor, devin, jules), Guardrails framework, and an Electron desktop app.
+> OmniRoute is a free, open-source AI Gateway that acts as a universal API proxy for multi-provider LLMs. It provides smart routing, automatic fallback, load balancing, and format translation across 351 AI providers — all through a single OpenAI-compatible endpoint. Includes a built-in MCP Server (110 tools), A2A v0.3 protocol, Memory/Skills systems, Cloud Agents (codex, cursor, devin, jules), Guardrails framework, and an Electron desktop app.
 
 ## Overview
 
@@ -281,7 +281,7 @@ OmniRoute solves the problem of managing multiple AI provider subscriptions, quo
 ## Key Features (v3.8.50)
 
 ### Core Proxy
-- **350 AI providers** with automatic format translation
+- **351 AI providers** with automatic format translation
 - **Provider categories**: Free (90+ free tiers), OAuth, API Key, Self-Hosted, Custom (OpenAI/Anthropic-compatible)
 - **18 routing strategies**: priority, weighted, round-robin, fill-first, p2c, random, least-used, cost-optimized, reset-aware, reset-window, headroom, strict-random, auto, lkgp, context-optimized, context-relay, fusion, pipeline
 - **4-tier fallback**: Subscription → API Key → Cheap → Free

docs/i18n/hu/llm.txt

@@ -4,7 +4,7 @@
 
 ---
 
-> OmniRoute is a free, open-source AI Gateway that acts as a universal API proxy for multi-provider LLMs. It provides smart routing, automatic fallback, load balancing, and format translation across 350 AI providers — all through a single OpenAI-compatible endpoint. Includes a built-in MCP Server (110 tools), A2A v0.3 protocol, Memory/Skills systems, Cloud Agents (codex, cursor, devin, jules), Guardrails framework, and an Electron desktop app.
+> OmniRoute is a free, open-source AI Gateway that acts as a universal API proxy for multi-provider LLMs. It provides smart routing, automatic fallback, load balancing, and format translation across 351 AI providers — all through a single OpenAI-compatible endpoint. Includes a built-in MCP Server (110 tools), A2A v0.3 protocol, Memory/Skills systems, Cloud Agents (codex, cursor, devin, jules), Guardrails framework, and an Electron desktop app.
 
 ## Overview
 
@@ -281,7 +281,7 @@ OmniRoute solves the problem of managing multiple AI provider subscriptions, quo
 ## Key Features (v3.8.50)
 
 ### Core Proxy
-- **350 AI providers** with automatic format translation
+- **351 AI providers** with automatic format translation
 - **Provider categories**: Free (90+ free tiers), OAuth, API Key, Self-Hosted, Custom (OpenAI/Anthropic-compatible)
 - **18 routing strategies**: priority, weighted, round-robin, fill-first, p2c, random, least-used, cost-optimized, reset-aware, reset-window, headroom, strict-random, auto, lkgp, context-optimized, context-relay, fusion, pipeline
 - **4-tier fallback**: Subscription → API Key → Cheap → Free

docs/i18n/id/llm.txt

@@ -4,7 +4,7 @@
 
 ---
 
-> OmniRoute is a free, open-source AI Gateway that acts as a universal API proxy for multi-provider LLMs. It provides smart routing, automatic fallback, load balancing, and format translation across 350 AI providers — all through a single OpenAI-compatible endpoint. Includes a built-in MCP Server (110 tools), A2A v0.3 protocol, Memory/Skills systems, Cloud Agents (codex, cursor, devin, jules), Guardrails framework, and an Electron desktop app.
+> OmniRoute is a free, open-source AI Gateway that acts as a universal API proxy for multi-provider LLMs. It provides smart routing, automatic fallback, load balancing, and format translation across 351 AI providers — all through a single OpenAI-compatible endpoint. Includes a built-in MCP Server (110 tools), A2A v0.3 protocol, Memory/Skills systems, Cloud Agents (codex, cursor, devin, jules), Guardrails framework, and an Electron desktop app.
 
 ## Overview
 
@@ -281,7 +281,7 @@ OmniRoute solves the problem of managing multiple AI provider subscriptions, quo
 ## Key Features (v3.8.50)
 
 ### Core Proxy
-- **350 AI providers** with automatic format translation
+- **351 AI providers** with automatic format translation
 - **Provider categories**: Free (90+ free tiers), OAuth, API Key, Self-Hosted, Custom (OpenAI/Anthropic-compatible)
 - **18 routing strategies**: priority, weighted, round-robin, fill-first, p2c, random, least-used, cost-optimized, reset-aware, reset-window, headroom, strict-random, auto, lkgp, context-optimized, context-relay, fusion, pipeline
 - **4-tier fallback**: Subscription → API Key → Cheap → Free

docs/i18n/in/llm.txt

@@ -4,7 +4,7 @@
 
 ---
 
-> OmniRoute is a free, open-source AI Gateway that acts as a universal API proxy for multi-provider LLMs. It provides smart routing, automatic fallback, load balancing, and format translation across 350 AI providers — all through a single OpenAI-compatible endpoint. Includes a built-in MCP Server (110 tools), A2A v0.3 protocol, Memory/Skills systems, Cloud Agents (codex, cursor, devin, jules), Guardrails framework, and an Electron desktop app.
+> OmniRoute is a free, open-source AI Gateway that acts as a universal API proxy for multi-provider LLMs. It provides smart routing, automatic fallback, load balancing, and format translation across 351 AI providers — all through a single OpenAI-compatible endpoint. Includes a built-in MCP Server (110 tools), A2A v0.3 protocol, Memory/Skills systems, Cloud Agents (codex, cursor, devin, jules), Guardrails framework, and an Electron desktop app.
 
 ## Overview
 
@@ -281,7 +281,7 @@ OmniRoute solves the problem of managing multiple AI provider subscriptions, quo
 ## Key Features (v3.8.50)
 
 ### Core Proxy
-- **350 AI providers** with automatic format translation
+- **351 AI providers** with automatic format translation
 - **Provider categories**: Free (90+ free tiers), OAuth, API Key, Self-Hosted, Custom (OpenAI/Anthropic-compatible)
 - **18 routing strategies**: priority, weighted, round-robin, fill-first, p2c, random, least-used, cost-optimized, reset-aware, reset-window, headroom, strict-random, auto, lkgp, context-optimized, context-relay, fusion, pipeline
 - **4-tier fallback**: Subscription → API Key → Cheap → Free

docs/i18n/it/llm.txt

@@ -4,7 +4,7 @@
 
 ---
 
-> OmniRoute is a free, open-source AI Gateway that acts as a universal API proxy for multi-provider LLMs. It provides smart routing, automatic fallback, load balancing, and format translation across 350 AI providers — all through a single OpenAI-compatible endpoint. Includes a built-in MCP Server (110 tools), A2A v0.3 protocol, Memory/Skills systems, Cloud Agents (codex, cursor, devin, jules), Guardrails framework, and an Electron desktop app.
+> OmniRoute is a free, open-source AI Gateway that acts as a universal API proxy for multi-provider LLMs. It provides smart routing, automatic fallback, load balancing, and format translation across 351 AI providers — all through a single OpenAI-compatible endpoint. Includes a built-in MCP Server (110 tools), A2A v0.3 protocol, Memory/Skills systems, Cloud Agents (codex, cursor, devin, jules), Guardrails framework, and an Electron desktop app.
 
 ## Overview
 
@@ -281,7 +281,7 @@ OmniRoute solves the problem of managing multiple AI provider subscriptions, quo
 ## Key Features (v3.8.50)
 
 ### Core Proxy
-- **350 AI providers** with automatic format translation
+- **351 AI providers** with automatic format translation
 - **Provider categories**: Free (90+ free tiers), OAuth, API Key, Self-Hosted, Custom (OpenAI/Anthropic-compatible)
 - **18 routing strategies**: priority, weighted, round-robin, fill-first, p2c, random, least-used, cost-optimized, reset-aware, reset-window, headroom, strict-random, auto, lkgp, context-optimized, context-relay, fusion, pipeline
 - **4-tier fallback**: Subscription → API Key → Cheap → Free

docs/i18n/ja/llm.txt

@@ -4,7 +4,7 @@
 
 ---
 
-> OmniRoute is a free, open-source AI Gateway that acts as a universal API proxy for multi-provider LLMs. It provides smart routing, automatic fallback, load balancing, and format translation across 350 AI providers — all through a single OpenAI-compatible endpoint. Includes a built-in MCP Server (110 tools), A2A v0.3 protocol, Memory/Skills systems, Cloud Agents (codex, cursor, devin, jules), Guardrails framework, and an Electron desktop app.
+> OmniRoute is a free, open-source AI Gateway that acts as a universal API proxy for multi-provider LLMs. It provides smart routing, automatic fallback, load balancing, and format translation across 351 AI providers — all through a single OpenAI-compatible endpoint. Includes a built-in MCP Server (110 tools), A2A v0.3 protocol, Memory/Skills systems, Cloud Agents (codex, cursor, devin, jules), Guardrails framework, and an Electron desktop app.
 
 ## Overview
 
@@ -281,7 +281,7 @@ OmniRoute solves the problem of managing multiple AI provider subscriptions, quo
 ## Key Features (v3.8.50)
 
 ### Core Proxy
-- **350 AI providers** with automatic format translation
+- **351 AI providers** with automatic format translation
 - **Provider categories**: Free (90+ free tiers), OAuth, API Key, Self-Hosted, Custom (OpenAI/Anthropic-compatible)
 - **18 routing strategies**: priority, weighted, round-robin, fill-first, p2c, random, least-used, cost-optimized, reset-aware, reset-window, headroom, strict-random, auto, lkgp, context-optimized, context-relay, fusion, pipeline
 - **4-tier fallback**: Subscription → API Key → Cheap → Free

docs/i18n/ko/llm.txt

@@ -4,7 +4,7 @@
 
 ---
 
-> OmniRoute is a free, open-source AI Gateway that acts as a universal API proxy for multi-provider LLMs. It provides smart routing, automatic fallback, load balancing, and format translation across 350 AI providers — all through a single OpenAI-compatible endpoint. Includes a built-in MCP Server (110 tools), A2A v0.3 protocol, Memory/Skills systems, Cloud Agents (codex, cursor, devin, jules), Guardrails framework, and an Electron desktop app.
+> OmniRoute is a free, open-source AI Gateway that acts as a universal API proxy for multi-provider LLMs. It provides smart routing, automatic fallback, load balancing, and format translation across 351 AI providers — all through a single OpenAI-compatible endpoint. Includes a built-in MCP Server (110 tools), A2A v0.3 protocol, Memory/Skills systems, Cloud Agents (codex, cursor, devin, jules), Guardrails framework, and an Electron desktop app.
 
 ## Overview
 
@@ -281,7 +281,7 @@ OmniRoute solves the problem of managing multiple AI provider subscriptions, quo
 ## Key Features (v3.8.50)
 
 ### Core Proxy
-- **350 AI providers** with automatic format translation
+- **351 AI providers** with automatic format translation
 - **Provider categories**: Free (90+ free tiers), OAuth, API Key, Self-Hosted, Custom (OpenAI/Anthropic-compatible)
 - **18 routing strategies**: priority, weighted, round-robin, fill-first, p2c, random, least-used, cost-optimized, reset-aware, reset-window, headroom, strict-random, auto, lkgp, context-optimized, context-relay, fusion, pipeline
 - **4-tier fallback**: Subscription → API Key → Cheap → Free

docs/i18n/mr/llm.txt

@@ -4,7 +4,7 @@
 
 ---
 
-> OmniRoute is a free, open-source AI Gateway that acts as a universal API proxy for multi-provider LLMs. It provides smart routing, automatic fallback, load balancing, and format translation across 350 AI providers — all through a single OpenAI-compatible endpoint. Includes a built-in MCP Server (110 tools), A2A v0.3 protocol, Memory/Skills systems, Cloud Agents (codex, cursor, devin, jules), Guardrails framework, and an Electron desktop app.
+> OmniRoute is a free, open-source AI Gateway that acts as a universal API proxy for multi-provider LLMs. It provides smart routing, automatic fallback, load balancing, and format translation across 351 AI providers — all through a single OpenAI-compatible endpoint. Includes a built-in MCP Server (110 tools), A2A v0.3 protocol, Memory/Skills systems, Cloud Agents (codex, cursor, devin, jules), Guardrails framework, and an Electron desktop app.
 
 ## Overview
 
@@ -281,7 +281,7 @@ OmniRoute solves the problem of managing multiple AI provider subscriptions, quo
 ## Key Features (v3.8.50)
 
 ### Core Proxy
-- **350 AI providers** with automatic format translation
+- **351 AI providers** with automatic format translation
 - **Provider categories**: Free (90+ free tiers), OAuth, API Key, Self-Hosted, Custom (OpenAI/Anthropic-compatible)
 - **18 routing strategies**: priority, weighted, round-robin, fill-first, p2c, random, least-used, cost-optimized, reset-aware, reset-window, headroom, strict-random, auto, lkgp, context-optimized, context-relay, fusion, pipeline
 - **4-tier fallback**: Subscription → API Key → Cheap → Free

docs/i18n/ms/llm.txt

@@ -4,7 +4,7 @@
 
 ---
 
-> OmniRoute is a free, open-source AI Gateway that acts as a universal API proxy for multi-provider LLMs. It provides smart routing, automatic fallback, load balancing, and format translation across 350 AI providers — all through a single OpenAI-compatible endpoint. Includes a built-in MCP Server (110 tools), A2A v0.3 protocol, Memory/Skills systems, Cloud Agents (codex, cursor, devin, jules), Guardrails framework, and an Electron desktop app.
+> OmniRoute is a free, open-source AI Gateway that acts as a universal API proxy for multi-provider LLMs. It provides smart routing, automatic fallback, load balancing, and format translation across 351 AI providers — all through a single OpenAI-compatible endpoint. Includes a built-in MCP Server (110 tools), A2A v0.3 protocol, Memory/Skills systems, Cloud Agents (codex, cursor, devin, jules), Guardrails framework, and an Electron desktop app.
 
 ## Overview
 
@@ -281,7 +281,7 @@ OmniRoute solves the problem of managing multiple AI provider subscriptions, quo
 ## Key Features (v3.8.50)
 
 ### Core Proxy
-- **350 AI providers** with automatic format translation
+- **351 AI providers** with automatic format translation
 - **Provider categories**: Free (90+ free tiers), OAuth, API Key, Self-Hosted, Custom (OpenAI/Anthropic-compatible)
 - **18 routing strategies**: priority, weighted, round-robin, fill-first, p2c, random, least-used, cost-optimized, reset-aware, reset-window, headroom, strict-random, auto, lkgp, context-optimized, context-relay, fusion, pipeline
 - **4-tier fallback**: Subscription → API Key → Cheap → Free

docs/i18n/nl/llm.txt

@@ -4,7 +4,7 @@
 
 ---
 
-> OmniRoute is a free, open-source AI Gateway that acts as a universal API proxy for multi-provider LLMs. It provides smart routing, automatic fallback, load balancing, and format translation across 350 AI providers — all through a single OpenAI-compatible endpoint. Includes a built-in MCP Server (110 tools), A2A v0.3 protocol, Memory/Skills systems, Cloud Agents (codex, cursor, devin, jules), Guardrails framework, and an Electron desktop app.
+> OmniRoute is a free, open-source AI Gateway that acts as a universal API proxy for multi-provider LLMs. It provides smart routing, automatic fallback, load balancing, and format translation across 351 AI providers — all through a single OpenAI-compatible endpoint. Includes a built-in MCP Server (110 tools), A2A v0.3 protocol, Memory/Skills systems, Cloud Agents (codex, cursor, devin, jules), Guardrails framework, and an Electron desktop app.
 
 ## Overview
 
@@ -281,7 +281,7 @@ OmniRoute solves the problem of managing multiple AI provider subscriptions, quo
 ## Key Features (v3.8.50)
 
 ### Core Proxy
-- **350 AI providers** with automatic format translation
+- **351 AI providers** with automatic format translation
 - **Provider categories**: Free (90+ free tiers), OAuth, API Key, Self-Hosted, Custom (OpenAI/Anthropic-compatible)
 - **18 routing strategies**: priority, weighted, round-robin, fill-first, p2c, random, least-used, cost-optimized, reset-aware, reset-window, headroom, strict-random, auto, lkgp, context-optimized, context-relay, fusion, pipeline
 - **4-tier fallback**: Subscription → API Key → Cheap → Free

docs/i18n/no/llm.txt

@@ -4,7 +4,7 @@
 
 ---
 
-> OmniRoute is a free, open-source AI Gateway that acts as a universal API proxy for multi-provider LLMs. It provides smart routing, automatic fallback, load balancing, and format translation across 350 AI providers — all through a single OpenAI-compatible endpoint. Includes a built-in MCP Server (110 tools), A2A v0.3 protocol, Memory/Skills systems, Cloud Agents (codex, cursor, devin, jules), Guardrails framework, and an Electron desktop app.
+> OmniRoute is a free, open-source AI Gateway that acts as a universal API proxy for multi-provider LLMs. It provides smart routing, automatic fallback, load balancing, and format translation across 351 AI providers — all through a single OpenAI-compatible endpoint. Includes a built-in MCP Server (110 tools), A2A v0.3 protocol, Memory/Skills systems, Cloud Agents (codex, cursor, devin, jules), Guardrails framework, and an Electron desktop app.
 
 ## Overview
 
@@ -281,7 +281,7 @@ OmniRoute solves the problem of managing multiple AI provider subscriptions, quo
 ## Key Features (v3.8.50)
 
 ### Core Proxy
-- **350 AI providers** with automatic format translation
+- **351 AI providers** with automatic format translation
 - **Provider categories**: Free (90+ free tiers), OAuth, API Key, Self-Hosted, Custom (OpenAI/Anthropic-compatible)
 - **18 routing strategies**: priority, weighted, round-robin, fill-first, p2c, random, least-used, cost-optimized, reset-aware, reset-window, headroom, strict-random, auto, lkgp, context-optimized, context-relay, fusion, pipeline
 - **4-tier fallback**: Subscription → API Key → Cheap → Free

docs/i18n/phi/llm.txt

@@ -4,7 +4,7 @@
 
 ---
 
-> OmniRoute is a free, open-source AI Gateway that acts as a universal API proxy for multi-provider LLMs. It provides smart routing, automatic fallback, load balancing, and format translation across 350 AI providers — all through a single OpenAI-compatible endpoint. Includes a built-in MCP Server (110 tools), A2A v0.3 protocol, Memory/Skills systems, Cloud Agents (codex, cursor, devin, jules), Guardrails framework, and an Electron desktop app.
+> OmniRoute is a free, open-source AI Gateway that acts as a universal API proxy for multi-provider LLMs. It provides smart routing, automatic fallback, load balancing, and format translation across 351 AI providers — all through a single OpenAI-compatible endpoint. Includes a built-in MCP Server (110 tools), A2A v0.3 protocol, Memory/Skills systems, Cloud Agents (codex, cursor, devin, jules), Guardrails framework, and an Electron desktop app.
 
 ## Overview
 
@@ -281,7 +281,7 @@ OmniRoute solves the problem of managing multiple AI provider subscriptions, quo
 ## Key Features (v3.8.50)
 
 ### Core Proxy
-- **350 AI providers** with automatic format translation
+- **351 AI providers** with automatic format translation
 - **Provider categories**: Free (90+ free tiers), OAuth, API Key, Self-Hosted, Custom (OpenAI/Anthropic-compatible)
 - **18 routing strategies**: priority, weighted, round-robin, fill-first, p2c, random, least-used, cost-optimized, reset-aware, reset-window, headroom, strict-random, auto, lkgp, context-optimized, context-relay, fusion, pipeline
 - **4-tier fallback**: Subscription → API Key → Cheap → Free

docs/i18n/pl/llm.txt

@@ -4,7 +4,7 @@
 
 ---
 
-> OmniRoute is a free, open-source AI Gateway that acts as a universal API proxy for multi-provider LLMs. It provides smart routing, automatic fallback, load balancing, and format translation across 350 AI providers — all through a single OpenAI-compatible endpoint. Includes a built-in MCP Server (110 tools), A2A v0.3 protocol, Memory/Skills systems, Cloud Agents (codex, cursor, devin, jules), Guardrails framework, and an Electron desktop app.
+> OmniRoute is a free, open-source AI Gateway that acts as a universal API proxy for multi-provider LLMs. It provides smart routing, automatic fallback, load balancing, and format translation across 351 AI providers — all through a single OpenAI-compatible endpoint. Includes a built-in MCP Server (110 tools), A2A v0.3 protocol, Memory/Skills systems, Cloud Agents (codex, cursor, devin, jules), Guardrails framework, and an Electron desktop app.
 
