feat: v2.4.0 — per-model circuit breaker, adaptive timeout tiers, structured logging

- Circuit breaker: consecutive timeouts (default 3) mark model as degraded for 60s,
  failing fast instead of blocking gateway with repeated timeout attempts
- Per-model timeout tiers: Opus 60s base, Sonnet 45s, Haiku 30s, plus adaptive
  scaling by prompt size (~15s/100k chars for Opus)
- Structured JSON logging for spawns, timeouts, breaker state changes
- Late close guard: prevents double-settle when timeout fires before proc.close

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

openclaw-claude-proxy/README.md

@@ -0,0 +1,182 @@
+# openclaw-claude-proxy v2.3.0
+
+Use your **Claude Pro / Max** subscription as an **OpenAI-compatible local endpoint**.
+
+`openclaw-claude-proxy` accepts OpenAI-style chat completion requests, then runs them through the local `claude` CLI. That means tools which only know how to talk to an OpenAI API can still use Claude models through a local base URL.
+
+## Why v2 matters
+
+v2 is not just a bugfix release. It changes the runtime model:
+
+- **On-demand spawning** instead of fragile warm pools
+- **Session resume** support for multi-turn conversations
+- **Faster fallback** with first-byte timeout + lower default request timeout
+- **Full tool access** via configurable allowed tools
+- **MCP config + system prompt pass-through**
+- **Health / sessions / diagnostics endpoints**
+- **Safe coexistence with Claude Code channel / interactive mode**
+
+## The short pitch
+
+If Claude's new channel workflow feels useful, OCP v2 now covers the same practical ground for many local agent/tooling setups:
+
+- multi-turn continuity
+- tool-enabled Claude runs
+- local orchestration
+- stable process isolation
+- coexistence with your normal Claude Code workflow
+
+And it adds a few advantages that channel users usually still want:
+
+- **OpenAI-compatible HTTP API** for existing tools
+- **Works with OpenClaw, Cursor, Continue, Open WebUI, LangChain, and anything with custom base URL support**
+- **Explicit health checks and diagnostics**
+- **Model/provider failover can happen outside Claude itself**
+- **No lock-in to a single client UX**
+
+## Coexistence with Claude Code channel
+
+This is the important part: **OCP v2 does not replace Claude Code channel, and it does not need to. They can coexist on the same machine.**
+
+### Claude Code channel / interactive mode
+- persistent interactive workflow
+- MCP protocol / in-process experience
+- great when you are directly driving Claude Code
+
+### OCP v2
+- local HTTP server on `localhost`
+- OpenAI-compatible API surface
+- per-request `claude -p` execution with session resume when you want continuity
+- ideal for external tools, routers, orchestrators, OpenClaw providers, and local automation
+
+### Practical takeaway
+Use both:
+- use **Claude Code channel** when you want Claude's native interactive workflow
+- use **OCP v2** when another app expects an OpenAI-style API but you still want to use Claude
+
+They solve adjacent problems, not identical ones.
+
+## Unique advantages of OCP v2
+
+1. **API compatibility**
+   - Drop into tools that already support OpenAI-compatible endpoints.
+   - No need to wait for each tool to add native Claude channel support.
+
+2. **Routing freedom**
+   - Put OCP behind OpenClaw or another router.
+   - Mix Claude with fallback providers outside the Claude client itself.
+
+3. **Operational visibility**
+   - `/health`, `/sessions`, recent errors, auth state, resolved binary path, timeout config.
+   - Much easier to debug than a black-box local integration.
+
+4. **Safer runtime model**
+   - v2 removes the old pre-spawn pool crash loop.
+   - No stale workers, no degraded warm pool states, fewer hidden failure modes.
+
+5. **Configurable tools and behavior**
+   - allowed tools
+   - skip permissions mode
+   - system prompt append
+   - MCP config passthrough
+   - session TTL
+   - concurrency limits
+
+## Install
+
+```bash
+git clone https://github.com/dtzp555-max/openclaw-claude-proxy
+cd openclaw-claude-proxy
+npm install
+node server.mjs
+```
+
+Default base URL:
+
+```text
+http://127.0.0.1:3456/v1
+```
+
+## Quick OpenAI-compatible config
+
+```json
+{
+  "baseURL": "http://127.0.0.1:3456/v1",
+  "apiKey": "anything"
+}
+```
+
+If `PROXY_API_KEY` is unset, auth is disabled. If you set it, pass it as a Bearer token.
+
+## Environment variables
+
+| Variable | Default | Purpose |
+|---|---:|---|
+| `CLAUDE_PROXY_PORT` | `3456` | Listen port |
+| `CLAUDE_BIN` | auto-detect | Claude CLI binary path |
+| `CLAUDE_TIMEOUT` | `120000` | Overall per-request timeout |
+| `CLAUDE_FIRST_BYTE_TIMEOUT` | `30000` | Abort if Claude produces no stdout quickly |
+| `CLAUDE_ALLOWED_TOOLS` | expanded set | Comma-separated allowed tools |
+| `CLAUDE_SKIP_PERMISSIONS` | `false` | Bypass permission checks |
+| `CLAUDE_SYSTEM_PROMPT` | unset | Append a system prompt to every request |
+| `CLAUDE_MCP_CONFIG` | unset | Path to MCP config JSON |
+| `CLAUDE_SESSION_TTL` | `3600000` | Session TTL |
+| `CLAUDE_MAX_CONCURRENT` | `5` | Max concurrent Claude processes |
+| `PROXY_API_KEY` | unset | Optional Bearer token auth |
+
+## Endpoints
+
+- `GET /health`
+- `GET /v1/models`
+- `POST /v1/chat/completions`
+- `GET /sessions`
+- `DELETE /sessions`
+
+## Example health response highlights
+
+`/health` reports useful operational state such as:
+
+- resolved Claude binary path
+- whether the binary is executable
+- auth status
+- timeouts
+- current sessions
+- recent errors
+- basic request stats
+
+## Version highlights
+
+### v2.3.0
+- clarified v2 positioning and coexistence story in docs
+- officially documents faster fallback defaults
+- recommends OCP v2 as the API bridge layer for Claude-powered tools
+
+### v2.2.0
+- first-byte timeout
+- reduced default timeout for faster fallback
+
+### v2.0.0
+- on-demand architecture
+- session management
+- full tool access
+- MCP + system prompt passthrough
+- concurrency control
+- coexistence with Claude Code interactive mode
+
+## When to use OCP v2 vs Claude channel
+
+Choose **OCP v2** when:
+- your app only supports OpenAI-compatible endpoints
+- you want routing / failover outside Claude
+- you want explicit health checks and local diagnostics
+- you want Claude available to multiple local tools through one endpoint
+
+Choose **Claude channel** when:
+- you are primarily living inside Claude Code itself
+- you want Claude's native interactive workflow directly
+
+Use **both together** when you want the best of both worlds.
+
+---
+
+If you already pay for Claude Pro or Max, OCP v2 turns that subscription into a practical local API bridge for the rest of your tooling stack.