 ## Overview
 
@@ -281,7 +281,7 @@ OmniRoute solves the problem of managing multiple AI provider subscriptions, quo
 ## Key Features (v3.8.50)
 
 ### Core Proxy
-- **350 AI providers** with automatic format translation
+- **351 AI providers** with automatic format translation
 - **Provider categories**: Free (90+ free tiers), OAuth, API Key, Self-Hosted, Custom (OpenAI/Anthropic-compatible)
 - **18 routing strategies**: priority, weighted, round-robin, fill-first, p2c, random, least-used, cost-optimized, reset-aware, reset-window, headroom, strict-random, auto, lkgp, context-optimized, context-relay, fusion, pipeline
 - **4-tier fallback**: Subscription → API Key → Cheap → Free

docs/i18n/pt-BR/llm.txt

@@ -4,7 +4,7 @@
 
 ---
 
-> OmniRoute is a free, open-source AI Gateway that acts as a universal API proxy for multi-provider LLMs. It provides smart routing, automatic fallback, load balancing, and format translation across 350 AI providers — all through a single OpenAI-compatible endpoint. Includes a built-in MCP Server (110 tools), A2A v0.3 protocol, Memory/Skills systems, Cloud Agents (codex, cursor, devin, jules), Guardrails framework, and an Electron desktop app.
+> OmniRoute is a free, open-source AI Gateway that acts as a universal API proxy for multi-provider LLMs. It provides smart routing, automatic fallback, load balancing, and format translation across 351 AI providers — all through a single OpenAI-compatible endpoint. Includes a built-in MCP Server (110 tools), A2A v0.3 protocol, Memory/Skills systems, Cloud Agents (codex, cursor, devin, jules), Guardrails framework, and an Electron desktop app.
 
 ## Overview
 
@@ -281,7 +281,7 @@ OmniRoute solves the problem of managing multiple AI provider subscriptions, quo
 ## Key Features (v3.8.50)
 
 ### Core Proxy
-- **350 AI providers** with automatic format translation
+- **351 AI providers** with automatic format translation
 - **Provider categories**: Free (90+ free tiers), OAuth, API Key, Self-Hosted, Custom (OpenAI/Anthropic-compatible)
 - **18 routing strategies**: priority, weighted, round-robin, fill-first, p2c, random, least-used, cost-optimized, reset-aware, reset-window, headroom, strict-random, auto, lkgp, context-optimized, context-relay, fusion, pipeline
 - **4-tier fallback**: Subscription → API Key → Cheap → Free

docs/i18n/pt/llm.txt

@@ -4,7 +4,7 @@
 
 ---
 
-> OmniRoute is a free, open-source AI Gateway that acts as a universal API proxy for multi-provider LLMs. It provides smart routing, automatic fallback, load balancing, and format translation across 350 AI providers — all through a single OpenAI-compatible endpoint. Includes a built-in MCP Server (110 tools), A2A v0.3 protocol, Memory/Skills systems, Cloud Agents (codex, cursor, devin, jules), Guardrails framework, and an Electron desktop app.
+> OmniRoute is a free, open-source AI Gateway that acts as a universal API proxy for multi-provider LLMs. It provides smart routing, automatic fallback, load balancing, and format translation across 351 AI providers — all through a single OpenAI-compatible endpoint. Includes a built-in MCP Server (110 tools), A2A v0.3 protocol, Memory/Skills systems, Cloud Agents (codex, cursor, devin, jules), Guardrails framework, and an Electron desktop app.
 
 ## Overview
 
@@ -281,7 +281,7 @@ OmniRoute solves the problem of managing multiple AI provider subscriptions, quo
 ## Key Features (v3.8.50)
 
 ### Core Proxy
-- **350 AI providers** with automatic format translation
+- **351 AI providers** with automatic format translation
 - **Provider categories**: Free (90+ free tiers), OAuth, API Key, Self-Hosted, Custom (OpenAI/Anthropic-compatible)
 - **18 routing strategies**: priority, weighted, round-robin, fill-first, p2c, random, least-used, cost-optimized, reset-aware, reset-window, headroom, strict-random, auto, lkgp, context-optimized, context-relay, fusion, pipeline
 - **4-tier fallback**: Subscription → API Key → Cheap → Free

docs/i18n/ro/llm.txt

@@ -4,7 +4,7 @@
 
 ---
 
-> OmniRoute is a free, open-source AI Gateway that acts as a universal API proxy for multi-provider LLMs. It provides smart routing, automatic fallback, load balancing, and format translation across 350 AI providers — all through a single OpenAI-compatible endpoint. Includes a built-in MCP Server (110 tools), A2A v0.3 protocol, Memory/Skills systems, Cloud Agents (codex, cursor, devin, jules), Guardrails framework, and an Electron desktop app.
+> OmniRoute is a free, open-source AI Gateway that acts as a universal API proxy for multi-provider LLMs. It provides smart routing, automatic fallback, load balancing, and format translation across 351 AI providers — all through a single OpenAI-compatible endpoint. Includes a built-in MCP Server (110 tools), A2A v0.3 protocol, Memory/Skills systems, Cloud Agents (codex, cursor, devin, jules), Guardrails framework, and an Electron desktop app.
 
 ## Overview
 
@@ -281,7 +281,7 @@ OmniRoute solves the problem of managing multiple AI provider subscriptions, quo
 ## Key Features (v3.8.50)
 
 ### Core Proxy
-- **350 AI providers** with automatic format translation
+- **351 AI providers** with automatic format translation
 - **Provider categories**: Free (90+ free tiers), OAuth, API Key, Self-Hosted, Custom (OpenAI/Anthropic-compatible)
 - **18 routing strategies**: priority, weighted, round-robin, fill-first, p2c, random, least-used, cost-optimized, reset-aware, reset-window, headroom, strict-random, auto, lkgp, context-optimized, context-relay, fusion, pipeline
 - **4-tier fallback**: Subscription → API Key → Cheap → Free

docs/i18n/ru/llm.txt

@@ -4,7 +4,7 @@
 
 ---
 
-> OmniRoute is a free, open-source AI Gateway that acts as a universal API proxy for multi-provider LLMs. It provides smart routing, automatic fallback, load balancing, and format translation across 350 AI providers — all through a single OpenAI-compatible endpoint. Includes a built-in MCP Server (110 tools), A2A v0.3 protocol, Memory/Skills systems, Cloud Agents (codex, cursor, devin, jules), Guardrails framework, and an Electron desktop app.
+> OmniRoute is a free, open-source AI Gateway that acts as a universal API proxy for multi-provider LLMs. It provides smart routing, automatic fallback, load balancing, and format translation across 351 AI providers — all through a single OpenAI-compatible endpoint. Includes a built-in MCP Server (110 tools), A2A v0.3 protocol, Memory/Skills systems, Cloud Agents (codex, cursor, devin, jules), Guardrails framework, and an Electron desktop app.
 
 ## Overview
 
@@ -281,7 +281,7 @@ OmniRoute solves the problem of managing multiple AI provider subscriptions, quo
 ## Key Features (v3.8.50)
 
 ### Core Proxy
-- **350 AI providers** with automatic format translation
+- **351 AI providers** with automatic format translation
 - **Provider categories**: Free (90+ free tiers), OAuth, API Key, Self-Hosted, Custom (OpenAI/Anthropic-compatible)
 - **18 routing strategies**: priority, weighted, round-robin, fill-first, p2c, random, least-used, cost-optimized, reset-aware, reset-window, headroom, strict-random, auto, lkgp, context-optimized, context-relay, fusion, pipeline
 - **4-tier fallback**: Subscription → API Key → Cheap → Free

docs/i18n/sk/llm.txt

@@ -4,7 +4,7 @@
 
 ---
 
-> OmniRoute is a free, open-source AI Gateway that acts as a universal API proxy for multi-provider LLMs. It provides smart routing, automatic fallback, load balancing, and format translation across 350 AI providers — all through a single OpenAI-compatible endpoint. Includes a built-in MCP Server (110 tools), A2A v0.3 protocol, Memory/Skills systems, Cloud Agents (codex, cursor, devin, jules), Guardrails framework, and an Electron desktop app.
+> OmniRoute is a free, open-source AI Gateway that acts as a universal API proxy for multi-provider LLMs. It provides smart routing, automatic fallback, load balancing, and format translation across 351 AI providers — all through a single OpenAI-compatible endpoint. Includes a built-in MCP Server (110 tools), A2A v0.3 protocol, Memory/Skills systems, Cloud Agents (codex, cursor, devin, jules), Guardrails framework, and an Electron desktop app.
 
 ## Overview
 
@@ -281,7 +281,7 @@ OmniRoute solves the problem of managing multiple AI provider subscriptions, quo
 ## Key Features (v3.8.50)
 
 ### Core Proxy
-- **350 AI providers** with automatic format translation
+- **351 AI providers** with automatic format translation
 - **Provider categories**: Free (90+ free tiers), OAuth, API Key, Self-Hosted, Custom (OpenAI/Anthropic-compatible)
 - **18 routing strategies**: priority, weighted, round-robin, fill-first, p2c, random, least-used, cost-optimized, reset-aware, reset-window, headroom, strict-random, auto, lkgp, context-optimized, context-relay, fusion, pipeline
 - **4-tier fallback**: Subscription → API Key → Cheap → Free

docs/i18n/sv/llm.txt

@@ -4,7 +4,7 @@
 
 ---
 
-> OmniRoute is a free, open-source AI Gateway that acts as a universal API proxy for multi-provider LLMs. It provides smart routing, automatic fallback, load balancing, and format translation across 350 AI providers — all through a single OpenAI-compatible endpoint. Includes a built-in MCP Server (110 tools), A2A v0.3 protocol, Memory/Skills systems, Cloud Agents (codex, cursor, devin, jules), Guardrails framework, and an Electron desktop app.
+> OmniRoute is a free, open-source AI Gateway that acts as a universal API proxy for multi-provider LLMs. It provides smart routing, automatic fallback, load balancing, and format translation across 351 AI providers — all through a single OpenAI-compatible endpoint. Includes a built-in MCP Server (110 tools), A2A v0.3 protocol, Memory/Skills systems, Cloud Agents (codex, cursor, devin, jules), Guardrails framework, and an Electron desktop app.
 
 ## Overview
 
@@ -281,7 +281,7 @@ OmniRoute solves the problem of managing multiple AI provider subscriptions, quo
 ## Key Features (v3.8.50)
 
 ### Core Proxy
-- **350 AI providers** with automatic format translation
+- **351 AI providers** with automatic format translation
 - **Provider categories**: Free (90+ free tiers), OAuth, API Key, Self-Hosted, Custom (OpenAI/Anthropic-compatible)
 - **18 routing strategies**: priority, weighted, round-robin, fill-first, p2c, random, least-used, cost-optimized, reset-aware, reset-window, headroom, strict-random, auto, lkgp, context-optimized, context-relay, fusion, pipeline
 - **4-tier fallback**: Subscription → API Key → Cheap → Free

docs/i18n/sw/llm.txt

@@ -4,7 +4,7 @@
 
 ---
 
-> OmniRoute is a free, open-source AI Gateway that acts as a universal API proxy for multi-provider LLMs. It provides smart routing, automatic fallback, load balancing, and format translation across 350 AI providers — all through a single OpenAI-compatible endpoint. Includes a built-in MCP Server (110 tools), A2A v0.3 protocol, Memory/Skills systems, Cloud Agents (codex, cursor, devin, jules), Guardrails framework, and an Electron desktop app.
+> OmniRoute is a free, open-source AI Gateway that acts as a universal API proxy for multi-provider LLMs. It provides smart routing, automatic fallback, load balancing, and format translation across 351 AI providers — all through a single OpenAI-compatible endpoint. Includes a built-in MCP Server (110 tools), A2A v0.3 protocol, Memory/Skills systems, Cloud Agents (codex, cursor, devin, jules), Guardrails framework, and an Electron desktop app.
 
 ## Overview
 
@@ -281,7 +281,7 @@ OmniRoute solves the problem of managing multiple AI provider subscriptions, quo
 ## Key Features (v3.8.50)
 
 ### Core Proxy
-- **350 AI providers** with automatic format translation
+- **351 AI providers** with automatic format translation
 - **Provider categories**: Free (90+ free tiers), OAuth, API Key, Self-Hosted, Custom (OpenAI/Anthropic-compatible)
 - **18 routing strategies**: priority, weighted, round-robin, fill-first, p2c, random, least-used, cost-optimized, reset-aware, reset-window, headroom, strict-random, auto, lkgp, context-optimized, context-relay, fusion, pipeline
 - **4-tier fallback**: Subscription → API Key → Cheap → Free

docs/i18n/ta/llm.txt

@@ -4,7 +4,7 @@
 
 ---
 
-> OmniRoute is a free, open-source AI Gateway that acts as a universal API proxy for multi-provider LLMs. It provides smart routing, automatic fallback, load balancing, and format translation across 350 AI providers — all through a single OpenAI-compatible endpoint. Includes a built-in MCP Server (110 tools), A2A v0.3 protocol, Memory/Skills systems, Cloud Agents (codex, cursor, devin, jules), Guardrails framework, and an Electron desktop app.
+> OmniRoute is a free, open-source AI Gateway that acts as a universal API proxy for multi-provider LLMs. It provides smart routing, automatic fallback, load balancing, and format translation across 351 AI providers — all through a single OpenAI-compatible endpoint. Includes a built-in MCP Server (110 tools), A2A v0.3 protocol, Memory/Skills systems, Cloud Agents (codex, cursor, devin, jules), Guardrails framework, and an Electron desktop app.
 
 ## Overview
 
@@ -281,7 +281,7 @@ OmniRoute solves the problem of managing multiple AI provider subscriptions, quo
 ## Key Features (v3.8.50)
 
 ### Core Proxy
-- **350 AI providers** with automatic format translation
+- **351 AI providers** with automatic format translation
 - **Provider categories**: Free (90+ free tiers), OAuth, API Key, Self-Hosted, Custom (OpenAI/Anthropic-compatible)
 - **18 routing strategies**: priority, weighted, round-robin, fill-first, p2c, random, least-used, cost-optimized, reset-aware, reset-window, headroom, strict-random, auto, lkgp, context-optimized, context-relay, fusion, pipeline
 - **4-tier fallback**: Subscription → API Key → Cheap → Free

docs/i18n/te/llm.txt

@@ -4,7 +4,7 @@
 
 ---
 
-> OmniRoute is a free, open-source AI Gateway that acts as a universal API proxy for multi-provider LLMs. It provides smart routing, automatic fallback, load balancing, and format translation across 350 AI providers — all through a single OpenAI-compatible endpoint. Includes a built-in MCP Server (110 tools), A2A v0.3 protocol, Memory/Skills systems, Cloud Agents (codex, cursor, devin, jules), Guardrails framework, and an Electron desktop app.
+> OmniRoute is a free, open-source AI Gateway that acts as a universal API proxy for multi-provider LLMs. It provides smart routing, automatic fallback, load balancing, and format translation across 351 AI providers — all through a single OpenAI-compatible endpoint. Includes a built-in MCP Server (110 tools), A2A v0.3 protocol, Memory/Skills systems, Cloud Agents (codex, cursor, devin, jules), Guardrails framework, and an Electron desktop app.
 
 ## Overview
 
@@ -281,7 +281,7 @@ OmniRoute solves the problem of managing multiple AI provider subscriptions, quo
 ## Key Features (v3.8.50)
 
 ### Core Proxy
-- **350 AI providers** with automatic format translation
+- **351 AI providers** with automatic format translation
 - **Provider categories**: Free (90+ free tiers), OAuth, API Key, Self-Hosted, Custom (OpenAI/Anthropic-compatible)
 - **18 routing strategies**: priority, weighted, round-robin, fill-first, p2c, random, least-used, cost-optimized, reset-aware, reset-window, headroom, strict-random, auto, lkgp, context-optimized, context-relay, fusion, pipeline
 - **4-tier fallback**: Subscription → API Key → Cheap → Free

docs/i18n/th/llm.txt

@@ -4,7 +4,7 @@
 
 ---
 
-> OmniRoute is a free, open-source AI Gateway that acts as a universal API proxy for multi-provider LLMs. It provides smart routing, automatic fallback, load balancing, and format translation across 350 AI providers — all through a single OpenAI-compatible endpoint. Includes a built-in MCP Server (110 tools), A2A v0.3 protocol, Memory/Skills systems, Cloud Agents (codex, cursor, devin, jules), Guardrails framework, and an Electron desktop app.
+> OmniRoute is a free, open-source AI Gateway that acts as a universal API proxy for multi-provider LLMs. It provides smart routing, automatic fallback, load balancing, and format translation across 351 AI providers — all through a single OpenAI-compatible endpoint. Includes a built-in MCP Server (110 tools), A2A v0.3 protocol, Memory/Skills systems, Cloud Agents (codex, cursor, devin, jules), Guardrails framework, and an Electron desktop app.
 
 ## Overview
 
@@ -281,7 +281,7 @@ OmniRoute solves the problem of managing multiple AI provider subscriptions, quo
 ## Key Features (v3.8.50)
 
 ### Core Proxy
-- **350 AI providers** with automatic format translation
+- **351 AI providers** with automatic format translation
 - **Provider categories**: Free (90+ free tiers), OAuth, API Key, Self-Hosted, Custom (OpenAI/Anthropic-compatible)
 - **18 routing strategies**: priority, weighted, round-robin, fill-first, p2c, random, least-used, cost-optimized, reset-aware, reset-window, headroom, strict-random, auto, lkgp, context-optimized, context-relay, fusion, pipeline
 - **4-tier fallback**: Subscription → API Key → Cheap → Free

docs/i18n/tr/llm.txt

@@ -4,7 +4,7 @@
 
 ---
 
-> OmniRoute is a free, open-source AI Gateway that acts as a universal API proxy for multi-provider LLMs. It provides smart routing, automatic fallback, load balancing, and format translation across 350 AI providers — all through a single OpenAI-compatible endpoint. Includes a built-in MCP Server (110 tools), A2A v0.3 protocol, Memory/Skills systems, Cloud Agents (codex, cursor, devin, jules), Guardrails framework, and an Electron desktop app.
+> OmniRoute is a free, open-source AI Gateway that acts as a universal API proxy for multi-provider LLMs. It provides smart routing, automatic fallback, load balancing, and format translation across 351 AI providers — all through a single OpenAI-compatible endpoint. Includes a built-in MCP Server (110 tools), A2A v0.3 protocol, Memory/Skills systems, Cloud Agents (codex, cursor, devin, jules), Guardrails framework, and an Electron desktop app.
 
 ## Overview
 
@@ -281,7 +281,7 @@ OmniRoute solves the problem of managing multiple AI provider subscriptions, quo
 ## Key Features (v3.8.50)
 
 ### Core Proxy
-- **350 AI providers** with automatic format translation
+- **351 AI providers** with automatic format translation
 - **Provider categories**: Free (90+ free tiers), OAuth, API Key, Self-Hosted, Custom (OpenAI/Anthropic-compatible)
 - **18 routing strategies**: priority, weighted, round-robin, fill-first, p2c, random, least-used, cost-optimized, reset-aware, reset-window, headroom, strict-random, auto, lkgp, context-optimized, context-relay, fusion, pipeline
 - **4-tier fallback**: Subscription → API Key → Cheap → Free

docs/i18n/uk-UA/llm.txt

@@ -4,7 +4,7 @@
 
 ---
 
-> OmniRoute is a free, open-source AI Gateway that acts as a universal API proxy for multi-provider LLMs. It provides smart routing, automatic fallback, load balancing, and format translation across 350 AI providers — all through a single OpenAI-compatible endpoint. Includes a built-in MCP Server (110 tools), A2A v0.3 protocol, Memory/Skills systems, Cloud Agents (codex, cursor, devin, jules), Guardrails framework, and an Electron desktop app.
+> OmniRoute is a free, open-source AI Gateway that acts as a universal API proxy for multi-provider LLMs. It provides smart routing, automatic fallback, load balancing, and format translation across 351 AI providers — all through a single OpenAI-compatible endpoint. Includes a built-in MCP Server (110 tools), A2A v0.3 protocol, Memory/Skills systems, Cloud Agents (codex, cursor, devin, jules), Guardrails framework, and an Electron desktop app.
 
 ## Overview
 
@@ -281,7 +281,7 @@ OmniRoute solves the problem of managing multiple AI provider subscriptions, quo
 ## Key Features (v3.8.50)
 
 ### Core Proxy
-- **350 AI providers** with automatic format translation
+- **351 AI providers** with automatic format translation
 - **Provider categories**: Free (90+ free tiers), OAuth, API Key, Self-Hosted, Custom (OpenAI/Anthropic-compatible)
 - **18 routing strategies**: priority, weighted, round-robin, fill-first, p2c, random, least-used, cost-optimized, reset-aware, reset-window, headroom, strict-random, auto, lkgp, context-optimized, context-relay, fusion, pipeline
 - **4-tier fallback**: Subscription → API Key → Cheap → Free

docs/i18n/ur/llm.txt

@@ -4,7 +4,7 @@
 
 ---
 
-> OmniRoute is a free, open-source AI Gateway that acts as a universal API proxy for multi-provider LLMs. It provides smart routing, automatic fallback, load balancing, and format translation across 350 AI providers — all through a single OpenAI-compatible endpoint. Includes a built-in MCP Server (110 tools), A2A v0.3 protocol, Memory/Skills systems, Cloud Agents (codex, cursor, devin, jules), Guardrails framework, and an Electron desktop app.
+> OmniRoute is a free, open-source AI Gateway that acts as a universal API proxy for multi-provider LLMs. It provides smart routing, automatic fallback, load balancing, and format translation across 351 AI providers — all through a single OpenAI-compatible endpoint. Includes a built-in MCP Server (110 tools), A2A v0.3 protocol, Memory/Skills systems, Cloud Agents (codex, cursor, devin, jules), Guardrails framework, and an Electron desktop app.
 
 ## Overview
 
@@ -281,7 +281,7 @@ OmniRoute solves the problem of managing multiple AI provider subscriptions, quo
 ## Key Features (v3.8.50)
 
 ### Core Proxy
-- **350 AI providers** with automatic format translation
+- **351 AI providers** with automatic format translation
 - **Provider categories**: Free (90+ free tiers), OAuth, API Key, Self-Hosted, Custom (OpenAI/Anthropic-compatible)
 - **18 routing strategies**: priority, weighted, round-robin, fill-first, p2c, random, least-used, cost-optimized, reset-aware, reset-window, headroom, strict-random, auto, lkgp, context-optimized, context-relay, fusion, pipeline
 - **4-tier fallback**: Subscription → API Key → Cheap → Free

docs/i18n/vi/llm.txt

@@ -4,7 +4,7 @@
 
 ---
 
-> OmniRoute is a free, open-source AI Gateway that acts as a universal API proxy for multi-provider LLMs. It provides smart routing, automatic fallback, load balancing, and format translation across 350 AI providers — all through a single OpenAI-compatible endpoint. Includes a built-in MCP Server (110 tools), A2A v0.3 protocol, Memory/Skills systems, Cloud Agents (codex, cursor, devin, jules), Guardrails framework, and an Electron desktop app.
+> OmniRoute is a free, open-source AI Gateway that acts as a universal API proxy for multi-provider LLMs. It provides smart routing, automatic fallback, load balancing, and format translation across 351 AI providers — all through a single OpenAI-compatible endpoint. Includes a built-in MCP Server (110 tools), A2A v0.3 protocol, Memory/Skills systems, Cloud Agents (codex, cursor, devin, jules), Guardrails framework, and an Electron desktop app.
 
 ## Overview
 
@@ -281,7 +281,7 @@ OmniRoute solves the problem of managing multiple AI provider subscriptions, quo
 ## Key Features (v3.8.50)
 
 ### Core Proxy
-- **350 AI providers** with automatic format translation
+- **351 AI providers** with automatic format translation
 - **Provider categories**: Free (90+ free tiers), OAuth, API Key, Self-Hosted, Custom (OpenAI/Anthropic-compatible)
 - **18 routing strategies**: priority, weighted, round-robin, fill-first, p2c, random, least-used, cost-optimized, reset-aware, reset-window, headroom, strict-random, auto, lkgp, context-optimized, context-relay, fusion, pipeline
 - **4-tier fallback**: Subscription → API Key → Cheap → Free

docs/i18n/zh-CN/llm.txt

@@ -4,7 +4,7 @@
 
 ---
 
-> OmniRoute is a free, open-source AI Gateway that acts as a universal API proxy for multi-provider LLMs. It provides smart routing, automatic fallback, load balancing, and format translation across 350 AI providers — all through a single OpenAI-compatible endpoint. Includes a built-in MCP Server (110 tools), A2A v0.3 protocol, Memory/Skills systems, Cloud Agents (codex, cursor, devin, jules), Guardrails framework, and an Electron desktop app.
+> OmniRoute is a free, open-source AI Gateway that acts as a universal API proxy for multi-provider LLMs. It provides smart routing, automatic fallback, load balancing, and format translation across 351 AI providers — all through a single OpenAI-compatible endpoint. Includes a built-in MCP Server (110 tools), A2A v0.3 protocol, Memory/Skills systems, Cloud Agents (codex, cursor, devin, jules), Guardrails framework, and an Electron desktop app.
 
 ## Overview
 
@@ -281,7 +281,7 @@ OmniRoute solves the problem of managing multiple AI provider subscriptions, quo
 ## Key Features (v3.8.50)
 
 ### Core Proxy
-- **350 AI providers** with automatic format translation
+- **351 AI providers** with automatic format translation
 - **Provider categories**: Free (90+ free tiers), OAuth, API Key, Self-Hosted, Custom (OpenAI/Anthropic-compatible)
 - **18 routing strategies**: priority, weighted, round-robin, fill-first, p2c, random, least-used, cost-optimized, reset-aware, reset-window, headroom, strict-random, auto, lkgp, context-optimized, context-relay, fusion, pipeline
 - **4-tier fallback**: Subscription → API Key → Cheap → Free

docs/i18n/zh-TW/llm.txt

@@ -4,7 +4,7 @@
 
 ---
 
-> OmniRoute is a free, open-source AI Gateway that acts as a universal API proxy for multi-provider LLMs. It provides smart routing, automatic fallback, load balancing, and format translation across 350 AI providers — all through a single OpenAI-compatible endpoint. Includes a built-in MCP Server (110 tools), A2A v0.3 protocol, Memory/Skills systems, Cloud Agents (codex, cursor, devin, jules), Guardrails framework, and an Electron desktop app.
+> OmniRoute is a free, open-source AI Gateway that acts as a universal API proxy for multi-provider LLMs. It provides smart routing, automatic fallback, load balancing, and format translation across 351 AI providers — all through a single OpenAI-compatible endpoint. Includes a built-in MCP Server (110 tools), A2A v0.3 protocol, Memory/Skills systems, Cloud Agents (codex, cursor, devin, jules), Guardrails framework, and an Electron desktop app.
 
 ## Overview
 
@@ -281,7 +281,7 @@ OmniRoute solves the problem of managing multiple AI provider subscriptions, quo
 ## Key Features (v3.8.50)
 
 ### Core Proxy
-- **350 AI providers** with automatic format translation
+- **351 AI providers** with automatic format translation
 - **Provider categories**: Free (90+ free tiers), OAuth, API Key, Self-Hosted, Custom (OpenAI/Anthropic-compatible)
 - **18 routing strategies**: priority, weighted, round-robin, fill-first, p2c, random, least-used, cost-optimized, reset-aware, reset-window, headroom, strict-random, auto, lkgp, context-optimized, context-relay, fusion, pipeline
 - **4-tier fallback**: Subscription → API Key → Cheap → Free

docs/reference/ENVIRONMENT.md

@@ -737,6 +737,12 @@ REQUEST_TIMEOUT_MS (global override)
 | `OMNIROUTE_AGENT_GOAL_READINESS_MAX_TIMEOUT_MS` | `600000`      | Maximum first-event readiness window for detected `/goal` agent runs or requests forced with `x-omniroute-agent-goal`.                                          |
 | `OMNIROUTE_AGENT_GOAL_STREAM_RECOVERY`    | `true`               | Enable early stream recovery automatically for detected `/goal` agent runs. Set `false`/`0`/`off` to disable the goal-specific opt-in. This can only ADD recovery on top of the operator default — it never overrides an explicit `STREAM_RECOVERY_ENABLED`/DB settings opt-out. |
 | `OMNIROUTE_CODEX_DROP_NONSTANDARD_EVENTS` | `true`               | Strip non-standard `codex.*` SSE events (e.g. `codex.rate_limits`) that break the OpenAI SDK's `responses.stream()` with a 502. Default ON (#11014). Set `0`/`false`/`no`/`off` to forward them. |
+| `OMNIROUTE_CODEX_APPSERVER_WS`            | _(unset)_            | Opt-in Codex app-server transport. WebSocket endpoint (`ws://`/`wss://`) of a local `codex app-server` sidecar. When set together with a token, Codex requests are routed over JSON-RPC to the sidecar instead of the HTTP Responses API. Also settable per-connection via `providerSpecificData.codexAppServerUrl`. Used by `open-sse/executors/codex/appServerConfig.ts`. |
+| `OMNIROUTE_CODEX_APPSERVER_WS_TOKEN`      | _(unset)_            | Inline capability/bearer token presented to the app-server. Per-connection override: `providerSpecificData.codexAppServerToken`. |
+| `OMNIROUTE_CODEX_APPSERVER_WS_TOKEN_FILE` | _(unset)_            | Path to a file holding the app-server capability token (from `codex app-server --ws-token-file`). Used when `OMNIROUTE_CODEX_APPSERVER_WS_TOKEN` is unset. Per-connection override: `providerSpecificData.codexAppServerTokenFile`. |
+| `OMNIROUTE_CODEX_APPSERVER_CWD`           | `/tmp`               | Working directory the app-server turn runs in. Per-connection override: `providerSpecificData.codexAppServerCwd`. |
+| `OMNIROUTE_CODEX_APPSERVER_APPROVAL`      | _(unset)_            | Approval policy passed to the app-server turn (e.g. `never`, `on-request`). Per-connection override: `providerSpecificData.codexAppServerApprovalPolicy`. |
+| `OMNIROUTE_CODEX_APPSERVER_SANDBOX`       | _(unset)_            | Sandbox policy passed to the app-server turn (e.g. `read-only`, `workspace-write`, `danger-full-access`). Per-connection override: `providerSpecificData.codexAppServerSandbox`. |
 | `FETCH_HEADERS_TIMEOUT_MS`                | = `FETCH_TIMEOUT_MS` | Time to receive response headers.                                                                                                                               |
 | `OMNIROUTE_DIRECT_HEADERS_TIMEOUT_MS`    | `30000` (30s)         | Maximum response-start wait (ms) for each direct no-proxy attempt. A timeout retries once on a fresh socket; set `0` to disable the bound and retain the previous behavior. |
 | `FETCH_BODY_TIMEOUT_MS`                   | = `FETCH_TIMEOUT_MS` | Time to receive the full response body.                                                                                                                         |

docs/reference/PROVIDER_REFERENCE.md

@@ -10,7 +10,7 @@ lastUpdated: 2026-08-23
 > Regenerate with: `npm run gen:provider-reference`
 > **Last generated:** 2026-08-23
 
-Total providers: **350**. See category breakdown below.
+Total providers: **351**. See category breakdown below.
 
 ## Categories
 
@@ -34,7 +34,7 @@ Use the dashboard at `/dashboard/providers` to enable, configure, and test each
 
 ---
 
-## No-auth Providers (no key required) (12)
+## No-auth Providers (no key required) (13)
 
 | ID | Alias | Name | Tags | Website | Notes | Tool calling |
 |----|-------|------|------|---------|-------|--------------|
@@ -42,6 +42,7 @@ Use the dashboard at `/dashboard/providers` to enable, configure, and test each
 | `auggie` | `aug` | Augment (Auggie CLI) | No-auth | [link](https://augmentcode.com) | No API key stored by OmniRoute. Install the Auggie CLI and run `auggie login` on this machine, then OmniRoute spawns it locally for each request. | — |
 | `chipotle` | `pepper` | Chipotle Pepper AI (Free) | No-auth | [link](https://amelia.chipotle.com) | No credentials required. Uses Chipotle's public support chatbot via reverse-engineered SockJS/STOMP protocol. | — |
 | `cloudflare-playground` | `cfp` | Cloudflare AI Playground | No-auth | [link](https://playground.ai.cloudflare.com) | No credentials required — anonymous browser sessions over a reverse-engineered cf_agent WebSocket protocol (Playwright transport). | — |
+| `codex-app-server` | `cxa` | OpenAI Codex (App-Server) | No-auth | [link](https://developers.openai.com/codex/cli) | No token stored by OmniRoute. The Codex CLI app-server manages its own ChatGPT sign-in (~/.codex/auth.json, auto-refreshed). Use “Sign in with ChatGPT” if the CLI is not yet authenticated. | — |
 | `devin-cli-agentic` | `dva` | Devin CLI Agentic Bridge | No-auth | [link](https://docs.devin.ai/work-with-devin/devin-cli) | Authentication is owned by the official Devin CLI in its isolated bridge volume. | emulated |
 | `duckduckgo-web` | `ddgw` | DuckDuckGo AI Chat | No-auth | [link](https://duckduckgo.com/duckchat) | No credentials required — DuckDuckGo AI Chat is anonymous and free. | emulated |
 | `felo-web` | `felo` | Felo | No-auth | [link](https://felo.ai) | No credentials required — Felo is a free, no-signup chat/search aggregator. | — |
@@ -438,7 +439,7 @@ Use the dashboard at `/dashboard/providers` to enable, configure, and test each
 
 - Catalog: [`src/shared/constants/providers.ts`](../../src/shared/constants/providers.ts)
 - Registry (per-model details): [`open-sse/config/providerRegistry.ts`](../../open-sse/config/providerRegistry.ts)
-- Executors: [`open-sse/executors/`](../../open-sse/executors/) (107 implementations)
+- Executors: [`open-sse/executors/`](../../open-sse/executors/) (108 implementations)
 - Translators: [`open-sse/translator/`](../../open-sse/translator/)
 
 ## See Also

llm.txt

@@ -1,6 +1,6 @@
 # OmniRoute
 
-> OmniRoute is a free, open-source AI Gateway that acts as a universal API proxy for multi-provider LLMs. It provides smart routing, automatic fallback, load balancing, and format translation across 350 AI providers — all through a single OpenAI-compatible endpoint. Includes a built-in MCP Server (110 tools), A2A v0.3 protocol, Memory/Skills systems, Cloud Agents (codex, cursor, devin, jules), Guardrails framework, and an Electron desktop app.
+> OmniRoute is a free, open-source AI Gateway that acts as a universal API proxy for multi-provider LLMs. It provides smart routing, automatic fallback, load balancing, and format translation across 351 AI providers — all through a single OpenAI-compatible endpoint. Includes a built-in MCP Server (110 tools), A2A v0.3 protocol, Memory/Skills systems, Cloud Agents (codex, cursor, devin, jules), Guardrails framework, and an Electron desktop app.
 
 ## Overview
 
@@ -277,7 +277,7 @@ OmniRoute solves the problem of managing multiple AI provider subscriptions, quo
 ## Key Features (v3.8.50)
 
 ### Core Proxy
-- **350 AI providers** with automatic format translation
+- **351 AI providers** with automatic format translation
 - **Provider categories**: Free (90+ free tiers), OAuth, API Key, Self-Hosted, Custom (OpenAI/Anthropic-compatible)
 - **18 routing strategies**: priority, weighted, round-robin, fill-first, p2c, random, least-used, cost-optimized, reset-aware, reset-window, headroom, strict-random, auto, lkgp, context-optimized, context-relay, fusion, pipeline
 - **4-tier fallback**: Subscription → API Key → Cheap → Free

open-sse/config/providers/index.ts

@@ -210,6 +210,7 @@ import { baiduProvider } from "./registry/baidu/index.ts";
 import { pollinationsProvider } from "./registry/pollinations/index.ts";
 import { veoaifree_webProvider } from "./registry/veoaifree-web/index.ts";
 import { codexProvider } from "./registry/codex/index.ts";
+import { codexAppServerProvider } from "./registry/codex-app-server/index.ts";
 import { veniceProvider } from "./registry/venice/index.ts";
 import { kiroProvider } from "./registry/kiro/index.ts";
 import { openadapterProvider } from "./registry/openadapter/index.ts";
@@ -476,6 +477,7 @@ export const REGISTRY: Record<string, RegistryEntry> = {
   pollinations: pollinationsProvider,
   "veoaifree-web": veoaifree_webProvider,
   codex: codexProvider,
+  "codex-app-server": codexAppServerProvider,
   venice: veniceProvider,
   kiro: kiroProvider,
   byteplus: byteplusProvider,

open-sse/config/providers/registry/codex-app-server/index.ts

@@ -0,0 +1,36 @@
+import type { RegistryEntry } from "../../shared.ts";
+import { codexProvider } from "../codex/index.ts";
+
+/**
+ * OpenAI Codex — App-Server transport (sibling of the `codex` provider).
+ *
+ * This provider drives the Codex CLI's own `codex app-server` over JSON-RPC/
+ * WebSocket (executor: "codex-app-server"). Unlike the `codex` provider — which
+ * replays the user's ChatGPT/OpenAI OAuth token directly to the Responses API —
+ * the app-server process OWNS and self-refreshes its OpenAI auth
+ * (~/.codex/auth.json), exactly like an interactive `codex` session. OmniRoute
+ * never receives or replays a token, so there is no `authType: "oauth"` and no
+ * usage-caveat: `authType: "none"`.
+ *
+ * The connection target (ws:// URL + capability token) is supplied per-connection
+ * via providerSpecificData (codexAppServerUrl / codexAppServerToken[File]) and
+ * resolved by resolveAppServerConfig — NOT from `baseUrl` below, which is a
+ * documentation sentinel only.
+ *
+ * Models are shared with the `codex` provider (same underlying ChatGPT Codex
+ * backend), imported from codexProvider so the two stay in lockstep.
+ */
+export const codexAppServerProvider: RegistryEntry = {
+  id: "codex-app-server",
+  alias: "cxa",
+  format: "openai-responses",
+  executor: "codex-app-server",
+  // Sentinel: the executor dials the WebSocket app-server URL from
+  // providerSpecificData, not this baseUrl. Kept for catalog/debug display.
+  baseUrl: "codex-app-server://cli/websocket",
+  reasoningTransport: "opaque",
+  authType: "none",
+  authHeader: "none",
+  defaultContextLength: 400000,
+  models: [...codexProvider.models],
+};