openclaw-claude-proxy/package.json

@@ -1,7 +1,7 @@
 {
   "name": "openclaw-claude-proxy",
-  "version": "2.2.0",
+  "version": "2.4.0",
-  "description": "OpenAI-compatible proxy that routes requests through Claude CLI \u2014 use your Claude Pro/Max subscription as an OpenClaw model provider",
+  "description": "OpenAI-compatible proxy for Claude CLI v2 — per-model circuit breaker, adaptive first-byte timeout, structured logging",
   "type": "module",
   "bin": {
     "openclaw-claude-proxy": "./server.mjs"

openclaw-claude-proxy/server.mjs

@@ -1,29 +1,29 @@
 #!/usr/bin/env node
 /**
- * openclaw-claude-proxy v2.2.0 — OpenAI-compatible proxy for Claude CLI
+ * openclaw-claude-proxy v2.4.0 — OpenAI-compatible proxy for Claude CLI
  *
  * Translates OpenAI chat/completions requests into `claude -p` CLI calls,
  * letting you use your Claude Pro/Max subscription as an OpenClaw model provider.
  *
- * v2.0.0 highlights:
+ * v2.4.0:
- *   - On-demand spawning: eliminates pool crash loops from v1.x
+ *   - Per-model circuit breaker: consecutive timeouts temporarily mark a model as degraded
- *   - Session management: --resume support reduces token waste on multi-turn
+ *   - Adaptive first-byte timeout: scales by model tier + prompt size
- *   - Full tool access: configurable allowedTools (expanded defaults)
+ *   - Structured JSON logging for key events (easier to parse/alert on)
- *   - System prompt & MCP config pass-through
+ *   - On-demand spawning (no pool), session management, full tool access
- *   - Concurrency control with queuing
- *   - Coexists safely with Claude Code interactive mode (Telegram, IDE, etc.)
  *
  * Env vars:
  *   CLAUDE_PROXY_PORT        — listen port (default: 3456)
  *   CLAUDE_BIN               — path to claude binary (default: auto-detect)
  *   CLAUDE_TIMEOUT           — per-request timeout in ms (default: 120000)
- *   CLAUDE_FIRST_BYTE_TIMEOUT — abort if no stdout within this ms (default: 30000)
+ *   CLAUDE_FIRST_BYTE_TIMEOUT — base first-byte timeout in ms (default: 45000)
  *   CLAUDE_ALLOWED_TOOLS     — comma-separated tools to allow (default: expanded set)
  *   CLAUDE_SKIP_PERMISSIONS  — "true" to bypass all permission checks (default: false)
  *   CLAUDE_SYSTEM_PROMPT     — system prompt appended to all requests
  *   CLAUDE_MCP_CONFIG        — path to MCP server config JSON file
  *   CLAUDE_SESSION_TTL       — session TTL in ms (default: 3600000 = 1h)
  *   CLAUDE_MAX_CONCURRENT    — max concurrent claude processes (default: 5)
+ *   CLAUDE_BREAKER_THRESHOLD — consecutive timeouts before circuit opens (default: 3)
+ *   CLAUDE_BREAKER_COOLDOWN  — ms to wait before retrying after circuit opens (default: 60000)
  *   PROXY_API_KEY            — Bearer token for API auth (optional)
  */
 import { createServer } from "node:http";
@@ -77,7 +77,7 @@ function resolveClaude() {
 const PORT = parseInt(process.env.CLAUDE_PROXY_PORT || "3456", 10);
 const CLAUDE = resolveClaude();
 const TIMEOUT = parseInt(process.env.CLAUDE_TIMEOUT || "120000", 10);
-const FIRST_BYTE_TIMEOUT = parseInt(process.env.CLAUDE_FIRST_BYTE_TIMEOUT || "30000", 10);