open-sse/executors/codex-app-server.ts

@@ -0,0 +1,448 @@
+import {
+  bridgeToResponsesSSE,
+  buildResponseJSON,
+} from "../vendor/codex-chatgpt-web/bridge.ts";
+import { AsyncEventQueue } from "../vendor/codex-chatgpt-web/event-queue.ts";
+import type { AdapterEvent } from "../vendor/codex-chatgpt-web/types.ts";
+import { sanitizeErrorMessage } from "../utils/error.ts";
+import { PROVIDERS } from "../config/constants.ts";
+import { BaseExecutor, type ExecuteInput, type ExecutorExecuteResult } from "./base.ts";
+import {
+  CodexAppServerClient,
+  type CodexAppServerClientOptions,
+} from "./codex/appServerClient.ts";
+import { resolveAppServerConfig, type CodexAppServerConfig } from "./codex/appServerConfig.ts";
+import {
+  translateNotification,
+  translateToolCall,
+  type DynamicToolCallLike,
+} from "./codex/appServerEvents.ts";
+
+const JSON_HEADERS = { "Content-Type": "application/json" };
+const SSE_HEADERS = {
+  "Cache-Control": "no-cache, no-transform",
+  Connection: "keep-alive",
+  "Content-Type": "text/event-stream; charset=utf-8",
+};
+
+/** A single text UserInput as accepted by turn/start (text_elements is required). */
+interface CodexTextUserInput {
+  type: "text";
+  text: string;
+  text_elements: [];
+}
+
+/**
+ * Flatten an OpenAI Responses request body into the plain prompt text the
+ * app-server turn expects. The body's `input` is a string, a single message item,
+ * or an array of message items with `content` parts; we concatenate the user-facing
+ * text. This is intentionally lossless-enough for a text turn (images/tool parts are
+ * out of scope for the initial app-server transport).
+ */
+export function extractPromptText(body: unknown): string {
+  if (!body || typeof body !== "object") return "";
+  const input = (body as Record<string, unknown>).input;
+  if (typeof input === "string") return input;
+  if (input == null) return "";
+  const items = Array.isArray(input) ? input : [input];
+  const chunks: string[] = [];
+  for (const item of items) {
+    collectText(item, chunks);
+  }
+  return chunks.join("\n").trim();
+}
+
+function collectText(item: unknown, out: string[]): void {
+  if (typeof item === "string") {
+    if (item.length > 0) out.push(item);
+    return;
+  }
+  if (!item || typeof item !== "object") return;
+  const rec = item as Record<string, unknown>;
+  if (typeof rec.text === "string" && rec.text.length > 0) {
+    out.push(rec.text);
+    return;
+  }
+  const content = rec.content;
+  if (typeof content === "string") {
+    if (content.length > 0) out.push(content);
+    return;
+  }
+  if (Array.isArray(content)) {
+    for (const part of content) {
+      if (part && typeof part === "object") {
+        const text = (part as Record<string, unknown>).text;
+        if (typeof text === "string" && text.length > 0) out.push(text);
+      } else if (typeof part === "string" && part.length > 0) {
+        out.push(part);
+      }
+    }
+  }
+}
+
+/** Optional reasoning effort carried on the Responses body (`reasoning.effort`). */
+function extractEffort(body: unknown): string | undefined {
+  if (!body || typeof body !== "object") return undefined;
+  const reasoning = (body as Record<string, unknown>).reasoning;
+  if (reasoning && typeof reasoning === "object") {
+    const effort = (reasoning as Record<string, unknown>).effort;
+    if (typeof effort === "string" && effort.length > 0) return effort;
+  }
+  return undefined;
+}
+
+/** A codex app-server DynamicToolSpec (experimental-api) advertised on thread/start. */
+interface DynamicToolFunctionSpec {
+  type: "function";
+  name: string;
+  description: string;
+  inputSchema: Record<string, unknown>;
+}
+
+interface AppServerToolMaps {
+  /** wireName -> {namespace, name} for restoring MCP namespaced calls in the bridge. */
+  namespace: Map<string, { namespace: string; name: string }>;
+  /** wireNames the bridge must relay as custom_tool_call (freeform, e.g. apply_patch). */
+  freeform: Set<string>;
+  /** wireNames the bridge must relay as tool_search_call. */
+  toolSearch: Set<string>;
+  /** DynamicToolSpecs to advertise to codex on thread/start (experimental-api). */
+  specs: DynamicToolFunctionSpec[];
+}
+
+const EMPTY_OBJECT_SCHEMA: Record<string, unknown> = { type: "object", properties: {} };
+const FREEFORM_INPUT_SCHEMA: Record<string, unknown> = {
+  type: "object",
+  properties: { input: { type: "string", description: "Raw tool input." } },
+  required: ["input"],
+};
+
+function asRecord(v: unknown): Record<string, unknown> | null {
+  return v && typeof v === "object" && !Array.isArray(v) ? (v as Record<string, unknown>) : null;
+}
+
+/**
+ * Build the bridge tool maps + the codex dynamicTools specs from the harness's
+ * Responses `tools` array. This mirrors chatgpt-web-codex.ts:toolMaps() /
+ * parser.ts:buildTools(): every harness tool is exposed to codex FLAT under its
+ * wire name ("<namespace>__<name>" for MCP tools) so the round-trip is
+ * namespace-preserving (codex echoes the call via item/tool/call; the bridge
+ * restores {namespace, name} from `toolNsMap`). Custom (freeform) and tool_search
+ * tools are tracked so the bridge relays them as custom_tool_call / tool_search_call.
+ */
+function buildAppServerToolMaps(body: unknown): AppServerToolMaps {
+  const namespace = new Map<string, { namespace: string; name: string }>();
+  const freeform = new Set<string>();
+  const toolSearch = new Set<string>();
+  const specs: DynamicToolFunctionSpec[] = [];
+
+  const rec = asRecord(body);
+  const tools = rec && Array.isArray(rec.tools) ? (rec.tools as unknown[]) : [];
+
+  const pushFn = (name: string, description: string, inputSchema: Record<string, unknown>) => {
+    specs.push({ type: "function", name, description, inputSchema });
+  };
+
+  for (const raw of tools) {
+    const t = asRecord(raw);
+    if (!t) continue;
+    const type = t.type;
+    const desc = typeof t.description === "string" ? t.description : "";
+
+    if (type === "function" && typeof t.name === "string") {
+      const wireName = t.name;
+      pushFn(wireName, desc, asRecord(t.parameters) ?? EMPTY_OBJECT_SCHEMA);
+    } else if (type === "namespace" && Array.isArray(t.tools) && typeof t.name === "string") {
+      const ns = t.name;
+      for (const innerRaw of t.tools as unknown[]) {
+        const inner = asRecord(innerRaw);
+        if (inner && inner.type === "function" && typeof inner.name === "string") {
+          const wireName = `${ns}__${inner.name}`;
+          namespace.set(wireName, { namespace: ns, name: inner.name });
+          const innerDesc = typeof inner.description === "string" ? inner.description : "";
+          pushFn(wireName, innerDesc, asRecord(inner.parameters) ?? EMPTY_OBJECT_SCHEMA);
+        }
+      }
+    } else if (type === "custom" && typeof t.name === "string") {
+      const wireName = t.name;
+      freeform.add(wireName);
+      pushFn(wireName, desc, FREEFORM_INPUT_SCHEMA);
+    } else if (type === "tool_search") {
+      const wireName = "tool_search";
+      toolSearch.add(wireName);
+      pushFn(
+        wireName,
+        desc || "Search for additional tools to load for the next turn.",
+        asRecord(t.parameters) ?? {
+          type: "object",
+          properties: { query: { type: "string" }, limit: { type: "number" } },
+          required: ["query"],
+        }
+      );
+    } else if (
+      typeof t.name === "string" &&
+      type !== "web_search" &&
+      type !== "image_generation" &&
+      type !== "web_search_preview"
+    ) {
+      // Any other named, client-executed tool → pass through as a function so the
+      // routed model can call it; the bridge relays its call as a function_call.
+      pushFn(t.name, desc, asRecord(t.parameters) ?? EMPTY_OBJECT_SCHEMA);
+    }
+    // web_search / image_generation are OpenAI-hosted server-side tools — not relayable.
+  }
+
+  return { namespace, freeform, toolSearch, specs };
+}
+
+/**
+ * Executor for the Codex app-server WS transport. Drives one turn against a local
+ * `codex app-server` over JSON-RPC and re-emits its notifications as OpenAI
+ * Responses SSE via the shared bridge.
+ *
+ * Errors are delivered IN-BAND (an `error` AdapterEvent → `response.failed` SSE
+ * frame for streaming, or an error field in the JSON body for non-streaming),
+ * never thrown out of execute().
+ */
+export class CodexAppServerExecutor extends BaseExecutor {
+  private readonly clientOptions: CodexAppServerClientOptions;
+
+  /**
+   * @param clientOptions transport options (websocketFn, timeouts).
+   * @param providerId which provider identity this executor reports as. Defaults
+   *   to "codex" so the existing per-connection `codexTransport==="app-server"`
+   *   flag path (routed through CodexExecutor for the `codex` provider) keeps its
+   *   original identity. The first-class `codex-app-server` sibling passes
+   *   "codex-app-server" so logs/quota scoping and the golden executor map reflect
+   *   the real provider. Falls back to PROVIDERS.codex when the sibling registry
+   *   entry is not present (defensive; both share the codex backend).
+   */
+  constructor(clientOptions: CodexAppServerClientOptions = {}, providerId = "codex") {
+    super(providerId, PROVIDERS[providerId] ?? PROVIDERS.codex);
+    this.clientOptions = clientOptions;
+  }
+
+  override async execute(input: ExecuteInput): Promise<ExecutorExecuteResult> {
+    const psd = input.credentials?.providerSpecificData;
+    const config = resolveAppServerConfig(psd);
+    if (!config) {
+      return errorResponse(
+        503,
+        "Codex app-server transport is not configured (missing url or token)",
+        "codex_app_server_unconfigured"
+      );
+    }
+
+    const promptText = extractPromptText(input.body);
+    const effort = extractEffort(input.body);
+    const toolMaps = buildAppServerToolMaps(input.body);
+    const hasTools = toolMaps.specs.length > 0;
+    const events = new AsyncEventQueue<AdapterEvent>();
+    const client = new CodexAppServerClient(this.clientOptions);
+
+    const run = async () => {
+      let terminated = false;
+      // Resolves when the turn reaches a terminal state (turn/completed, error,
+      // or an item/tool/call passthrough). `turn/start` resolving only means the
+      // turn was ACCEPTED (status: inProgress) — the model's output arrives later
+      // as notifications. run() MUST await this before the finally-block closes
+      // the client, otherwise the socket is torn down mid-turn and the event
+      // queue never receives its terminal event (the request then hangs until the
+      // caller's timeout). See translateNotification: it returns true on the
+      // terminal notification, which is where we settle this.
+      let settleTurn!: () => void;
+      const turnDone = new Promise<void>((resolve) => {
+        settleTurn = resolve;
+      });
+      const markTerminated = () => {
+        if (terminated) return;
+        terminated = true;
+        settleTurn();
+      };
+      const finishTurn = () => {
+        if (terminated) return;
+        events.push({ type: "done", endTurn: true });
+        events.close();
+        markTerminated();
+      };
+      try {
+        await client.connect(config.url, config.token);
+        await client.request("initialize", {
+          clientInfo: {
+            name: "omniroute-codex-app-server",
+            title: null,
+            version: "1.0",
+          },
+          // Harness function tools are advertised via thread/start's `dynamicTools`,
+          // which is an EXPERIMENTAL app-server field: opt into experimental API so
+          // codex accepts it (and can emit the item/tool/call ServerRequest).
+          capabilities: hasTools
+            ? { experimentalApi: true, requestAttestation: false }
+            : null,
+        });
+        const threadResult = (await client.request("thread/start", {
+          cwd: config.cwd,
+          // OmniRoute is a router: the HARNESS that consumes OmniRoute owns tool
+          // execution and policy. codex must therefore NEVER block a turn waiting
+          // on its own interactive approval, and its own sandbox must not gate the
+          // model — the harness decides what actually runs. So we pair
+          // approvalPolicy:"never" (non-interactive; codex never prompts) with
+          // sandbox:"danger-full-access" (codex's own sandbox imposes no
+          // restriction), mirroring codexInstructions.ts:50 ("never +
+          // danger-full-access = take advantage of it"). Any server→client
+          // approval request that still arrives is auto-APPROVED by the client
+          // (see CodexAppServerClient), never denied — denial would sabotage the
+          // harness's tool calls. Callers can override both via providerSpecificData.
+          approvalPolicy: config.approvalPolicy ?? "never",
+          sandbox: config.sandbox ?? "danger-full-access",
+          // INBOUND harness tools → codex. The client tells the app-server which
+          // function tools are available for the thread via the `dynamicTools`
+          // field on thread/start (a DynamicToolSpec[] under the experimental API,
+          // verified from the real codex binary; see appServerEvents.ts). codex
+          // then invokes them by sending the `item/tool/call` ServerRequest back
+          // to the client (DynamicToolCallParams), which we PASS THROUGH.
+          ...(hasTools ? { dynamicTools: toolMaps.specs } : {}),
+        })) as { thread?: { id?: unknown }; threadId?: unknown };
+        // The live app-server (codex 0.149.0) returns the thread under
+        // result.thread.id — NOT a top-level threadId (verified against the real
+        // binary 2026-08-22). Keep the top-level fallback for forward/back compat.
+        const threadId =
+          threadResult && typeof threadResult.thread?.id === "string"
+            ? threadResult.thread.id
+            : threadResult && typeof threadResult.threadId === "string"
+              ? threadResult.threadId
+              : "";
+
+        client.onNotification((method, params) => {
+          if (terminated) return;
+          const isTerminal = translateNotification(method, params, (event) => events.push(event));
+          if (isTerminal) {
+            events.close();
+            markTerminated();
+          }
+        });
+
+        // OUTBOUND codex tool call → harness. codex asks us to execute a harness
+        // tool via the `item/tool/call` ServerRequest. OmniRoute is a STATELESS
+        // ROUTER and CANNOT execute the harness's tool (the tool body lives in the
+        // harness downstream). So we PASS IT THROUGH: emit tool_call_* AdapterEvents
+        // (the bridge renders a Responses function_call / custom_tool_call /
+        // tool_search_call), settle the app-server request with a benign
+        // DynamicToolCallResponse so codex does not hang, and COMPLETE the turn.
+        // The harness runs the tool and replays the result in a fresh /v1/responses
+        // request (the stateless-full-history contract every OmniRoute provider uses).
+        client.onToolCall((_id, params, api) => {
+          if (terminated) return;
+          const toolParams = (params && typeof params === "object" ? params : {}) as DynamicToolCallLike;
+          translateToolCall(toolParams, (event) => events.push(event));
+          // Settle the app-server request so the socket does not stall. The router
+          // does not have the tool output (the harness will produce it next turn),
+          // so we report the passthrough as an unsuccessful in-line result and end
+          // the turn — the function_call has already been surfaced to the harness.
+          api.respond({
+            contentItems: [
+              {
+                type: "inputText",
+                text: "router: tool executed by harness; call surfaced as function_call",
+              },
+            ],
+            success: false,
+          });
+          finishTurn();
+        });
+
+        const onAbort = () => {
+          try {
+            client.notify("turn/interrupt", { threadId, turnId: "" });
+          } catch {
+            /* interrupt best-effort */
+          }
+          // Unblock run() so the finally-block can tear down the client. Without
+          // this, an aborted request would wait on turnDone until the terminal
+          // notification that will never come.
+          if (!terminated) {
+            events.close();
+            markTerminated();
+          }
+        };
+        input.signal?.addEventListener("abort", onAbort, { once: true });
+
+        const turnInput: CodexTextUserInput[] = [
+          { type: "text", text: promptText, text_elements: [] },
+        ];
+        await client.request("turn/start", {
+          threadId,
+          input: turnInput,
+          model: input.model,
+          ...(effort ? { effort } : {}),
+        });
+        // `turn/start` resolving only ACCEPTS the turn (status: inProgress). The
+        // model's output (agentMessage deltas) and the terminal turn/completed
+        // arrive AFTER, as notifications. Wait for the terminal signal before
+        // falling through to the finally-block — otherwise client.close() tears
+        // down the socket mid-turn and the queue never closes (request hangs).
+        await turnDone;
+      } catch (err) {
+        if (!terminated) {
+          events.push({
+            type: "error",
+            message: sanitizeErrorMessage(err instanceof Error ? err.message : err),
+            status: 502,
+            errorType: "provider_error",
+            code: "codex_app_server_turn_failed",
+          });
+          events.close();
+          markTerminated();
+        }
+      } finally {
+        client.close();
+      }
+    };
+
+    if (!input.stream) {
+      const running = run();
+      const collected = await events.collect();
+      await running;
+      const response = buildResponseJSON(collected, input.model, {
+        toolNsMap: toolMaps.namespace,
+        freeformToolNames: toolMaps.freeform,
+        toolSearchToolNames: toolMaps.toolSearch,
+      });
+      return {
+        response: new Response(JSON.stringify(response), { status: 200, headers: JSON_HEADERS }),
+        url: config.url,
+      };
+    }
+
+    void run();
+    const stream = bridgeToResponsesSSE(
+      events,
+      input.model,
+      toolMaps.namespace,
+      toolMaps.freeform,
+      toolMaps.toolSearch,
+      () => client.close(),
+      2_000
+    );
+    return {
+      response: new Response(stream, { status: 200, headers: SSE_HEADERS }),
+      url: config.url,
+    };
+  }
+}
+
+function errorResponse(status: number, message: string, code: string): Response {
+  return new Response(
+    JSON.stringify({
+      error: {
+        code,
+        message: sanitizeErrorMessage(message),
+        type: status >= 500 ? "provider_error" : "invalid_request_error",
+      },
+    }),
+    { status, headers: JSON_HEADERS }
+  );
+}
+
+// re-export config type for consumers/tests
+export type { CodexAppServerConfig };

open-sse/executors/codex.ts

@@ -58,6 +58,8 @@ import {
   type CodexEffortLevel as EffortLevel,
 } from "./codex/reasoningSuffix.ts";
 import { repairMissingCodexToolCallOutputs } from "./codex/toolCallRepair.ts";
+import { resolveAppServerConfig } from "./codex/appServerConfig.ts";
+import { CodexAppServerExecutor } from "./codex-app-server.ts";
 // Re-exported for external importers (tests + provider services).
 export { isCodexFreePlan, normalizeCodexTools } from "./codex/tools.ts";
 
@@ -102,6 +104,12 @@ export function __setCodexWebSocketTransportForTesting(
   _websocketOverride = websocket;
 }
 
+// Exposed for the app-server transport, which needs the same wreq-js websocket
+// factory (with the testing override honored) to open its JSON-RPC socket.
+export function getCodexAppServerWebsocketTransport(): WebsocketFn | null {
+  return getCodexWebSocketTransport();
+}
+
 function codexWebSocketUnavailableResponse(): Response {
   return new Response(
     JSON.stringify({
@@ -395,6 +403,34 @@ function isCodexWsGloballyEnabled(): boolean {
   }
 }
 
+/**
+ * Global Codex app-server kill-switch (feature flag OMNIROUTE_CODEX_APP_SERVER_ENABLED,
+ * default ON). Fail-open, mirroring isCodexWsGloballyEnabled.
+ */
+function isCodexAppServerGloballyEnabled(): boolean {
+  try {
+    return isFeatureFlagEnabled("OMNIROUTE_CODEX_APP_SERVER_ENABLED");
+  } catch {
+    return true;
+  }
+}
+
+/**
+ * True when the connection opted into the app-server transport
+ * (providerSpecificData.codexTransport === "app-server") AND the app-server is
+ * configured (URL + token resolvable) AND the global flag is on. Selected BEFORE
+ * the websocket check so it wins when configured.
+ */
+export function isCodexAppServerRequired(credentials: unknown): boolean {
+  if (!isCodexAppServerGloballyEnabled()) return false;
+  const providerSpecificData =
+    credentials && typeof credentials === "object"
+      ? (credentials as { providerSpecificData?: Record<string, unknown> }).providerSpecificData
+      : null;
+  if (providerSpecificData?.codexTransport !== "app-server") return false;
+  return !!resolveAppServerConfig(providerSpecificData);
+}
+
 export function isCodexResponsesWebSocketRequired(_model: string, credentials: unknown): boolean {
   // Global kill-switch (default ON). When disabled, Codex never uses the WS
   // transport — even per-connection codexTransport=websocket falls back to the
@@ -760,6 +796,8 @@ function normalizeCodexWsHeaders(headers: Record<string, string>): Record<string
  * IMPORTANT: Includes chatgpt-account-id header for workspace binding.
  */
 export class CodexExecutor extends BaseExecutor {
+  private appServer: CodexAppServerExecutor | null = null;
+
   constructor() {
     super("codex", PROVIDERS.codex);
   }
@@ -778,6 +816,15 @@ export class CodexExecutor extends BaseExecutor {
     );
     const nextInput = { ...requestInput, credentials };
 
+    if (isCodexAppServerRequired(nextInput.credentials)) {
+      if (!this.appServer) {
+        this.appServer = new CodexAppServerExecutor({
+          websocketFn: getCodexAppServerWebsocketTransport(),
+        });
+      }
+      return this.appServer.execute(nextInput);
+    }
+
     if (!isCodexResponsesWebSocketRequired(nextInput.model, nextInput.credentials)) {
       const httpResult = await super.execute(nextInput);
       if (codexDropNonstandardEvents()) {

open-sse/executors/codex/appServerAuthProbe.ts

@@ -0,0 +1,102 @@
+/**
+ * Layer-2 auth-status probe for the Codex app-server transport.
+ *
+ * The HTTP `/readyz` endpoint only proves the app-server PROCESS is up — not that
+ * its Codex CLI is signed in. A public user whose CLI is not yet authenticated
+ * would otherwise see a green "ready" badge and then fail on the first turn with
+ * an upstream auth error. This probe opens the same JSON-RPC/WebSocket the
+ * executor uses and calls `account/read` (verified against codex 0.149.0): an
+ * authenticated server returns `{ account: { type, email, planType }, ... }`;
+ * a logged-out server returns no account (or an error). So the presence of
+ * `result.account` is the "authenticated" signal.
+ *
+ * Kept separate from the executor turn path so the health check pulls in only the
+ * lightweight client + transport, and so it is independently unit-testable with a
+ * fake websocketFn.
+ */
+import {
+  CodexAppServerClient,
+  type CodexAppServerWebsocketFn,
+} from "./appServerClient.ts";
+import type { CodexAppServerConfig } from "./appServerConfig.ts";
+
+export type CodexAppServerAuthStatus =
+  | { state: "authenticated"; account: { type?: string; email?: string; planType?: string } }
+  | { state: "logged_out"; reason: string }
+  | { state: "unknown"; reason: string };
+
+interface AccountReadResult {
+  account?: { type?: unknown; email?: unknown; planType?: unknown } | null;
+  requiresOpenaiAuth?: unknown;
+}
+
+function str(v: unknown): string | undefined {
+  return typeof v === "string" && v.length > 0 ? v : undefined;
+}
+
+/**
+ * Open a short-lived WS to the app-server, initialize, and read the account.
+ * Returns an auth status; never throws (maps failures to state "unknown").
+ *
+ * @param config resolved app-server config (url + capability token).
+ * @param websocketFn the wreq-js websocket factory
+ *   (getCodexAppServerWebsocketTransport()); when null, returns "unknown".
+ * @param timeoutMs overall budget for connect + account/read.
+ */
+export async function probeCodexAppServerAuth(
+  config: CodexAppServerConfig,
+  websocketFn: CodexAppServerWebsocketFn | null,
+  timeoutMs = 8000
+): Promise<CodexAppServerAuthStatus> {
+  if (!websocketFn) {
+    return { state: "unknown", reason: "websocket transport unavailable" };
+  }
+  const client = new CodexAppServerClient({ websocketFn, defaultTimeoutMs: timeoutMs });
+  const deadline = new Promise<CodexAppServerAuthStatus>((resolve) =>
+    setTimeout(() => resolve({ state: "unknown", reason: "auth probe timed out" }), timeoutMs)
+  );
+
+  const run = (async (): Promise<CodexAppServerAuthStatus> => {
+    try {
+      await client.connect(config.url, config.token);
+      await client.request(
+        "initialize",
+        {
+          clientInfo: { name: "omniroute-codex-app-server-health", title: null, version: "1.0" },
+          capabilities: null,
+        },
+        timeoutMs
+      );
+      // account/read: authenticated → { account: {...} }; logged out → no account.
+      const result = (await client.request("account/read", {}, timeoutMs)) as AccountReadResult;
+      const account = result?.account;
+      if (account && typeof account === "object") {
+        return {
+          state: "authenticated",
+          account: {
+            type: str(account.type),
+            email: str(account.email),
+            planType: str(account.planType),
+          },
+        };
+      }
+      return {
+        state: "logged_out",
+        reason: "app-server reachable but its Codex CLI is not signed in",
+      };
+    } catch (err) {
+      const message = err instanceof Error ? err.message : String(err);
+      // A JSON-RPC error on account/read (e.g. AuthRequiredError) also means
+      // "up but not authenticated" — surface it as logged_out, not unknown, so
+      // the dashboard offers "Sign in with ChatGPT" rather than a scary error.
+      if (/auth|login|sign|unauthor|401/i.test(message)) {
+        return { state: "logged_out", reason: message };
+      }
+      return { state: "unknown", reason: message };
+    } finally {
+      client.close();
+    }
+  })();
+
+  return Promise.race([run, deadline]);
+}

open-sse/executors/codex/appServerClient.ts

@@ -0,0 +1,289 @@
+/**
+ * Id-correlated JSON-RPC 2.0 client over a single WebSocket, for the Codex
+ * app-server transport.
+ *
+ * Ported from the stdio JSON-RPC pattern in `devin-cli-agentic.ts` (monotonic id,
+ * pending-request map settled on responses, notification vs response
+ * discrimination, settle-once) onto the wreq-js WebSocket transport used by the
+ * existing Codex WS path.
+ *
+ * The critical addition over the other transports is a catch-all handler for
+ * server -> client ServerRequests: the app-server can ask the client to approve a
+ * command / patch / permission. OmniRoute is a ROUTER — the harness that consumes
+ * it owns tool execution and policy — so codex must never stall a turn on its own
+ * interactive approval. Every inbound ServerRequest is always answered: approval
+ * prompts are auto-APPROVED (so the model's agentic tool calls proceed; the harness
+ * decides what really runs), and anything else we can't service gets a JSON-RPC
+ * error so the id is always settled and the turn never hangs.
+ */
+
+// wreq-js WebSocket surface (mirrors the private type in codex.ts:71-77).
+export type CodexWreqWebSocket = {
+  send: (data: string) => void;
+  close: (code?: number, reason?: string) => void;
+  onmessage: ((event: { data: unknown }) => void) | null;
+  onerror: ((event: { message?: string }) => void) | null;
+  onclose: (() => void) | null;
+};
+
+export type CodexAppServerWebsocketFn = (
+  url: string,
+  opts?: Record<string, unknown>
+) => Promise<CodexWreqWebSocket>;
+
+interface PendingReq {
+  resolve: (result: unknown) => void;
+  reject: (err: Error) => void;
+}
+
+// The set of ServerRequest methods that are approval prompts (see PROTOCOL-DIGEST
+// "Server -> client REQUESTS"). All of these get an auto-denial decision.
+const APPROVAL_REQUEST_METHODS = new Set<string>([
+  "item/commandExecution/requestApproval",
+  "item/fileChange/requestApproval",
+  "item/permissions/requestApproval",
+  "applyPatchApproval",
+  "execCommandApproval",
+]);
+
+const ROUTER_APPROVAL_NOTE = "router: harness-controlled execution";
+
+export interface CodexAppServerClientOptions {
+  /** Transport factory. Defaults to the shared wreq-js websocket() when omitted. */
+  websocketFn?: CodexAppServerWebsocketFn | null;
+  /** Default per-request timeout (ms). */
+  defaultTimeoutMs?: number;
+}
+
+/**
+ * The app-server → client REQUEST method by which codex invokes a harness-defined
+ * (dynamic) function tool. See appServerEvents.ts:CODEX_APPSERVER_TOOL_CALL_METHOD.
+ * A stateless router cannot execute the harness's tool, so this is handled by a
+ * PASSTHROUGH handler (surface it as a Responses function_call and complete the
+ * turn) rather than by the default -32601 rejection.
+ */
+const TOOL_CALL_REQUEST_METHOD = "item/tool/call";
+
+/**
+ * Handler for a server → client `item/tool/call` ServerRequest. It receives the
+ * JSON-RPC id and raw params (DynamicToolCallParams). It OWNS settling the id
+ * (call `respond`/`respondError`) so the socket never hangs. Returning lets the
+ * executor emit tool_call_* AdapterEvents + complete the turn.
+ */
+export type CodexAppServerToolCallHandler = (
+  id: number,
+  params: unknown,
+  api: {
+    /** Settle the request id with a JSON-RPC result (a DynamicToolCallResponse). */
+    respond: (result: unknown) => void;
+    /** Settle the request id with a JSON-RPC error. */
+    respondError: (code: number, message: string) => void;
+  }
+) => void;
+
+export class CodexAppServerClient {
+  private ws: CodexWreqWebSocket | null = null;
+  private nextId = 1;
+  private readonly pending = new Map<number, PendingReq>();
+  private notificationHandler: (method: string, params: unknown) => void = () => {};
+  private toolCallHandler: CodexAppServerToolCallHandler | null = null;
+  private readonly websocketFn: CodexAppServerWebsocketFn | null;
+  private readonly defaultTimeoutMs: number;
+  private closed = false;
+
+  constructor(options: CodexAppServerClientOptions = {}) {
+    this.websocketFn = options.websocketFn ?? null;
+    this.defaultTimeoutMs = options.defaultTimeoutMs ?? 120_000;
+  }
+
+  /**
+   * Open the WebSocket and attach the capability token as `Authorization: Bearer`.
+   * Do NOT add any chatgpt.com Origin/WS header normalization here — the local
+   * app-server wants only the Authorization header.
+   */
+  async connect(url: string, token: string): Promise<void> {
+    if (!this.websocketFn) {
+      throw new Error("Codex app-server websocket transport unavailable");
+    }
+    // wreq-js's websocket() REQUIRES a browser/os impersonation profile alongside
+    // headers — the same shape the existing Codex WS path uses (codex.ts:980).
+    // Omitting browser/os makes the native call hang/throw, so the app-server
+    // turn never connects. The local app-server ignores the impersonation
+    // fingerprint; only the Authorization bearer matters for its ws-auth.
+    this.ws = await this.websocketFn(url, {
+      browser: "chrome_142",
+      os: "windows",
+      headers: { Authorization: `Bearer ${token}` },
+    });
+    this.ws.onmessage = (event) => this.onFrame(event.data);
+    this.ws.onerror = (event) => this.failAll(event?.message ?? "app-server socket error");
+    this.ws.onclose = () => this.failAll("app-server connection closed");
+  }
+
+  /** Send a ClientRequest and resolve when its id-matched response arrives. */
+  request<T = unknown>(method: string, params: unknown, timeoutMs = this.defaultTimeoutMs): Promise<T> {
+    const id = this.nextId++;
+    return new Promise<T>((resolve, reject) => {
+      if (!this.ws || this.closed) {
+        reject(new Error(`Cannot send ${method}: app-server connection is not open`));
+        return;
+      }
+      const timer = setTimeout(() => {
+        this.pending.delete(id);
+        reject(new Error(`Codex app-server request "${method}" timed out after ${timeoutMs}ms`));
+      }, timeoutMs);
+      this.pending.set(id, {
+        resolve: (result) => {
+          clearTimeout(timer);
+          resolve(result as T);
+        },
+        reject: (err) => {
+          clearTimeout(timer);
+          reject(err);
+        },
+      });
+      this.ws.send(JSON.stringify({ jsonrpc: "2.0", id, method, params }));
+    });
+  }
+
+  /** Send a ClientNotification (no id, no reply expected — e.g. turn/interrupt). */
+  notify(method: string, params: unknown): void {
+    if (!this.ws || this.closed) return;
+    this.ws.send(JSON.stringify({ jsonrpc: "2.0", method, params }));
+  }
+
+  /** Register the handler that receives server -> client NOTIFICATIONS (no id). */
+  onNotification(fn: (method: string, params: unknown) => void): void {
+    this.notificationHandler = fn;
+  }
+
+  /**
+   * Register the handler for the `item/tool/call` server → client ServerRequest
+   * (a harness function-tool invocation). When set, `item/tool/call` is routed to
+   * this handler INSTEAD of the default -32601 rejection; the handler must settle
+   * the id via the provided `respond`/`respondError`. When unset, `item/tool/call`
+   * falls through to the default rejection (keeps the turn unstuck).
+   */
+  onToolCall(fn: CodexAppServerToolCallHandler): void {
+    this.toolCallHandler = fn;
+  }
+
+  close(): void {
+    if (this.closed) return;
+    this.closed = true;
+    try {
+      this.ws?.close(1000, "done");
+    } catch {
+      /* socket close race — ignore */
+    }
+  }
+
+  /** Parse one inbound frame and dispatch by JSON-RPC shape. */
+  private onFrame(raw: unknown): void {
+    let msg: Record<string, unknown>;
+    try {
+      const line = typeof raw === "string" ? raw : Buffer.from(raw as Uint8Array).toString("utf8");
+      msg = JSON.parse(line) as Record<string, unknown>;
+    } catch {
+      // A non-JSON frame is unusable; drop it rather than crash the socket.
+      return;
+    }
+
+    const hasId = msg.id !== undefined && msg.id !== null;
+    const hasMethod = typeof msg.method === "string";
+
+    if (hasId && !hasMethod) {
+      // A RESPONSE to one of our ClientRequests → settle the pending map.
+      const id = msg.id as number;
+      const pending = this.pending.get(id);
+      if (!pending) return;
+      this.pending.delete(id);
+      if (msg.error) {
+        const err = msg.error as { code?: unknown; message?: unknown };
+        pending.reject(new Error(`${String(err.code ?? "error")}: ${String(err.message ?? "unknown")}`));
+      } else {
+        pending.resolve(msg.result);
+      }
+      return;
+    }
+
+    if (hasMethod && hasId) {
+      // A server -> client REQUEST → we MUST reply with the matching id or the turn stalls.
+      const id = msg.id as number;
+      const method = msg.method as string;
+      // A harness function-tool invocation is routed to the passthrough handler
+      // (if registered) so the executor can surface it as a Responses function_call
+      // and complete the turn. The handler owns settling the id.
+      if (method === TOOL_CALL_REQUEST_METHOD && this.toolCallHandler) {
+        this.toolCallHandler(id, msg.params, {
+          respond: (result) => this.respondToRequest(id, result),
+          respondError: (code, message) => this.respondErrorToRequest(id, code, message),
+        });
+        return;
+      }
+      this.answerServerRequest(id, method);
+      return;
+    }
+
+    if (hasMethod) {
+      // A server -> client NOTIFICATION → hand to the stream.
+      this.notificationHandler(msg.method as string, msg.params);
+    }
+  }
+
+  /**
+   * Always answer an inbound ServerRequest so its id is settled. Approval prompts
+   * are auto-APPROVED (OmniRoute is a router; the harness that consumes it owns
+   * execution policy, so codex's own approval must not block the turn). Anything
+   * we cannot service gets a JSON-RPC error so the id is still settled.
+   */
+  private answerServerRequest(id: number, method: string): void {
+    if (!this.ws || this.closed) return;
+    if (APPROVAL_REQUEST_METHODS.has(method)) {
+      // ReviewDecision "approved" — let the model's agentic action proceed. The
+      // harness downstream of OmniRoute is the real gate. Note the note field is
+      // advisory; the decision string is what codex acts on.
+      this.ws.send(
+        JSON.stringify({
+          jsonrpc: "2.0",
+          id,
+          result: { decision: "approved", note: ROUTER_APPROVAL_NOTE },
+        })
+      );
+      return;
+    }
+    // Non-approval server request we do not service here: reject the id so the
+    // app-server does not wait on us (belt-and-suspenders; keeps turns unstuck).
+    this.ws.send(
+      JSON.stringify({
+        jsonrpc: "2.0",
+        id,
+        error: {
+          code: -32601,
+          message: `router: unsupported server request "${method}"`,
+        },
+      })
+    );
+  }
+
+  /** Settle an inbound ServerRequest id with a JSON-RPC result. */
+  private respondToRequest(id: number, result: unknown): void {
+    if (!this.ws || this.closed) return;
+    this.ws.send(JSON.stringify({ jsonrpc: "2.0", id, result }));
+  }
+
+  /** Settle an inbound ServerRequest id with a JSON-RPC error. */
+  private respondErrorToRequest(id: number, code: number, message: string): void {
+    if (!this.ws || this.closed) return;
+    this.ws.send(JSON.stringify({ jsonrpc: "2.0", id, error: { code, message } }));
+  }
+
+  private failAll(reason: string): void {
+    const err = new Error(reason);
+    for (const [id, pending] of this.pending.entries()) {
+      this.pending.delete(id);
+      pending.reject(err);
+    }
+    this.notificationHandler("__transport_closed__", { reason });
+  }
+}

open-sse/executors/codex/appServerConfig.ts

@@ -0,0 +1,94 @@
+import { readFileSync } from "node:fs";
+
+/**
+ * Resolved connection config for the Codex app-server WS transport.
+ *
+ * The app-server is a locally-running `codex app-server` process reachable over a
+ * single WebSocket speaking JSON-RPC 2.0. It self-manages OpenAI auth + model
+ * routing; the ONLY credential OmniRoute presents is the capability token, sent as
+ * `Authorization: Bearer <hex>` on the WS handshake.
+ */
+export interface CodexAppServerConfig {
+  /** ws:// or wss:// URL of the app-server (e.g. "ws://ts-egress:1456"). */
+  url: string;
+  /** Capability token (hex string) sent as `Authorization: Bearer <token>`. */
+  token: string;
+  /** Working directory passed to `thread/start { cwd }` inside the codex container. */
+  cwd: string;
+  /**
+   * Optional codex approval policy override (AskForApproval). Defaults to "never"
+   * in the executor so codex runs non-interactively and never blocks the turn on
+   * its own approval — the harness that consumes OmniRoute owns execution policy.
+   */
+  approvalPolicy?: string;
+  /**
+   * Optional codex sandbox override (SandboxMode). Defaults to "danger-full-access"
+   * in the executor so codex's own sandbox does not gate the model; the harness is
+   * the real gate. Callers may tighten this per request via providerSpecificData.
+   */
+  sandbox?: string;
+}
+
+type ProviderSpecificData = Record<string, unknown> | null | undefined;
+
+function firstString(...values: unknown[]): string | null {
+  for (const value of values) {
+    if (typeof value === "string" && value.trim().length > 0) return value.trim();
+  }
+  return null;
+}
+
+/**
+ * Read the capability token, preferring an inline token, then a token FILE path.
+ * The token file (produced by `codex app-server --ws-token-file <path>`) holds the
+ * same hex string that is presented as the bearer token.
+ */
+function resolveToken(psd: ProviderSpecificData): string | null {
+  const inline = firstString(
+    psd?.codexAppServerToken,
+    process.env.OMNIROUTE_CODEX_APPSERVER_WS_TOKEN
+  );
+  if (inline) return inline;
+
+  const tokenFile = firstString(
+    psd?.codexAppServerTokenFile,
+    process.env.OMNIROUTE_CODEX_APPSERVER_WS_TOKEN_FILE
+  );
+  if (!tokenFile) return null;
+  try {
+    const contents = readFileSync(tokenFile, "utf8").trim();
+    return contents.length > 0 ? contents : null;
+  } catch {
+    return null;
+  }
+}
+
+function isWebSocketUrl(url: string): boolean {
+  return url.startsWith("ws://") || url.startsWith("wss://");
+}
+
+/**
+ * Resolve the app-server connection config from providerSpecificData with env
+ * fallbacks. Returns `null` when not fully configured (URL + token both required)
+ * so the gating predicate `isCodexAppServerRequired` stays false and Codex falls
+ * back to its other transports.
+ */
+export function resolveAppServerConfig(psd: ProviderSpecificData): CodexAppServerConfig | null {
+  const url = firstString(psd?.codexAppServerUrl, process.env.OMNIROUTE_CODEX_APPSERVER_WS);
+  if (!url || !isWebSocketUrl(url)) return null;
+
+  const token = resolveToken(psd);
+  if (!token) return null;
+
+  const cwd =
+    firstString(psd?.codexAppServerCwd, process.env.OMNIROUTE_CODEX_APPSERVER_CWD) ?? "/tmp";
+
+  const approvalPolicy =
+    firstString(psd?.codexAppServerApprovalPolicy, process.env.OMNIROUTE_CODEX_APPSERVER_APPROVAL) ??
+    undefined;
+  const sandbox =
+    firstString(psd?.codexAppServerSandbox, process.env.OMNIROUTE_CODEX_APPSERVER_SANDBOX) ??
+    undefined;
+
+  return { url, token, cwd, ...(approvalPolicy ? { approvalPolicy } : {}), ...(sandbox ? { sandbox } : {}) };
+}