+const BASE_FIRST_BYTE_TIMEOUT = parseInt(process.env.CLAUDE_FIRST_BYTE_TIMEOUT || "45000", 10);
 const PROXY_API_KEY = process.env.PROXY_API_KEY || "";
 const SKIP_PERMISSIONS = process.env.CLAUDE_SKIP_PERMISSIONS === "true";
 const ALLOWED_TOOLS = (process.env.CLAUDE_ALLOWED_TOOLS ||
@@ -87,10 +87,65 @@ const SYSTEM_PROMPT = process.env.CLAUDE_SYSTEM_PROMPT || "";
 const MCP_CONFIG = process.env.CLAUDE_MCP_CONFIG || "";
 const SESSION_TTL = parseInt(process.env.CLAUDE_SESSION_TTL || "3600000", 10);
 const MAX_CONCURRENT = parseInt(process.env.CLAUDE_MAX_CONCURRENT || "5", 10);
+const BREAKER_THRESHOLD = parseInt(process.env.CLAUDE_BREAKER_THRESHOLD || "3", 10);
+const BREAKER_COOLDOWN = parseInt(process.env.CLAUDE_BREAKER_COOLDOWN || "60000", 10);
 
 const VERSION = _pkg.version;
 const START_TIME = Date.now();
 
+// ── Structured logging helper ───────────────────────────────────────────
+function logEvent(level, event, data = {}) {
+  const entry = { ts: new Date().toISOString(), level, event, ...data };
+  if (level === "error" || level === "warn") {
+    console.error(JSON.stringify(entry));
+  } else {
+    console.log(JSON.stringify(entry));
+  }
+}
+
+// ── Per-model circuit breaker ───────────────────────────────────────────
+// Tracks consecutive timeouts per model. When threshold is reached, the
+// model is marked "open" (degraded) for BREAKER_COOLDOWN ms. During that
+// window, requests for this model fail fast with a clear error instead of
+// waiting for yet another timeout that would block the gateway.
+const breakers = new Map(); // cliModel → { failures, state, openedAt }
+
+function getBreakerState(cliModel) {
+  if (!breakers.has(cliModel)) {
+    breakers.set(cliModel, { failures: 0, state: "closed", openedAt: 0 });
+  }
+  const b = breakers.get(cliModel);
+
+  // Auto-recover: if cooldown has elapsed, transition to half-open
+  if (b.state === "open" && Date.now() - b.openedAt >= BREAKER_COOLDOWN) {
+    b.state = "half-open";
+    logEvent("info", "breaker_half_open", { model: cliModel, cooldownMs: BREAKER_COOLDOWN });
+  }
+  return b;
+}
+
+function breakerRecordSuccess(cliModel) {
+  const b = getBreakerState(cliModel);
+  if (b.failures > 0 || b.state !== "closed") {
+    logEvent("info", "breaker_reset", { model: cliModel, previousFailures: b.failures, previousState: b.state });
+  }
+  b.failures = 0;
+  b.state = "closed";
+  b.openedAt = 0;
+}
+
+function breakerRecordTimeout(cliModel) {
+  const b = getBreakerState(cliModel);
+  b.failures++;
+  logEvent("warn", "breaker_failure", { model: cliModel, consecutiveFailures: b.failures, threshold: BREAKER_THRESHOLD });
+
+  if (b.failures >= BREAKER_THRESHOLD && b.state !== "open") {
+    b.state = "open";
+    b.openedAt = Date.now();
+    logEvent("error", "breaker_open", { model: cliModel, failures: b.failures, cooldownMs: BREAKER_COOLDOWN });
+  }
+}
+
 // ── Model mapping ───────────────────────────────────────────────────────
 // Maps request model IDs and aliases to canonical claude CLI model IDs.
 const MODEL_MAP = {
@@ -210,6 +265,26 @@ function messagesToPrompt(messages) {
   }).join("\n\n");
 }
 
+// Model tier multipliers for first-byte timeout.
+// Opus is much slower to produce first token, especially with large contexts.
+const MODEL_TIMEOUT_TIERS = {
+  "opus": { base: 60000, perPromptChar: 0.00015 },    // 60s base + ~15s per 100k chars
+  "sonnet": { base: 45000, perPromptChar: 0.00008 },  // 45s base + ~8s per 100k chars
+  "haiku": { base: 30000, perPromptChar: 0.00005 },   // 30s base + ~5s per 100k chars
+};
+
+function getModelTier(cliModel) {
+  if (cliModel.includes("opus")) return "opus";
+  if (cliModel.includes("haiku")) return "haiku";
+  return "sonnet";
+}
+
+function computeFirstByteTimeout(cliModel, promptLength) {
+  const tier = MODEL_TIMEOUT_TIERS[getModelTier(cliModel)];
+  const timeout = tier.base + Math.floor(promptLength * tier.perPromptChar);
+  return Math.min(timeout, Math.max(TIMEOUT - 5000, 10000));
+}
+
 // ── Call claude CLI ─────────────────────────────────────────────────────
 // On-demand spawning: each request spawns a fresh `claude -p` process.
 // No pool = no crash loops, no stale workers, no degraded states.
@@ -219,10 +294,20 @@ function callClaude(model, messages, conversationId) {
     if (stats.activeRequests >= MAX_CONCURRENT) {
       return reject(new Error(`concurrency limit reached (${stats.activeRequests}/${MAX_CONCURRENT})`));
     }
+
+    const cliModel = MODEL_MAP[model] || model;
+
+    // Circuit breaker check: fail fast if model is in open state
+    const breaker = getBreakerState(cliModel);
+    if (breaker.state === "open") {
+      const remainingMs = BREAKER_COOLDOWN - (Date.now() - breaker.openedAt);
+      logEvent("warn", "breaker_rejected", { model: cliModel, remainingCooldownMs: remainingMs });
+      return reject(new Error(`circuit breaker open for ${cliModel}: ${breaker.failures} consecutive timeouts, retry in ${Math.ceil(remainingMs / 1000)}s`));
+    }
+
     stats.activeRequests++;
     stats.totalRequests++;
 
-    const cliModel = MODEL_MAP[model] || model;
     let sessionInfo = null;
     let prompt;
 
@@ -271,6 +356,7 @@ function callClaude(model, messages, conversationId) {
     let stdout = "";
     let stderr = "";
     const t0 = Date.now();
+    const firstByteTimeoutMs = computeFirstByteTimeout(cliModel, prompt.length);
     let settled = false;
     let gotFirstByte = false;
 
@@ -306,13 +392,18 @@ function callClaude(model, messages, conversationId) {
     });
     proc.stderr.on("data", (d) => (stderr += d));
 
-    proc.on("close", (code) => {
+    proc.on("close", (code, signal) => {
       const elapsed = Date.now() - t0;
+      if (settled) {
+        logEvent("warn", "late_close", { model: cliModel, code, signal: signal || "none", elapsed });
+        return;
+      }
       if (code !== 0) {
-        console.error(`[claude] exit=${code} model=${cliModel} elapsed=${elapsed}ms stderr=${stderr.slice(0, 500)}`);
+        logEvent("error", "claude_exit", { model: cliModel, code, signal: signal || "none", elapsed, stderr: stderr.slice(0, 300) });
         settle(new Error(stderr.slice(0, 300) || stdout.slice(0, 300) || `claude exit ${code}`));
       } else {
-        console.log(`[claude] ok model=${cliModel} chars=${stdout.length} elapsed=${elapsed}ms session=${conversationId ? conversationId.slice(0, 12) + "..." : "none"}`);
+        breakerRecordSuccess(cliModel);
+        logEvent("info", "claude_ok", { model: cliModel, chars: stdout.length, elapsed, session: conversationId ? conversationId.slice(0, 12) + "..." : "none" });
         settle(null, stdout.trim());
       }
     });
@@ -326,24 +417,27 @@ function callClaude(model, messages, conversationId) {
     proc.stdin.write(prompt);
     proc.stdin.end();
 