open-sse/executors/codex/appServerEvents.ts

@@ -0,0 +1,208 @@
+import type { AdapterEvent, CodexUsage } from "../../vendor/codex-chatgpt-web/types.ts";
+
+/**
+ * Map Codex app-server JSON-RPC notifications onto the AdapterEvent stream that
+ * `bridgeToResponsesSSE` / `buildResponseJSON` consume.
+ *
+ * Wire method names are the slash-notation ServerNotification variants verified
+ * from the real codex binary (see PROTOCOL-DIGEST.md). Only the handful needed for
+ * a plain text turn are mapped; everything else is ignored.
+ *
+ * The `*Notification` param TYPES referenced below (adapted from the ts-rs bindings):
+ *   AgentMessageDeltaNotification   { threadId, turnId, itemId, delta }
+ *   ReasoningTextDeltaNotification  { threadId, turnId, itemId, delta, contentIndex }
+ *   TurnCompletedNotification       { threadId, turn }  (turn carries usage)
+ *   ErrorNotification               { error, willRetry, threadId, turnId }
+ */
+
+// Wire method names (slash-notation) → intent. Kept as named constants so a typo
+// can't silently break the mapping.
+export const CODEX_APPSERVER_METHODS = {
+  agentMessageDelta: "item/agentMessage/delta",
+  reasoningTextDelta: "item/reasoning/textDelta",
+  reasoningSummaryTextDelta: "item/reasoning/summaryTextDelta",
+  turnCompleted: "turn/completed",
+  error: "error",
+} as const;
+
+/**
+ * The app-server → client REQUEST method by which codex invokes a harness-defined
+ * (dynamic) function tool. It is NOT a notification: it is a server→client
+ * ServerRequest that BLOCKS the codex turn waiting for a `DynamicToolCallResponse`
+ * with the tool's output.
+ *
+ * `params` shape = `DynamicToolCallParams` (ts-rs binding):
+ *   { threadId, turnId, callId, namespace: string | null, tool: string, arguments: JsonValue }
+ *
+ * OmniRoute is a STATELESS ROUTER: it cannot execute the harness's tool (the tool
+ * body lives in the harness downstream, not here). So instead of "executing" the
+ * call, we PASS IT THROUGH: emit tool_call_* AdapterEvents so the bridge renders a
+ * Responses `function_call` output item, then complete the turn. The harness runs
+ * the tool and replays the result in a fresh /v1/responses request (the same
+ * stateless-full-history contract every other OmniRoute provider uses).
+ */
+export const CODEX_APPSERVER_TOOL_CALL_METHOD = "item/tool/call";
+
+/** Minimal shape of the DynamicToolCallParams we consume for the passthrough. */
+export interface DynamicToolCallLike {
+  callId?: unknown;
+  namespace?: unknown;
+  tool?: unknown;
+  arguments?: unknown;
+}
+
+/**
+ * The wire name the bridge's `toolNsMap` is keyed by: namespaced (MCP) tools are
+ * flattened to "<namespace>__<name>". codex sends the namespace + tool separately
+ * on DynamicToolCallParams, so we reconstruct the flat name for the round-trip.
+ */
+export function dynamicToolWireName(namespace: unknown, tool: unknown): string {
+  const name = typeof tool === "string" ? tool : "";
+  return typeof namespace === "string" && namespace.length > 0
+    ? `${namespace}__${name}`
+    : name;
+}
+
+/**
+ * Translate ONE codex `item/tool/call` ServerRequest into the tool_call_* AdapterEvent
+ * triple the bridge already knows how to turn into a Responses function_call /
+ * custom_tool_call / tool_search_call (see bridge.ts:700-784). The `arguments` are
+ * serialized to a JSON string (the bridge accumulates `tool_call_delta.arguments`
+ * as a string and JSON.parses it at close).
+ *
+ * This emits the COMPLETE call in one shot (start → delta → end) because the
+ * server-request carries the fully-formed arguments (codex does not stream dynamic
+ * tool-call arguments to the client the way the chatgpt-web adapter streams native
+ * ones). The caller is responsible for then completing the turn.
+ */
+export function translateToolCall(
+  params: DynamicToolCallLike,
+  push: (event: AdapterEvent) => void
+): void {
+  const callId =
+    typeof params.callId === "string" && params.callId.length > 0
+      ? params.callId
+      : `call_${Math.random().toString(36).slice(2)}`;
+  const name = dynamicToolWireName(params.namespace, params.tool);
+  let argsStr = "{}";
+  const rawArgs = params.arguments;
+  if (typeof rawArgs === "string") {
+    argsStr = rawArgs.length > 0 ? rawArgs : "{}";
+  } else if (rawArgs !== undefined && rawArgs !== null) {
+    try {
+      argsStr = JSON.stringify(rawArgs);
+    } catch {
+      argsStr = "{}";
+    }
+  }
+  push({ type: "tool_call_start", id: callId, name });
+  if (argsStr.length > 0) push({ type: "tool_call_delta", arguments: argsStr });
+  push({ type: "tool_call_end" });
+}
+
+interface RawUsage {
+  input_tokens?: number;
+  cached_input_tokens?: number;
+  output_tokens?: number;
+  reasoning_output_tokens?: number;
+  total_tokens?: number;
+}
+
+/** Extract a numeric field defensively (the wire may omit or null it). */
+function num(value: unknown): number | undefined {
+  return typeof value === "number" && Number.isFinite(value) ? value : undefined;
+}
+
+/**
+ * Convert the app-server usage shape (snake_case token counts) into the canonical
+ * CodexUsage the bridge expects. Returns undefined when nothing usable is present.
+ */
+export function mapUsage(raw: unknown): CodexUsage | undefined {
+  if (!raw || typeof raw !== "object") return undefined;
+  const u = raw as RawUsage;
+  const inputTokens = num(u.input_tokens) ?? 0;
+  const outputTokens = num(u.output_tokens) ?? 0;
+  const usage: CodexUsage = { inputTokens, outputTokens };
+  const cached = num(u.cached_input_tokens);
+  if (cached !== undefined) {
+    usage.cachedInputTokens = cached;
+    usage.cacheReadInputTokens = cached;
+  }
+  const reasoning = num(u.reasoning_output_tokens);
+  if (reasoning !== undefined) usage.reasoningOutputTokens = reasoning;
+  const total = num(u.total_tokens);
+  if (total !== undefined) usage.totalTokens = total;
+  return usage;
+}
+
+/**
+ * Pull a usage object out of a `turn/completed` param. The Turn payload carries
+ * token counts; different app-server builds nest it under `usage` or `tokenUsage`,
+ * so probe both before giving up.
+ */
+function extractTurnUsage(params: Record<string, unknown>): CodexUsage | undefined {
+  const turn = params.turn;
+  if (turn && typeof turn === "object") {
+    const t = turn as Record<string, unknown>;
+    return mapUsage(t.usage) ?? mapUsage(t.tokenUsage) ?? mapUsage(t.token_usage);
+  }
+  return mapUsage(params.usage);
+}
+
+function errorMessage(params: Record<string, unknown>): string {
+  const err = params.error;
+  if (err && typeof err === "object") {
+    const m = (err as Record<string, unknown>).message;
+    if (typeof m === "string" && m.length > 0) return m;
+  }
+  if (typeof params.message === "string" && params.message.length > 0) return params.message;
+  return "Codex app-server reported an error";
+}
+
+/**
+ * Translate one notification into AdapterEvent(s) and push them into the queue.
+ *
+ * Returns `true` when the notification is terminal (turn/completed or error), so
+ * the caller can close the event queue after draining.
+ */
+export function translateNotification(
+  method: string,
+  params: unknown,
+  push: (event: AdapterEvent) => void
+): boolean {
+  const p = (params && typeof params === "object" ? params : {}) as Record<string, unknown>;
+
+  switch (method) {
+    case CODEX_APPSERVER_METHODS.agentMessageDelta: {
+      const delta = p.delta;
+      if (typeof delta === "string" && delta.length > 0) {
+        push({ type: "text_delta", text: delta });
+      }
+      return false;
+    }
+    case CODEX_APPSERVER_METHODS.reasoningTextDelta:
+    case CODEX_APPSERVER_METHODS.reasoningSummaryTextDelta: {
+      const delta = p.delta;
+      if (typeof delta === "string" && delta.length > 0) {
+        push({ type: "thinking_delta", thinking: delta });
+      }
+      return false;
+    }
+    case CODEX_APPSERVER_METHODS.turnCompleted: {
+      push({ type: "done", usage: extractTurnUsage(p), endTurn: true });
+      return true;
+    }
+    case CODEX_APPSERVER_METHODS.error: {
+      push({
+        type: "error",
+        message: errorMessage(p),
+        status: 502,
+        errorType: "provider_error",
+        code: "codex_app_server_turn_failed",
+      });
+      return true;
+    }
+    default:
+      return false;
+  }
+}

open-sse/executors/index.ts

@@ -7,6 +7,7 @@ import { GheCopilotExecutor } from "./ghe-copilot.ts";
 import { QoderExecutor } from "./qoder.ts";
 import { KiroExecutor } from "./kiro.ts";
 import { CodexExecutor } from "./codex.ts";
+import { CodexAppServerExecutor } from "./codex-app-server.ts";
 import { CursorExecutor } from "./cursor.ts";
 import { TraeExecutor } from "./trae.ts";
 import { DefaultExecutor } from "./default.ts";
@@ -97,6 +98,7 @@ const executors = {
   "amazon-q": new KiroExecutor("amazon-q"),
   bedrock: new BedrockExecutor(),
   codex: new CodexExecutor(),
+  "codex-app-server": new CodexAppServerExecutor({}, "codex-app-server"),
   "chatgpt-web-codex": new ChatGptWebCodexExecutor(),
   "cgpt-codex": new ChatGptWebCodexExecutor(),
   cursor: new CursorExecutor(),

open-sse/services/compression/harness/benchmark.ts

@@ -187,6 +187,12 @@ export function engineToCompressFn(engineId: string): CompressFn {
   return async (text: string): Promise<string> => {
     const body: Record<string, unknown> = {
       messages: [{ role: "user", content: text }],
+      // #7746 follow-up: CCR only compresses for callers that advertise the
+      // omniroute_ccr_retrieve tool (otherwise its content-addressed marker is
+      // unresolvable). Real CCR traffic always carries this tool, so the
+      // benchmark must too, or CCR measures as a no-op. Other engines ignore
+      // the `tools` field, so this is inert for them.
+      tools: [{ type: "function", function: { name: "omniroute_ccr_retrieve" } }],
     };
 
     try {
@@ -199,6 +205,16 @@ export function engineToCompressFn(engineId: string): CompressFn {
 
       const messages = result.body["messages"];
       if (Array.isArray(messages) && messages.length > 0) {
+        // CCR may inject a leading [CCR protocol] system instruction, so the
+        // compressed user text is not necessarily messages[0]. Prefer the LAST
+        // message with string content (the user turn we fed in); fall back to
+        // the first string content otherwise.
+        for (let i = messages.length - 1; i >= 0; i--) {
+          const c = (messages[i] as Record<string, unknown>)["content"];
+          if (typeof c === "string" && (messages[i] as Record<string, unknown>)["role"] !== "system") {
+            return c;
+          }
+        }
         const content = (messages[0] as Record<string, unknown>)["content"];
         if (typeof content === "string") return content;
       }

package.json

@@ -1,7 +1,7 @@
 {
   "name": "omniroute",
   "version": "3.8.50",
-  "description": "Unified AI router with 350 providers, RTK+Caveman compression, auto fallback, MCP/A2A, desktop, PWA, and OpenAI-compatible APIs.",
+  "description": "Unified AI router with 351 providers, RTK+Caveman compression, auto fallback, MCP/A2A, desktop, PWA, and OpenAI-compatible APIs.",
   "type": "module",
   "bin": {
     "omniroute": "bin/omniroute.mjs",

src/app/(dashboard)/dashboard/providers/[id]/components/ConnectionsListPanel.tsx

@@ -409,7 +409,7 @@ export default function ConnectionsListPanel({
                     ? (enabled) => handleToggleConnectionAutoSync(conn.id, enabled)
                     : undefined
                 }
-                isCodex={providerId === "codex"}
+                isCodex={providerId === "codex" || providerId === "codex-app-server"}
                 isCcCompatible={isCcCompatible}
                 cliproxyapiEnabled={cpaProviderEnabled}
                 onToggleCliproxyapiMode={(enabled) => handleToggleCliproxyapiMode(conn.id, enabled)}
@@ -610,7 +610,7 @@ export default function ConnectionsListPanel({
                         ? (enabled) => handleToggleConnectionAutoSync(conn.id, enabled)
                         : undefined
                     }
-                    isCodex={providerId === "codex"}
+                    isCodex={providerId === "codex" || providerId === "codex-app-server"}
                     isCcCompatible={isCcCompatible}
                     cliproxyapiEnabled={cpaProviderEnabled}
                     onToggleCliproxyapiMode={(enabled) =>

src/app/api/providers/[id]/codex-auth/apply-local/route.ts

@@ -1,10 +1,22 @@
 import { NextResponse } from "next/server";
+import { z } from "zod";
 import { requireManagementAuth } from "@/lib/api/requireManagementAuth";
 import { ensureCliConfigWriteAllowed } from "@/shared/services/cliRuntime";
-import { CodexAuthFileError, writeCodexAuthFileToLocalCli } from "@/lib/oauth/utils/codexAuthFile";
+import {
+  CodexAuthFileError,
+  writeCodexAuthFileToLocalCliIfNeeded,
+} from "@/lib/oauth/utils/codexAuthFile";
 import { getAuditRequestContext, logAuditEvent } from "@/lib/compliance/index";
 import { sanitizeErrorMessage } from "@omniroute/open-sse/utils/error";
 
+// Optional body { force?: boolean }. Unknown keys are stripped rather than
+// rejected so the endpoint stays tolerant of the empty/no-body calls it
+// historically accepted. Non-boolean `force` is coerced away to the default.
+const ApplyLocalBodySchema = z
+  .object({ force: z.boolean().optional() })
+  .partial()
+  .passthrough();
+
 function toErrorResponse(error: unknown) {
   if (error instanceof CodexAuthFileError) {
     return NextResponse.json(
@@ -33,7 +45,21 @@ export async function POST(request: Request, { params }: { params: Promise<{ id:
     }
 
     const { id } = await params;
-    const result = await writeCodexAuthFileToLocalCli(id);
+
+    // Optional { force?: boolean } body. By default we DON'T clobber an existing,
+    // fresh ~/.codex/auth.json (a session the user may be managing themselves);
+    // force overwrites it (a backup is always taken regardless). Malformed/empty
+    // bodies are tolerated — this endpoint historically took no body.
+    let force = false;
+    try {
+      const parsed = ApplyLocalBodySchema.safeParse(await request.json());
+      force = parsed.success ? parsed.data.force === true : false;
+    } catch {
+      /* no body — default force=false */
+    }
+
+    const applied = await writeCodexAuthFileToLocalCliIfNeeded(id, { force });
+    const result = applied.result;
 
     logAuditEvent({
       action: "provider.credentials.applied",
@@ -45,18 +71,21 @@ export async function POST(request: Request, { params }: { params: Promise<{ id:
       requestId: auditContext.requestId,
       metadata: {
         provider: "codex",
-        authPath: result.authPath,
-        savedBakPath: result.savedBakPath,
+        decision: applied.decision,
+        authPath: applied.authPath,
+        savedBakPath: result?.savedBakPath,
       },
     });
 
     return NextResponse.json({
       success: true,
       connectionId: id,
-      connectionLabel: result.connectionLabel,
-      authPath: result.authPath,
-      savedBakPath: result.savedBakPath,
-      centralizedBackupPath: result.centralizedBackupPath,
+      // "skipped_present_fresh" means an existing healthy auth.json was kept.
+      decision: applied.decision,
+      connectionLabel: result?.connectionLabel,
+      authPath: applied.authPath,
+      savedBakPath: result?.savedBakPath,
+      centralizedBackupPath: result?.centralizedBackupPath,
       writtenAt: new Date().toISOString(),
     });
   } catch (error) {

src/app/api/providers/[id]/test/codexAppServerHealth.ts

@@ -0,0 +1,136 @@
+/**
+ * Build the structured diagnosis object the connection-test route returns.
+ * Lives here (rather than inline in test/route.ts) so both the route and the
+ * codex-app-server health probe share one definition. Pure.
+ */
+export function makeDiagnosis(
+  type: string,
+  source: string,
+  message: string | null,
+  code: string | null = null
+) {
+  return {
+    type,
+    source,
+    message: message || null,
+    code: code ?? null,
+  };
+}
+
+export type CodexAppServerHealth = {
+  valid: boolean;
+  error?: string;
+  diagnosis: unknown;
+  refreshed: boolean;
+};
+
+/**
+ * A codex "app-server" connection (providerSpecificData.codexTransport ===
+ * "app-server") does NOT carry a validatable OpenAI token: it drives the codex
+ * CLI's own `codex app-server` process over JSON-RPC/WebSocket, and THAT process
+ * self-manages its OpenAI OAuth (its own ~/.codex/auth.json), exactly like an
+ * interactive codex session. So the ordinary OAuth token probe is meaningless for
+ * these connections — it validates a placeholder and reports a false "Token
+ * invalid or revoked" 401 (which then trips the rate-limit cooldown on retest).
+ *
+ * The correct health signal for this transport is whether the app-server itself
+ * is reachable and ready. The app-server exposes an unauthenticated liveness
+ * endpoint at <httpBase>/readyz (200 = ready) alongside its ws:// listener, so we
+ * derive the http(s) origin from the configured ws(s):// URL and probe /readyz.
+ * Returns null when this connection is NOT an app-server connection (so the caller
+ * falls through to the normal token validation).
+ */
+export async function testCodexAppServerConnection(
+  connection: any
+): Promise<CodexAppServerHealth | null> {
+  const psd = (connection?.providerSpecificData as Record<string, unknown> | undefined) || undefined;
+  // Fire the /readyz probe when EITHER (a) the connection opted into the
+  // app-server transport via the per-connection flag (a `codex` provider
+  // connection with codexTransport==="app-server"), OR (b) this is the
+  // first-class `codex-app-server` provider, which is app-server by definition
+  // and needs no flag. Otherwise return null so the caller falls through to the
+  // normal OAuth/apikey token validation.
+  const isAppServerProvider = connection?.provider === "codex-app-server";
+  const isAppServerFlag = psd?.codexTransport === "app-server";
+  if (!isAppServerProvider && !isAppServerFlag) return null;
+
+  // Dynamic import (not a static top-level import) so this executor-config module
+  // stays behind the open-sse boundary the no-restricted-imports lint rule enforces.
+  const { resolveAppServerConfig } = await import(
+    "@omniroute/open-sse/executors/codex/appServerConfig.ts"
+  );
+  const config = resolveAppServerConfig(psd);
+  if (!config) {
+    const error = "Codex app-server transport is not configured (missing url or token)";
+    return {
+      valid: false,
+      error,
+      refreshed: false,
+      diagnosis: makeDiagnosis("validation_error", "local", error, "app_server_unconfigured"),
+    };
+  }
+
+  // ws://host:port → http://host:port/readyz ; wss:// → https://.
+  const httpBase = config.url.replace(/^ws(s?):\/\//i, (_m, s) => `http${s}://`).replace(/\/+$/, "");
+  const readyzUrl = `${httpBase}/readyz`;
+
+  const controller = new AbortController();
+  const timer = setTimeout(() => controller.abort(), 8000);
+  try {
+    const res = await fetch(readyzUrl, {
+      method: "GET",
+      headers: { Authorization: `Bearer ${config.token}` },
+      signal: controller.signal,
+    });
+    if (res.status !== 200) {
+      const error = `Codex app-server not ready (${readyzUrl} → HTTP ${res.status})`;
+      return {
+        valid: false,
+        error,
+        refreshed: false,
+        diagnosis: makeDiagnosis("provider_error", "app_server", error, "app_server_not_ready"),
+      };
+    }
+    // The server PROCESS is up. Now confirm its Codex CLI is actually SIGNED IN —
+    // /readyz alone would show green for a logged-out CLI, which then fails on the
+    // first real turn. Probe account/read over the JSON-RPC WebSocket.
+    let authStatus;
+    try {
+      const [{ probeCodexAppServerAuth }, { getCodexAppServerWebsocketTransport }] =
+        await Promise.all([
+          import("@omniroute/open-sse/executors/codex/appServerAuthProbe.ts"),
+          import("@omniroute/open-sse/executors/codex.ts"),
+        ]);
+      authStatus = await probeCodexAppServerAuth(config, getCodexAppServerWebsocketTransport(), 8000);
+    } catch (probeErr: any) {
+      // If the auth probe itself fails to load/run, don't fail the whole health
+      // check — the server IS reachable. Treat as unknown-but-reachable (valid).
+      authStatus = { state: "unknown", reason: probeErr?.message ?? "auth probe failed" } as const;
+    }
+
+    if (authStatus.state === "logged_out") {
+      const error =
+        "Codex app-server is running but its Codex CLI is not signed in. Use \u201cSign in with ChatGPT\u201d to authenticate.";
+      return {
+        valid: false,
+        error,
+        refreshed: false,
+        diagnosis: makeDiagnosis("auth_required", "app_server", error, "app_server_login_required"),
+      };
+    }
+    // "authenticated" → healthy; "unknown" (probe unavailable/timed out) → treat
+    // the reachable server as healthy rather than blocking on an inconclusive probe.
+    return { valid: true, refreshed: false, diagnosis: null };
+  } catch (err: any) {
+    const reason = err?.name === "AbortError" ? "timed out" : (err?.message ?? "unreachable");
+    const error = `Codex app-server unreachable (${readyzUrl}: ${reason})`;
+    return {
+      valid: false,
+      error,
+      refreshed: false,
+      diagnosis: makeDiagnosis("provider_error", "app_server", error, "app_server_unreachable"),
+    };
+  } finally {
+    clearTimeout(timer);
+  }
+}

src/app/api/providers/[id]/test/route.ts

@@ -28,6 +28,7 @@ import {
 } from "@/lib/oauth/gitlab";
 import { providerAllowsOptionalApiKey } from "@/shared/constants/providers";
 import { shouldUseApiKeyConnectionTest } from "./webSessionTestDispatch";
+import { testCodexAppServerConnection, makeDiagnosis } from "./codexAppServerHealth";
 import { removeConnectionHealth } from "@omniroute/open-sse/services/apiKeyRotator.ts";
 import { isConnectionUnavailableToAuxiliaryActivity } from "@/lib/exclusiveLeaseIsolation";
 import { classifyAmbiguousOrAuthError, type ClassifyFailureArgs } from "./mistralAmbiguousAuth";
@@ -52,20 +53,6 @@ function toSafeMessage(value: any, fallback = "Unknown error"): string {
   return trimmed || fallback;
 }
 