-    console.log(`[claude] spawned model=${cliModel} prompt_chars=${prompt.length} session=${conversationId ? conversationId.slice(0, 12) + "..." : "none"}`);
+    logEvent("info", "claude_spawned", { model: cliModel, promptChars: prompt.length, firstByteTimeout: firstByteTimeoutMs, tier: getModelTier(cliModel), session: conversationId ? conversationId.slice(0, 12) + "..." : "none" });
 
     // First-byte timeout: abort early if Claude CLI produces no output
     const firstByteTimer = setTimeout(() => {
-      if (!gotFirstByte) {
+      if (!gotFirstByte && !settled) {
         stats.timeouts++;
-        console.error(`[claude] first-byte timeout after ${FIRST_BYTE_TIMEOUT}ms model=${cliModel} — aborting`);
+        breakerRecordTimeout(cliModel);
-        proc.kill("SIGTERM");
+        logEvent("error", "first_byte_timeout", { model: cliModel, timeoutMs: firstByteTimeoutMs, promptChars: prompt.length });
+        try { proc.kill("SIGTERM"); } catch {}
         setTimeout(() => { try { proc.kill("SIGKILL"); } catch {} }, 5000);
-        settle(new Error(`first-byte timeout after ${FIRST_BYTE_TIMEOUT}ms`));
+        settle(new Error(`first-byte timeout after ${firstByteTimeoutMs}ms`));
       }
-    }, FIRST_BYTE_TIMEOUT);
+    }, firstByteTimeoutMs);
 
     // Overall request timeout with graceful kill
     const timer = setTimeout(() => {
+      if (settled) return;
       stats.timeouts++;
-      console.error(`[claude] timeout after ${TIMEOUT}ms model=${cliModel}`);
+      breakerRecordTimeout(cliModel);
-      proc.kill("SIGTERM");
+      logEvent("error", "request_timeout", { model: cliModel, timeoutMs: TIMEOUT });
+      try { proc.kill("SIGTERM"); } catch {}
       setTimeout(() => { try { proc.kill("SIGKILL"); } catch {} }, 5000);
       settle(new Error(`timeout after ${TIMEOUT}ms`));
     }, TIMEOUT);
@@ -492,7 +586,7 @@ const server = createServer(async (req, res) => {
       auth: authStatus,
       config: {
         timeout: TIMEOUT,
-        firstByteTimeout: FIRST_BYTE_TIMEOUT,
+        firstByteTimeout: BASE_FIRST_BYTE_TIMEOUT,
         maxConcurrent: MAX_CONCURRENT,
         sessionTTL: SESSION_TTL,
         allowedTools: SKIP_PERMISSIONS ? "all (skip-permissions)" : ALLOWED_TOOLS,
@@ -535,7 +629,7 @@ server.listen(PORT, "0.0.0.0", () => {
   console.log(`Architecture: on-demand spawning (no pool)`);
   console.log(`Models: ${MODELS.map((m) => m.id).join(", ")}`);
   console.log(`Claude binary: ${CLAUDE}`);
-  console.log(`Timeout: ${TIMEOUT}ms (first-byte: ${FIRST_BYTE_TIMEOUT}ms) | Max concurrent: ${MAX_CONCURRENT}`);
+  console.log(`Timeout: ${TIMEOUT}ms (base first-byte: ${BASE_FIRST_BYTE_TIMEOUT}ms, adaptive by model/prompt) | Max concurrent: ${MAX_CONCURRENT}`);
   console.log(`Tools: ${SKIP_PERMISSIONS ? "all (skip-permissions)" : ALLOWED_TOOLS.join(", ")}`);
   console.log(`Sessions: TTL=${SESSION_TTL / 1000}s`);
   if (SYSTEM_PROMPT) console.log(`System prompt: "${SYSTEM_PROMPT.slice(0, 80)}..."`);

package.json

@@ -1,7 +1,7 @@
 {
   "name": "openclaw-claude-proxy",
-  "version": "2.2.0",
+  "version": "2.4.0",
-  "description": "OpenAI-compatible proxy that routes requests through Claude CLI \u2014 use your Claude Pro/Max subscription as an OpenClaw model provider",
+  "description": "OpenAI-compatible proxy for Claude CLI v2 — per-model circuit breaker, adaptive first-byte timeout, structured logging",
   "type": "module",
   "bin": {
     "openclaw-claude-proxy": "./server.mjs"