-function makeDiagnosis(
-  type: string,
-  source: string,
-  message: string | null,
-  code: string | null = null
-) {
-  return {
-    type,
-    source,
-    message: message || null,
-    code: code ?? null,
-  };
-}
-
 /**
  * A provider/account that the upstream has deactivated (vs. a revoked/expired token).
  * #1444: a Codex account can have a perfectly healthy OAuth refresh while its ChatGPT
@@ -1024,6 +1011,13 @@ export async function testSingleConnection(connectionId: string, validationModel
   const startTime = Date.now();
   const runtime = await getProviderRuntimeStatus(connection);
 
+  // Codex app-server connections carry no validatable OpenAI token (the codex
+  // app-server process self-manages its own OAuth). Probe the app-server's
+  // /readyz liveness endpoint instead of the meaningless token check — otherwise
+  // every sweep reports a false "Token invalid or revoked" 401 and cools the
+  // connection down. Returns null for non-app-server connections (fall through).
+  const appServerResult = await testCodexAppServerConnection(connection);
+
   if ((runtime as any)?.diagnosis) {
     result = {
       valid: false,
@@ -1031,6 +1025,10 @@ export async function testSingleConnection(connectionId: string, validationModel
       refreshed: false,
       diagnosis: (runtime as any).diagnosis,
     };
+  } else if (appServerResult) {
+    result = await runWithProxyContext(proxyInfo?.proxy || null, () =>
+      Promise.resolve(appServerResult)
+    );
   } else if (shouldUseApiKeyConnectionTest(connection.authType, provider)) {
     const enrichedConnection = validationModelId
       ? {

src/i18n/messages/pt-BR.json

@@ -13873,5 +13873,12 @@
     "toolsMismatch": "O provedor nao suporta chamada de ferramentas",
     "structuredOutputMismatch": "O provedor nao suporta saida estruturada",
     "contextWindowMismatch": "A requisicao excede a janela de contexto do provedor"
+  },
+  "cheaperInferenceSponsorBanner": {
+    "title": "Cheaper Inference é um Amigo Open Source do OmniRoute",
+    "description": "Um gateway ordenado por custo que revende dezenas de modelos de fronteira atrás de um único endpoint compatível com OpenAI, roteando cada requisição para o provedor elegível mais barato, nunca acima do preço de tabela.",
+    "cta": "Obter uma chave de API",
+    "partnerLinkNote": "Link de parceiro",
+    "dismissAriaLabel": "Dispensar"
   }
 }

src/lib/db/adapters/nodeSqliteShared.ts

@@ -6,6 +6,10 @@ export interface NodeSqliteDatabaseLike {
     run(...p: unknown[]): { changes: number | bigint; lastInsertRowid: number | bigint };
     get(...p: unknown[]): unknown;
     all(...p: unknown[]): unknown[];
+    // node:sqlite (DatabaseSync) statements expose these tuning setters. They
+    // are optional here so the shared adapter also accepts lighter test doubles.
+    setAllowUnknownNamedParameters?(enabled: boolean): void;
+    setAllowBareNamedParameters?(enabled: boolean): void;
   };
   exec(sql: string): void;
   close(): void;
@@ -13,6 +17,46 @@ export interface NodeSqliteDatabaseLike {
 
 const MAX_STMT_CACHE_SIZE = 200;
 
+// node:sqlite hands back rows whose prototype is `null` (Object.create(null)),
+// whereas better-sqlite3 (the driver we ship and run in production/CI) returns
+// ordinary Object.prototype rows. The difference is invisible for normal
+// property access but breaks callers that compare rows with structural
+// equality that also checks the prototype (e.g. Node's assert.deepStrictEqual,
+// used by unit tests written against the better-sqlite3 row shape). Normalize
+// every row to a plain object so the node:sqlite fallback is behaviourally
+// identical to the native better-sqlite3 path.
+function toPlainRow<T>(row: T): T {
+  if (row === null || typeof row !== "object") return row;
+  return { ...(row as Record<string, unknown>) } as T;
+}
+
+// better-sqlite3 (the production/CI driver) and sql.js both accept `undefined`
+// as a bound value and treat it as SQL NULL. node:sqlite is stricter and throws
+// "Provided value cannot be bound to SQLite parameter N" for undefined. Several
+// call sites pass undefined for absent optional columns (e.g. a capability sync
+// that omits modalities_input), so coerce undefined -> null here to keep the
+// node:sqlite fallback behaviourally compatible with the native driver. This
+// handles both positional params and a single named-params object.
+function normalizeBindParams(params: unknown[]): unknown[] {
+  const [first] = params;
+  const isLoneNamedParamsObject =
+    params.length === 1 &&
+    first !== null &&
+    typeof first === "object" &&
+    !Array.isArray(first) &&
+    !Buffer.isBuffer(first) &&
+    !(first instanceof Uint8Array);
+  if (isLoneNamedParamsObject) {
+    const source = first as Record<string, unknown>;
+    const normalized: Record<string, unknown> = {};
+    for (const key of Object.keys(source)) {
+      normalized[key] = source[key] === undefined ? null : source[key];
+    }
+    return [normalized];
+  }
+  return params.map((value) => (value === undefined ? null : value));
+}
+
 export function createNodeSqliteAdapterFromDatabase(
   db: NodeSqliteDatabaseLike,
   filePath: string,
@@ -41,6 +85,14 @@ export function createNodeSqliteAdapterFromDatabase(
       stmtCache.set(sql, entry);
     } else {
       const stmt = db.prepare(sql);
+      // better-sqlite3 (the production/CI driver) silently ignores named
+      // parameters supplied in the bind object that the SQL text does not
+      // reference. node:sqlite instead throws "Unknown named parameter '<x>'".
+      // Several call sites deliberately pass a superset params object (e.g. an
+      // UPDATE that omits @createdAt while the shared params builder still
+      // includes it), so relax node:sqlite to match better-sqlite3 and keep the
+      // fallback driver behaviourally compatible.
+      stmt.setAllowUnknownNamedParameters?.(true);
       if (stmtCache.size >= MAX_STMT_CACHE_SIZE) {
         const oldestKey = stmtCache.keys().next().value;
         if (oldestKey !== undefined) {
@@ -119,17 +171,19 @@ export function createNodeSqliteAdapterFromDatabase(
       const stmt = getCached(sql);
       return {
         run(...params: unknown[]): RunResult {
-          const r = stmt.run(...params);
+          const r = stmt.run(...normalizeBindParams(params));
           return {
             changes: Number(r.changes ?? 0),
             lastInsertRowid: Number(r.lastInsertRowid ?? 0),
           };
         },
         get(...params: unknown[]): unknown {
-          return stmt.get(...params);
+          return toPlainRow(stmt.get(...normalizeBindParams(params)));
         },
         all(...params: unknown[]): unknown[] {
-          return stmt.all(...params);
+          return (stmt.all(...normalizeBindParams(params)) as unknown[]).map((row) =>
+            toPlainRow(row)
+          );
         },
       };
     },

src/lib/db/adapters/sqljsAdapter.ts

@@ -6,6 +6,19 @@ import type { SqliteAdapter, PreparedStatement, RunResult } from "./types";
 const SAVE_DEBOUNCE_MS = 100;
 const CHECKPOINT_INTERVAL_MS = 60_000;
 
+// sql.js's stmt.getAsObject() returns rows whose prototype is `null`
+// (Object.create(null)), whereas better-sqlite3 (the driver we ship and run in
+// production/CI) hands back ordinary Object.prototype rows. That difference is
+// invisible for normal property access but breaks callers that compare rows
+// with structural equality that also checks the prototype (e.g. Node's
+// assert.deepStrictEqual, used by several unit tests written against the
+// better-sqlite3 row shape). Normalize every row to a plain object so the
+// sql.js fallback is behaviourally identical to the native better-sqlite3 path.
+function toPlainRow<T>(row: T): T {
+  if (row === null || typeof row !== "object") return row;
+  return { ...(row as Record<string, unknown>) } as T;
+}
+
 let _sqlJsLib: Awaited<ReturnType<(typeof import("sql.js"))["default"]>> | null = null;
 
 function resolveSqlJsWasmPath(): string {
@@ -240,7 +253,7 @@ export async function createSqlJsAdapter(filePath: string): Promise<SqliteAdapte
         try {
           const bindValue = toBindValue(params);
           if (bindValue !== undefined) stmt.bind(bindValue);
-          if (stmt.step()) return stmt.getAsObject();
+          if (stmt.step()) return toPlainRow(stmt.getAsObject());
           return undefined;
         } finally {
           stmt.free();
@@ -252,7 +265,7 @@ export async function createSqlJsAdapter(filePath: string): Promise<SqliteAdapte
           const bindValue = toBindValue(params);
           if (bindValue !== undefined) stmt.bind(bindValue);
           const rows: unknown[] = [];
-          while (stmt.step()) rows.push(stmt.getAsObject());
+          while (stmt.step()) rows.push(toPlainRow(stmt.getAsObject()));
           return rows;
         } finally {
           stmt.free();

src/lib/oauth/utils/codexAuthFile.ts

@@ -350,3 +350,89 @@ export async function writeCodexAuthFileToLocalCli(connectionId: string) {
     centralizedBackupPath,
   };
 }
+
+/**
+ * Decision for the guarded write (see writeCodexAuthFileToLocalCliIfNeeded).
+ */
+export type CodexAuthWriteDecision =
+  | "written" // wrote a fresh auth.json (was absent, stale, or force)
+  | "skipped_present_fresh"; // an existing, non-stale auth.json was left untouched
+
+/**
+ * Guarded variant of writeCodexAuthFileToLocalCli for the codex-app-server
+ * "Sign in with ChatGPT" flow. Per the design decision (William, Q2):
+ *
+ *   - Write ONLY when ~/.codex/auth.json is ABSENT, or STALE (its token is at/
+ *     past the refresh buffer), or when `force` is set.
+ *   - NEVER clobber an existing, healthy (non-stale) auth.json — a user may be
+ *     managing the CLI session themselves. (The underlying writer always makes a
+ *     backup regardless, so even a forced overwrite is recoverable.)
+ *
+ * Staleness is read from the existing file's `last_refresh` + the token's own
+ * expiry claim (JWT `exp` on the access_token) when present; if neither is
+ * readable we treat the file as fresh (do not clobber).
+ *
+ * Returns the write decision plus (when written) the underlying write result.
+ */
+export async function writeCodexAuthFileToLocalCliIfNeeded(
+  connectionId: string,
+  options: { force?: boolean } = {}
+): Promise<{ decision: CodexAuthWriteDecision; authPath: string | null; result?: Awaited<ReturnType<typeof writeCodexAuthFileToLocalCli>> }> {
+  const paths = getCliConfigPaths("codex");
+  const authPath = paths?.auth ?? null;
+
+  if (!options.force && authPath) {
+    const existing = await readExistingCodexAuth(authPath);
+    if (existing && !isCodexAuthStale(existing)) {
+      // Present and healthy — do not clobber a session we didn't (or don't need
+      // to) manage. The connection can still authenticate turns via this file.
+      return { decision: "skipped_present_fresh", authPath };
+    }
+  }
+
+  const result = await writeCodexAuthFileToLocalCli(connectionId);
+  return { decision: "written", authPath: result.authPath, result };
+}
+
+/** Read + parse an existing ~/.codex/auth.json; null when absent/unreadable. */
+async function readExistingCodexAuth(authPath: string): Promise<CodexAuthFilePayload | null> {
+  try {
+    const raw = await fs.readFile(authPath, "utf8");
+    const parsed = JSON.parse(raw) as unknown;
+    const rec = toRecord(parsed);
+    const tokens = toRecord(rec.tokens);
+    if (!toNonEmptyString(tokens.access_token)) return null;
+    return parsed as CodexAuthFilePayload;
+  } catch {
+    return null;
+  }
+}
+
+/**
+ * A stored auth.json is "stale" when its access_token is at/past the refresh
+ * buffer. Prefer the JWT `exp` claim on the access_token; fall back to
+ * `last_refresh` + a conservative validity window; if neither is parseable,
+ * treat as NOT stale (never clobber on ambiguity).
+ */
+function isCodexAuthStale(payload: CodexAuthFilePayload): boolean {
+  const accessToken = toNonEmptyString(payload?.tokens?.access_token);
+  if (accessToken) {
+    const claims = decodeJwtPayload(accessToken);
+    const exp = claims && typeof claims.exp === "number" ? claims.exp : null;
+    if (exp) {
+      const expiresAtMs = exp * 1000;
+      return expiresAtMs - Date.now() <= CODEX_REFRESH_BUFFER_MS;
+    }
+  }
+  // No usable exp claim — fall back to last_refresh age. Codex access tokens are
+  // short-lived (~hours); if the file hasn't refreshed in > 6h, consider it stale.
+  const lastRefresh = toNonEmptyString(payload?.last_refresh);
+  if (lastRefresh) {
+    const refreshedMs = new Date(lastRefresh).getTime();
+    if (!Number.isNaN(refreshedMs)) {
+      const SIX_HOURS_MS = 6 * 60 * 60 * 1000;
+      return Date.now() - refreshedMs >= SIX_HOURS_MS;
+    }
+  }
+  return false;
+}

src/lib/providers/imageValidation.ts

@@ -1,6 +1,7 @@
 import { getImageProvider } from "@omniroute/open-sse/config/imageRegistry";
 
 import { getProviderOutboundGuard } from "@/shared/network/outboundUrlGuardPolicy";
+import { isSecurityBlockError } from "@/lib/providers/validation/transport";
 import {
   SAFE_OUTBOUND_FETCH_PRESETS,
   SafeOutboundFetchError,
@@ -62,7 +63,7 @@ function toValidationErrorResult(error: unknown) {
     ...(error instanceof SafeOutboundFetchError && error.code === "TIMEOUT"
       ? { timeout: true }
       : {}),
-    ...(statusCode === 400 ? { securityBlocked: true } : {}),
+    ...(isSecurityBlockError(error) ? { securityBlocked: true } : {}),
   };
 }
 

src/shared/constants/codexClient.ts

@@ -1,4 +1,9 @@
-export const DEFAULT_CODEX_CLIENT_VERSION = "0.146.0";
+// Kept in lockstep with the codex CLI actually installed in the OmniRoute image
+// (bin/omniroute-fix.Containerfile installs `codex` latest; app-server runtime is
+// 0.149.0 as of 2026-08-22). When the image's codex is bumped, refresh this so the
+// fingerprint OpenAI sees from the OAuth/Responses face matches the real client
+// version. Overridable per-deployment via the CODEX_CLIENT_VERSION env.
+export const DEFAULT_CODEX_CLIENT_VERSION = "0.149.0";
 export const CODEX_CLI_RS_ORIGINATOR = "codex_cli_rs";
 
 export function getCodexCliRsHeaders(

src/shared/constants/featureFlagDefinitions.ts

@@ -401,6 +401,18 @@ export const FEATURE_FLAG_DEFINITIONS: FeatureFlagDefinition[] = [
     requiresRestart: false,
     warningLevel: "info",
   },
+  {
+    key: "OMNIROUTE_CODEX_APP_SERVER_ENABLED",
+    label: "Codex App-Server Transport",
+    description:
+      "Allow Codex to use the local app-server WebSocket JSON-RPC transport (codexTransport=app-server). When off, connections opted into app-server fall back to Codex's other transports.",
+    descriptionI18nKey: "featureFlagOmnirouteCodexAppServerEnabledDescription",
+    category: "runtime",
+    defaultValue: "true",
+    type: "boolean",
+    requiresRestart: false,
+    warningLevel: "info",
+  },
   {
     key: "OMNIROUTE_EMERGENCY_FALLBACK",
     label: "Emergency Fallback",

src/shared/constants/providers/noauth.ts

@@ -175,6 +175,30 @@ export const NOAUTH_PROVIDERS = {
       text: "ZCode runs locally through its native app-server. OmniRoute never receives or stores the Z.ai credential.",
     },
   },
+  "codex-app-server": {
+    id: "codex-app-server",
+    alias: "cxa",
+    name: "OpenAI Codex (App-Server)",
+    icon: "code",
+    color: "#10A37F",
+    textIcon: "CA",
+    website: "https://developers.openai.com/codex/cli",
+    noAuth: true,
+    hasFree: false,
+    serviceKinds: ["llm"],
+    isLocalCli: true,
+    // No subscriptionRisk / riskNoticeVariant: unlike the `codex` provider (which
+    // replays your ChatGPT/OpenAI session token to the API), this transport drives
+    // the Codex CLI's own `codex app-server` over JSON-RPC/WebSocket. The CLI owns
+    // and self-refreshes its OAuth (~/.codex/auth.json) exactly like an interactive
+    // `codex` session — OmniRoute never replays a token to the API — so the
+    // "official session not authorized for proxy use" caveat does not apply.
+    authHint:
+      "No token stored by OmniRoute. The Codex CLI app-server manages its own ChatGPT sign-in (~/.codex/auth.json, auto-refreshed). Use \u201cSign in with ChatGPT\u201d if the CLI is not yet authenticated.",
+    notice: {
+      text: "OpenAI Codex (App-Server) drives the Codex CLI's local app-server (JSON-RPC over WebSocket). The CLI self-manages its OpenAI OAuth, so OmniRoute never sees or replays your token. Requires the codex CLI reachable at the configured app-server URL; sign in via the CLI or the dashboard \u201cSign in with ChatGPT\u201d action.",
+    },
+  },
   uncloseai: {
     id: "uncloseai",
     alias: "unc",

src/shared/services/cliRuntime.ts

@@ -338,6 +338,7 @@ export const CLI_TOOL_ALIASES: Readonly<Record<string, string>> = {
   "claude-code": "claude",
   "openai-codex": "codex",
   openai: "codex",
+  "codex-app-server": "codex",
   cn: "continue",
   qodercli: "qoder",
 };

stryker.conf.json

@@ -72,6 +72,8 @@
       "tests/unit/antigravity-prefer-stored-project.test.ts",
       "tests/unit/api-key-policy-noauth-allowed-connections.test.ts",
       "tests/unit/api-key-rotator-health.test.ts",
+      "tests/unit/chat-routing-synced-inventory-11089.test.ts",
+      "tests/unit/security-route-guard-tiers.test.ts",
       "tests/unit/egress-ip-lock-10880.test.ts",
       "tests/unit/egress-lock-allowlist-10880.test.ts",
       "tests/unit/api/jobs.test.ts",

tests/integration/search-providers-catalog.test.ts

@@ -48,10 +48,10 @@ const route = await import("../../src/app/api/search/providers/route.ts");
 // Constants
 // ---------------------------------------------------------------------------
 
-// 16 search-kind providers: serper, brave, perplexity, exa, tavily, firecrawl,
+// 17 search-kind providers: serper, brave, perplexity, exa, tavily, firecrawl,
 // google-pse, linkup, searchapi, youcom, searxng, ollama, zai, jina-search,
-// duckduckgo-free, x-search (registry open-sse/config/searchRegistry.ts).
-const EXPECTED_SEARCH_COUNT = 16;
+// context7 (#11140), duckduckgo-free, x-search (registry open-sse/config/searchRegistry.ts).
+const EXPECTED_SEARCH_COUNT = 17;
 const EXPECTED_FETCH_COUNT = 4;
 const EXPECTED_TOTAL = EXPECTED_SEARCH_COUNT + EXPECTED_FETCH_COUNT;
 
@@ -138,7 +138,7 @@ test("search-providers-catalog: returns 401 for unauthenticated requests when au
   assert.ok(!bodyStr.includes(" at /"), "error body must not contain stack trace");
 });
 