server.mjs

@@ -1,29 +1,29 @@
 #!/usr/bin/env node
 /**
- * openclaw-claude-proxy v2.2.0 — OpenAI-compatible proxy for Claude CLI
+ * openclaw-claude-proxy v2.4.0 — OpenAI-compatible proxy for Claude CLI
  *
  * Translates OpenAI chat/completions requests into `claude -p` CLI calls,
  * letting you use your Claude Pro/Max subscription as an OpenClaw model provider.
  *
- * v2.0.0 highlights:
+ * v2.4.0:
- *   - On-demand spawning: eliminates pool crash loops from v1.x
+ *   - Per-model circuit breaker: consecutive timeouts temporarily mark a model as degraded
- *   - Session management: --resume support reduces token waste on multi-turn
+ *   - Adaptive first-byte timeout: scales by model tier + prompt size
- *   - Full tool access: configurable allowedTools (expanded defaults)
+ *   - Structured JSON logging for key events (easier to parse/alert on)
- *   - System prompt & MCP config pass-through
+ *   - On-demand spawning (no pool), session management, full tool access
- *   - Concurrency control with queuing
- *   - Coexists safely with Claude Code interactive mode (Telegram, IDE, etc.)
  *
  * Env vars:
  *   CLAUDE_PROXY_PORT        — listen port (default: 3456)
  *   CLAUDE_BIN               — path to claude binary (default: auto-detect)
  *   CLAUDE_TIMEOUT           — per-request timeout in ms (default: 120000)
- *   CLAUDE_FIRST_BYTE_TIMEOUT — abort if no stdout within this ms (default: 30000)
+ *   CLAUDE_FIRST_BYTE_TIMEOUT — base first-byte timeout in ms (default: 45000)
  *   CLAUDE_ALLOWED_TOOLS     — comma-separated tools to allow (default: expanded set)
  *   CLAUDE_SKIP_PERMISSIONS  — "true" to bypass all permission checks (default: false)
  *   CLAUDE_SYSTEM_PROMPT     — system prompt appended to all requests
  *   CLAUDE_MCP_CONFIG        — path to MCP server config JSON file
  *   CLAUDE_SESSION_TTL       — session TTL in ms (default: 3600000 = 1h)
  *   CLAUDE_MAX_CONCURRENT    — max concurrent claude processes (default: 5)
+ *   CLAUDE_BREAKER_THRESHOLD — consecutive timeouts before circuit opens (default: 3)
+ *   CLAUDE_BREAKER_COOLDOWN  — ms to wait before retrying after circuit opens (default: 60000)
  *   PROXY_API_KEY            — Bearer token for API auth (optional)
  */
 import { createServer } from "node:http";
@@ -77,7 +77,7 @@ function resolveClaude() {
 const PORT = parseInt(process.env.CLAUDE_PROXY_PORT || "3456", 10);
 const CLAUDE = resolveClaude();
 const TIMEOUT = parseInt(process.env.CLAUDE_TIMEOUT || "120000", 10);
-const FIRST_BYTE_TIMEOUT = parseInt(process.env.CLAUDE_FIRST_BYTE_TIMEOUT || "30000", 10);
+const BASE_FIRST_BYTE_TIMEOUT = parseInt(process.env.CLAUDE_FIRST_BYTE_TIMEOUT || "45000", 10);
 const PROXY_API_KEY = process.env.PROXY_API_KEY || "";
 const SKIP_PERMISSIONS = process.env.CLAUDE_SKIP_PERMISSIONS === "true";
 const ALLOWED_TOOLS = (process.env.CLAUDE_ALLOWED_TOOLS ||
@@ -87,10 +87,65 @@ const SYSTEM_PROMPT = process.env.CLAUDE_SYSTEM_PROMPT || "";
 const MCP_CONFIG = process.env.CLAUDE_MCP_CONFIG || "";
 const SESSION_TTL = parseInt(process.env.CLAUDE_SESSION_TTL || "3600000", 10);
 const MAX_CONCURRENT = parseInt(process.env.CLAUDE_MAX_CONCURRENT || "5", 10);
+const BREAKER_THRESHOLD = parseInt(process.env.CLAUDE_BREAKER_THRESHOLD || "3", 10);
+const BREAKER_COOLDOWN = parseInt(process.env.CLAUDE_BREAKER_COOLDOWN || "60000", 10);
 
 const VERSION = _pkg.version;
 const START_TIME = Date.now();
 