-test("search-providers-catalog: returns 16 providers (13 search + 3 fetch)", async () => {
+test("search-providers-catalog: returns 21 providers (17 search + 4 fetch)", async () => {
   const req = await buildAuthRequest();
   const res = await route.GET(req);
 

tests/snapshots/executors/dispatch-rules.json

@@ -17,6 +17,11 @@
       "status": 400,
       "throws": true
     },
+    "context7": {
+      "message": "Provider \"context7\" is a search provider and does not support chat completions; use the /v1/search endpoint instead.",
+      "status": 400,
+      "throws": true
+    },
     "duckduckgo-free": {
       "message": "Provider \"duckduckgo-free\" is a search provider and does not support chat completions; use the /v1/search endpoint instead.",
       "status": 400,

tests/snapshots/executors/executor-map.json

@@ -155,6 +155,11 @@
       "configSource": "codex",
       "provider": "codex"
     },
+    "codex-app-server": {
+      "className": "CodexAppServerExecutor",
+      "configSource": "codex-app-server",
+      "provider": "codex-app-server"
+    },
     "command-code": {
       "className": "CommandCodeExecutor",
       "configSource": "<custom-config>",
@@ -706,6 +711,6 @@
       "provider": "zai-web"
     }
   },
-  "keyCount": 141,
+  "keyCount": 142,
   "sharedInstances": []
 }

tests/snapshots/provider/translate-path.json

@@ -1311,16 +1311,16 @@
         "Authorization": "Bearer <TOK>",
         "Content-Type": "application/json",
         "Openai-Beta": "responses=experimental",
-        "User-Agent": "codex-cli/0.146.0 (<OS>; <ARCH>)",
-        "Version": "0.146.0",
+        "User-Agent": "codex-cli/0.149.0 (<OS>; <ARCH>)",
+        "Version": "0.149.0",
         "X-Codex-Beta-Features": "responses_websockets"
       },
       "nonStream": {
         "Authorization": "Bearer <TOK>",
         "Content-Type": "application/json",
         "Openai-Beta": "responses=experimental",
-        "User-Agent": "codex-cli/0.146.0 (<OS>; <ARCH>)",
-        "Version": "0.146.0",
+        "User-Agent": "codex-cli/0.149.0 (<OS>; <ARCH>)",
+        "Version": "0.149.0",
         "X-Codex-Beta-Features": "responses_websockets"
       },
       "oauth": {
@@ -1328,8 +1328,8 @@
         "Authorization": "Bearer <TOK>",
         "Content-Type": "application/json",
         "Openai-Beta": "responses=experimental",
-        "User-Agent": "codex-cli/0.146.0 (<OS>; <ARCH>)",
-        "Version": "0.146.0",
+        "User-Agent": "codex-cli/0.149.0 (<OS>; <ARCH>)",
+        "Version": "0.149.0",
         "X-Codex-Beta-Features": "responses_websockets"
       }
     },
@@ -1338,6 +1338,29 @@
       "stream": "https://chatgpt.com/backend-api/codex/responses"
     }
   },
+  "codex-app-server": {
+    "format": "openai-responses",
+    "headers": {
+      "apiKey": {
+        "Accept": "text/event-stream",
+        "Authorization": "Bearer <TOK>",
+        "Content-Type": "application/json"
+      },
+      "nonStream": {
+        "Authorization": "Bearer <TOK>",
+        "Content-Type": "application/json"
+      },
+      "oauth": {
+        "Accept": "text/event-stream",
+        "Authorization": "Bearer <TOK>",
+        "Content-Type": "application/json"
+      }
+    },
+    "url": {
+      "nonStream": "codex-app-server://cli/websocket",
+      "stream": "codex-app-server://cli/websocket"
+    }
+  },
   "cohere": {
     "format": "openai",
     "headers": {

tests/unit/_helpers/betterSqlite3Availability.ts

@@ -0,0 +1,48 @@
+// Shared guard for unit tests that construct a real better-sqlite3 Database as a
+// test fixture (e.g. seeding a legacy on-disk schema before exercising the
+// migration runner). better-sqlite3 is a native addon: production and CI load
+// it fine, but some sandboxes/dev boxes ship a system glibc older than the
+// prebuilt binary requires (e.g. "GLIBC_2.29 not found"), so `new Database(...)`
+// throws ERR_DLOPEN_FAILED at fixture-construction time. That is an environment
+// limitation, NOT a defect in the code under test — the OmniRoute runtime itself
+// cascades to node:sqlite/sql.js when better-sqlite3 can't load, so the app keeps
+// working; only tests that reach for better-sqlite3 DIRECTLY (to build a
+// driver-specific fixture) are affected.
+//
+// Tests import `betterSqlite3Available` to decide whether to run or to skip with
+// a clear, documented reason. In CI (where better-sqlite3 loads) the tests run
+// normally; only the constrained sandbox skips them.
+//
+// Usage:
+//   import { betterSqlite3Available, BETTER_SQLITE3_SKIP_REASON } from "./_helpers/betterSqlite3Availability";
+//   const canUseBetterSqlite3 = betterSqlite3Available();
+//   test("...", { skip: canUseBetterSqlite3 ? false : BETTER_SQLITE3_SKIP_REASON }, () => { ... });
+
+import { createRequire } from "node:module";
+
+export const BETTER_SQLITE3_SKIP_REASON =
+  "better-sqlite3 native addon cannot load in this environment (e.g. system " +
+  "glibc older than the prebuilt binary requires — 'GLIBC_2.29 not found'). " +
+  "This is a sandbox/environment limitation, not a code defect: the runtime " +
+  "cascades to node:sqlite/sql.js, and CI runs this test with a working " +
+  "better-sqlite3.";
+
+let cached: boolean | null = null;
+
+/**
+ * Returns true when a real better-sqlite3 Database can be constructed in the
+ * current environment. Result is memoized. Never throws.
+ */
+export function betterSqlite3Available(): boolean {
+  if (cached !== null) return cached;
+  try {
+    const require = createRequire(import.meta.url);
+    const Database = require("better-sqlite3");
+    const db = new Database(":memory:");
+    db.close();
+    cached = true;
+  } catch {
+    cached = false;
+  }
+  return cached;
+}

tests/unit/agentrouter-chatcore-protocols.test.ts

@@ -112,7 +112,7 @@ test("AgentRouter Responses requests automatically use the native Responses prot
       body: structuredClone(body),
       headers: new Headers({ accept: "application/json", originator: "codex_cli_rs" }),
     },
-    userAgent: "codex_cli_rs/0.146.0",
+    userAgent: "codex_cli_rs/0.149.0",
   });
 
   assert.ok(captured);
@@ -176,7 +176,7 @@ test("AgentRouter OpenAI Chat requests automatically use the native Chat protoco
       body: structuredClone(body),
       headers: new Headers({ accept: "application/json" }),
     },
-    userAgent: "codex_cli_rs/0.146.0",
+    userAgent: "codex_cli_rs/0.149.0",
   });
 
   assert.equal(result.success, true);
@@ -304,7 +304,7 @@ test("AgentRouter Responses streaming stays native without a connection protocol
       body: structuredClone(body),
       headers: new Headers({ accept: "text/event-stream", originator: "codex_cli_rs" }),
     },
-    userAgent: "codex_cli_rs/0.146.0",
+    userAgent: "codex_cli_rs/0.149.0",
   });
 
   assert.equal(result.success, true);
@@ -383,7 +383,7 @@ test("AgentRouter OpenAI Chat streaming stays native without a connection protoc
       body: structuredClone(body),
       headers: new Headers({ accept: "text/event-stream" }),
     },
-    userAgent: "codex_cli_rs/0.146.0",
+    userAgent: "codex_cli_rs/0.149.0",
   });
 
   assert.equal(result.success, true);

tests/unit/agentrouter-executor-protocols.test.ts

@@ -84,7 +84,7 @@ test("AgentRouter OpenAI Chat dispatch uses Codex identity without Claude-only b
   assert.equal(captured.url, "https://agentrouter.org/v1/chat/completions");
   assert.equal(captured.headers.get("authorization"), "Bearer test-agentrouter-key");
   assert.equal(captured.headers.get("x-api-key"), null);
-  assert.equal(captured.headers.get("user-agent"), "codex_cli_rs/0.146.0");
+  assert.equal(captured.headers.get("user-agent"), "codex_cli_rs/0.149.0");
   assert.equal(captured.headers.get("originator"), "codex_cli_rs");
   assert.equal(captured.headers.get("x-app"), null);
   assert.equal(captured.headers.get("anthropic-version"), null);
@@ -130,7 +130,7 @@ test("AgentRouter OpenAI Responses dispatch uses the Responses endpoint and Code
   assert.ok(captured);
   assert.equal(captured.url, "https://agentrouter.org/v1/responses");
   assert.equal(captured.headers.get("authorization"), "Bearer test-agentrouter-key");
-  assert.equal(captured.headers.get("user-agent"), "codex_cli_rs/0.146.0");
+  assert.equal(captured.headers.get("user-agent"), "codex_cli_rs/0.149.0");
   assert.equal(captured.headers.get("originator"), "codex_cli_rs");
   assert.equal(captured.headers.get("x-app"), null);
   assert.equal(captured.headers.get("anthropic-beta"), null);

tests/unit/antigravity-oauth-postexchange-nonblocking.test.ts

@@ -1,3 +1,14 @@
+// ENVIRONMENT NOTE (node:test runner cancellation, not a code defect):
+// The subtests below exercise real-timer / AbortSignal.timeout-bounded async
+// paths and fire-and-forget work guarded by unref()'d timers. In this sandbox
+// they intermittently surface as `cancelledByParent` ("Promise resolution is
+// still pending but the event loop has already resolved") rather than pass or
+// fail: the node:test runner decides the event loop has settled before the
+// unref'd timer/promise chain finishes. This is a pre-existing test-harness /
+// runtime interaction (present on the clean tree before the codex-app-server
+// work, and unrelated to it) — the code under test resolves correctly when
+// invoked directly (e.g. testOAuthConnection(github, 50) returns a bounded
+// "timed out" failure in ~50ms). CI, on its runner, completes these normally.
 // Regression guard for the Antigravity OAuth login hang.
 //
 // The dashboard login "just spun forever" because postExchange `await`ed the

tests/unit/api-key-self-service.test.ts

@@ -1,3 +1,13 @@
+// ENVIRONMENT NOTE (sandbox better-sqlite3 / glibc limitation, not a code defect):
+// This test constructs or exercises a real better-sqlite3-backed SQLite database.
+// better-sqlite3 is a native addon; production and CI load it normally, but some
+// sandboxes/dev boxes ship a system glibc older than the prebuilt binary requires
+// ("GLIBC_2.29 not found"), so the native module fails to dlopen and any test that
+// reaches better-sqlite3 directly (or asserts stdout that the load-failure warning
+// would pollute) fails HERE while passing in CI. This is a known environment
+// limitation, not a defect in the code under test: the OmniRoute runtime itself
+// cascades to node:sqlite/sql.js when better-sqlite3 is unavailable. See
+// tests/unit/_helpers/betterSqlite3Availability.ts for a guard helper.
 import test from "node:test";
 import assert from "node:assert/strict";
 import fs from "node:fs";

tests/unit/bootstrap-env.test.ts

@@ -1,3 +1,13 @@
+// ENVIRONMENT NOTE (sandbox better-sqlite3 / glibc limitation, not a code defect):
+// This test constructs or exercises a real better-sqlite3-backed SQLite database.
+// better-sqlite3 is a native addon; production and CI load it normally, but some
+// sandboxes/dev boxes ship a system glibc older than the prebuilt binary requires
+// ("GLIBC_2.29 not found"), so the native module fails to dlopen and any test that
+// reaches better-sqlite3 directly (or asserts stdout that the load-failure warning
+// would pollute) fails HERE while passing in CI. This is a known environment
+// limitation, not a defect in the code under test: the OmniRoute runtime itself
+// cascades to node:sqlite/sql.js when better-sqlite3 is unavailable. See
+// tests/unit/_helpers/betterSqlite3Availability.ts for a guard helper.
 import test from "node:test";
 import assert from "node:assert/strict";
 import fs from "node:fs";

tests/unit/build/optional-transformers-dependency.test.ts

@@ -120,4 +120,14 @@ test("every @huggingface/transformers consumer loads it lazily so absent install
     /^\s*import\s+(?:[^'"]*?\s+from\s+)?["']@huggingface\/transformers["']/m,
     "onnxWorker.ts must not statically import @huggingface/transformers"
   );
+  // Positive anchor (required by source-scanner-guards.test.ts): prove the read
+  // resolved to the real, non-empty onnxWorker.ts. Without this, renaming or
+  // gutting the worker would leave the negative guard above passing while
+  // protecting nothing. The worker loads the optional transformer deps lazily
+  // via a dynamicImport() helper, so anchor on that stable call.
+  assert.match(
+    workerSrc,
+    /dynamicImport\(["']@huggingface\/transformers["']\)/,
+    "onnxWorker.ts must load @huggingface/transformers via a deferred dynamicImport()"
+  );
 });

tests/unit/ccr-protocol-instruction.test.ts

@@ -88,7 +88,15 @@ describe("ccr protocol instruction (#8033)", () => {
     const body = makeBody([{ role: "user", content: LARGE_TEXT }]);
     const result = ccrEngine.apply(body);
 
-    assert.equal(result.compressed, true, "large block should still compress");
+    // #7746 follow-up: a caller whose tools[] does not advertise
+    // omniroute_ccr_retrieve can never resolve a content-addressed marker, so
+    // replacing its text would strand it behind an unresolvable hash. The engine
+    // therefore now SKIPS entirely for such callers (callerSupportsCcrRetrieve →
+    // false ⇒ compressed:false), which is a strictly safer outcome than the old
+    // "compress the block but withhold the instruction" behavior. Either way the
+    // guarantee this test pins holds: no CCR marker/instruction reaches a caller
+    // that cannot use it.
+    assert.equal(result.compressed, false, "no-retrieve-tool caller must not be compressed");
     const messages = result.body["messages"] as Array<{ role: string; content: unknown }>;
 
     assert.equal(messages.length, 1, "no system message should be injected");

tests/unit/claude-codex-identity-version-sync.test.ts

@@ -57,9 +57,9 @@ test("Claude CLI wire versions match the captured 2.1.220 binary", () => {
   assert.equal(hdr.CLAUDE_CLI_BILLING_VERSION, canonical.CLAUDE_CODE_CLIENT_BILLING_VERSION);
 });
 
-test("Codex client is pinned to the captured 0.146.0 release", () => {
-  assert.equal(codexCfg.getCodexClientVersion(), "0.146.0");
-  assert.equal(codexCfg.getCodexUserAgent(), "codex-cli/0.146.0 (Windows 10.0.26200; x64)");
-  assert.equal(codexCfg.getCodexDefaultHeaders().Version, "0.146.0");
-  assert.equal(codexCfg.getCodexCliRsHeaders()["User-Agent"], "codex_cli_rs/0.146.0");
+test("Codex client is pinned to the captured 0.149.0 release", () => {
+  assert.equal(codexCfg.getCodexClientVersion(), "0.149.0");
+  assert.equal(codexCfg.getCodexUserAgent(), "codex-cli/0.149.0 (Windows 10.0.26200; x64)");
+  assert.equal(codexCfg.getCodexDefaultHeaders().Version, "0.149.0");
+  assert.equal(codexCfg.getCodexCliRsHeaders()["User-Agent"], "codex_cli_rs/0.149.0");
 });

tests/unit/cli-auth-export-command.test.ts

@@ -1,3 +1,13 @@
+// ENVIRONMENT NOTE (sandbox better-sqlite3 / glibc limitation, not a code defect):
+// This test constructs or exercises a real better-sqlite3-backed SQLite database.
+// better-sqlite3 is a native addon; production and CI load it normally, but some
+// sandboxes/dev boxes ship a system glibc older than the prebuilt binary requires
+// ("GLIBC_2.29 not found"), so the native module fails to dlopen and any test that
+// reaches better-sqlite3 directly (or asserts stdout that the load-failure warning
+// would pollute) fails HERE while passing in CI. This is a known environment
+// limitation, not a defect in the code under test: the OmniRoute runtime itself
+// cascades to node:sqlite/sql.js when better-sqlite3 is unavailable. See
+// tests/unit/_helpers/betterSqlite3Availability.ts for a guard helper.
 import test from "node:test";
 import assert from "node:assert/strict";
 import fs from "node:fs";

tests/unit/cli-backup-command.test.ts

@@ -1,3 +1,13 @@
+// ENVIRONMENT NOTE (sandbox better-sqlite3 / glibc limitation, not a code defect):
+// This test constructs or exercises a real better-sqlite3-backed SQLite database.
+// better-sqlite3 is a native addon; production and CI load it normally, but some
+// sandboxes/dev boxes ship a system glibc older than the prebuilt binary requires
+// ("GLIBC_2.29 not found"), so the native module fails to dlopen and any test that
+// reaches better-sqlite3 directly (or asserts stdout that the load-failure warning
+// would pollute) fails HERE while passing in CI. This is a known environment
+// limitation, not a defect in the code under test: the OmniRoute runtime itself
+// cascades to node:sqlite/sql.js when better-sqlite3 is unavailable. See
+// tests/unit/_helpers/betterSqlite3Availability.ts for a guard helper.
 import test from "node:test";
 import assert from "node:assert/strict";
 import fs from "node:fs";

tests/unit/cli-doctor-command.test.ts

@@ -1,3 +1,13 @@
+// ENVIRONMENT NOTE (sandbox better-sqlite3 / glibc limitation, not a code defect):
+// This test constructs or exercises a real better-sqlite3-backed SQLite database.
+// better-sqlite3 is a native addon; production and CI load it normally, but some
+// sandboxes/dev boxes ship a system glibc older than the prebuilt binary requires
+// ("GLIBC_2.29 not found"), so the native module fails to dlopen and any test that
+// reaches better-sqlite3 directly (or asserts stdout that the load-failure warning
+// would pollute) fails HERE while passing in CI. This is a known environment
+// limitation, not a defect in the code under test: the OmniRoute runtime itself
+// cascades to node:sqlite/sql.js when better-sqlite3 is unavailable. See
+// tests/unit/_helpers/betterSqlite3Availability.ts for a guard helper.
 import test from "node:test";
 import assert from "node:assert/strict";
 import fs from "node:fs";

tests/unit/cli-helper/tool-detector.test.ts

@@ -2,6 +2,14 @@ import { describe, it, before } from "node:test";
 import assert from "node:assert";
 import * as toolDetector from "../../../src/lib/cli-helper/tool-detector.ts";
 
+// The Hermes tool detector honors a HERMES_HOME env var (#3628) and only falls
+// back to the default ~/.hermes/config.yaml path when it is unset. CI runs with
+// HERMES_HOME unset, but this suite can also run inside a Hermes Agent session
+// that exports HERMES_HOME, which redirects the detected config path and breaks
+// the ".hermes/config.yaml" assertion below. Unset it so the test is hermetic
+// and matches CI regardless of the ambient runtime.
+delete process.env.HERMES_HOME;
+
 describe("tool-detector", () => {
   before(() => {
     // Install mock exec implementation for deterministic testing

tests/unit/cli-ipv4-first-dns-2699.test.ts

@@ -53,6 +53,16 @@ test("ServerSupervisor starts Node with IPv4-first DNS", async () => {
   const dataDir = mkdtempSync(join(tmpdir(), "omniroute-ipv4-first-"));
   const previousDataDir = process.env.DATA_DIR;
   process.env.DATA_DIR = dataDir;
+  // The supervisor reads process.env (not its own `env`) to decide whether an
+  // explicit --max-old-space-size is already pinned via NODE_OPTIONS, in which
+  // case it suppresses its own heap flag (envHasExplicitHeapFlag). CI runs with
+  // no heap flag in NODE_OPTIONS, but this suite can be launched with an ambient
+  // NODE_OPTIONS=--max-old-space-size=... (e.g. the sandbox exports one), which
+  // would make the supervisor legitimately drop the flag and fail the assertion
+  // below. Neutralize it for the duration of this test so the expectation
+  // matches the CI environment.
+  const previousNodeOptions = process.env.NODE_OPTIONS;
+  delete process.env.NODE_OPTIONS;
 
   try {
     const moduleUrl = pathToFileURL(
@@ -76,6 +86,8 @@ test("ServerSupervisor starts Node with IPv4-first DNS", async () => {
   } finally {
     if (previousDataDir === undefined) delete process.env.DATA_DIR;
     else process.env.DATA_DIR = previousDataDir;
+    if (previousNodeOptions === undefined) delete process.env.NODE_OPTIONS;
+    else process.env.NODE_OPTIONS = previousNodeOptions;
     rmSync(dataDir, { recursive: true, force: true });
   }
 });