+// ── Structured logging helper ───────────────────────────────────────────
+function logEvent(level, event, data = {}) {
+  const entry = { ts: new Date().toISOString(), level, event, ...data };
+  if (level === "error" || level === "warn") {
+    console.error(JSON.stringify(entry));
+  } else {
+    console.log(JSON.stringify(entry));
+  }
+}
+
+// ── Per-model circuit breaker ───────────────────────────────────────────
+// Tracks consecutive timeouts per model. When threshold is reached, the
+// model is marked "open" (degraded) for BREAKER_COOLDOWN ms. During that
+// window, requests for this model fail fast with a clear error instead of
+// waiting for yet another timeout that would block the gateway.
+const breakers = new Map(); // cliModel → { failures, state, openedAt }
+
+function getBreakerState(cliModel) {
+  if (!breakers.has(cliModel)) {
+    breakers.set(cliModel, { failures: 0, state: "closed", openedAt: 0 });
+  }
+  const b = breakers.get(cliModel);
+
+  // Auto-recover: if cooldown has elapsed, transition to half-open
+  if (b.state === "open" && Date.now() - b.openedAt >= BREAKER_COOLDOWN) {
+    b.state = "half-open";
+    logEvent("info", "breaker_half_open", { model: cliModel, cooldownMs: BREAKER_COOLDOWN });
+  }
+  return b;
+}
+
+function breakerRecordSuccess(cliModel) {
+  const b = getBreakerState(cliModel);
+  if (b.failures > 0 || b.state !== "closed") {
+    logEvent("info", "breaker_reset", { model: cliModel, previousFailures: b.failures, previousState: b.state });
+  }
+  b.failures = 0;
+  b.state = "closed";
+  b.openedAt = 0;
+}
+
+function breakerRecordTimeout(cliModel) {
+  const b = getBreakerState(cliModel);
+  b.failures++;
+  logEvent("warn", "breaker_failure", { model: cliModel, consecutiveFailures: b.failures, threshold: BREAKER_THRESHOLD });
+
+  if (b.failures >= BREAKER_THRESHOLD && b.state !== "open") {
+    b.state = "open";
+    b.openedAt = Date.now();
+    logEvent("error", "breaker_open", { model: cliModel, failures: b.failures, cooldownMs: BREAKER_COOLDOWN });
+  }
+}
+
 // ── Model mapping ───────────────────────────────────────────────────────
 // Maps request model IDs and aliases to canonical claude CLI model IDs.
 const MODEL_MAP = {
@@ -210,6 +265,26 @@ function messagesToPrompt(messages) {
   }).join("\n\n");
 }
 
+// Model tier multipliers for first-byte timeout.
+// Opus is much slower to produce first token, especially with large contexts.
+const MODEL_TIMEOUT_TIERS = {
+  "opus": { base: 60000, perPromptChar: 0.00015 },    // 60s base + ~15s per 100k chars
+  "sonnet": { base: 45000, perPromptChar: 0.00008 },  // 45s base + ~8s per 100k chars
+  "haiku": { base: 30000, perPromptChar: 0.00005 },   // 30s base + ~5s per 100k chars
+};
+
+function getModelTier(cliModel) {
+  if (cliModel.includes("opus")) return "opus";
+  if (cliModel.includes("haiku")) return "haiku";
+  return "sonnet";
+}
+
+function computeFirstByteTimeout(cliModel, promptLength) {
+  const tier = MODEL_TIMEOUT_TIERS[getModelTier(cliModel)];
+  const timeout = tier.base + Math.floor(promptLength * tier.perPromptChar);
+  return Math.min(timeout, Math.max(TIMEOUT - 5000, 10000));
+}
+
 // ── Call claude CLI ─────────────────────────────────────────────────────
 // On-demand spawning: each request spawns a fresh `claude -p` process.
 // No pool = no crash loops, no stale workers, no degraded states.
@@ -219,10 +294,20 @@ function callClaude(model, messages, conversationId) {
     if (stats.activeRequests >= MAX_CONCURRENT) {
       return reject(new Error(`concurrency limit reached (${stats.activeRequests}/${MAX_CONCURRENT})`));
     }
+
+    const cliModel = MODEL_MAP[model] || model;
+
+    // Circuit breaker check: fail fast if model is in open state
+    const breaker = getBreakerState(cliModel);
+    if (breaker.state === "open") {
+      const remainingMs = BREAKER_COOLDOWN - (Date.now() - breaker.openedAt);
+      logEvent("warn", "breaker_rejected", { model: cliModel, remainingCooldownMs: remainingMs });
+      return reject(new Error(`circuit breaker open for ${cliModel}: ${breaker.failures} consecutive timeouts, retry in ${Math.ceil(remainingMs / 1000)}s`));
+    }
+
     stats.activeRequests++;
     stats.totalRequests++;
 
-    const cliModel = MODEL_MAP[model] || model;
     let sessionInfo = null;
     let prompt;
 
@@ -271,6 +356,7 @@ function callClaude(model, messages, conversationId) {
     let stdout = "";
     let stderr = "";
     const t0 = Date.now();
+    const firstByteTimeoutMs = computeFirstByteTimeout(cliModel, prompt.length);
     let settled = false;
     let gotFirstByte = false;
 
@@ -306,13 +392,18 @@ function callClaude(model, messages, conversationId) {
     });
     proc.stderr.on("data", (d) => (stderr += d));
 
-    proc.on("close", (code) => {
+    proc.on("close", (code, signal) => {
       const elapsed = Date.now() - t0;
+      if (settled) {
+        logEvent("warn", "late_close", { model: cliModel, code, signal: signal || "none", elapsed });
+        return;
+      }
       if (code !== 0) {
-        console.error(`[claude] exit=${code} model=${cliModel} elapsed=${elapsed}ms stderr=${stderr.slice(0, 500)}`);
+        logEvent("error", "claude_exit", { model: cliModel, code, signal: signal || "none", elapsed, stderr: stderr.slice(0, 300) });
         settle(new Error(stderr.slice(0, 300) || stdout.slice(0, 300) || `claude exit ${code}`));
       } else {
-        console.log(`[claude] ok model=${cliModel} chars=${stdout.length} elapsed=${elapsed}ms session=${conversationId ? conversationId.slice(0, 12) + "..." : "none"}`);
+        breakerRecordSuccess(cliModel);
+        logEvent("info", "claude_ok", { model: cliModel, chars: stdout.length, elapsed, session: conversationId ? conversationId.slice(0, 12) + "..." : "none" });
         settle(null, stdout.trim());
       }
     });
@@ -326,24 +417,27 @@ function callClaude(model, messages, conversationId) {
     proc.stdin.write(prompt);
     proc.stdin.end();
 
-    console.log(`[claude] spawned model=${cliModel} prompt_chars=${prompt.length} session=${conversationId ? conversationId.slice(0, 12) + "..." : "none"}`);
+    logEvent("info", "claude_spawned", { model: cliModel, promptChars: prompt.length, firstByteTimeout: firstByteTimeoutMs, tier: getModelTier(cliModel), session: conversationId ? conversationId.slice(0, 12) + "..." : "none" });
 
     // First-byte timeout: abort early if Claude CLI produces no output
     const firstByteTimer = setTimeout(() => {
-      if (!gotFirstByte) {
+      if (!gotFirstByte && !settled) {
         stats.timeouts++;
-        console.error(`[claude] first-byte timeout after ${FIRST_BYTE_TIMEOUT}ms model=${cliModel} — aborting`);
+        breakerRecordTimeout(cliModel);
-        proc.kill("SIGTERM");
+        logEvent("error", "first_byte_timeout", { model: cliModel, timeoutMs: firstByteTimeoutMs, promptChars: prompt.length });
+        try { proc.kill("SIGTERM"); } catch {}
         setTimeout(() => { try { proc.kill("SIGKILL"); } catch {} }, 5000);
-        settle(new Error(`first-byte timeout after ${FIRST_BYTE_TIMEOUT}ms`));
+        settle(new Error(`first-byte timeout after ${firstByteTimeoutMs}ms`));
       }
-    }, FIRST_BYTE_TIMEOUT);
+    }, firstByteTimeoutMs);
 
     // Overall request timeout with graceful kill
     const timer = setTimeout(() => {
+      if (settled) return;
       stats.timeouts++;
-      console.error(`[claude] timeout after ${TIMEOUT}ms model=${cliModel}`);
+      breakerRecordTimeout(cliModel);
-      proc.kill("SIGTERM");
+      logEvent("error", "request_timeout", { model: cliModel, timeoutMs: TIMEOUT });
+      try { proc.kill("SIGTERM"); } catch {}
       setTimeout(() => { try { proc.kill("SIGKILL"); } catch {} }, 5000);
       settle(new Error(`timeout after ${TIMEOUT}ms`));
     }, TIMEOUT);
@@ -492,7 +586,7 @@ const server = createServer(async (req, res) => {
       auth: authStatus,
       config: {
         timeout: TIMEOUT,
-        firstByteTimeout: FIRST_BYTE_TIMEOUT,
+        firstByteTimeout: BASE_FIRST_BYTE_TIMEOUT,
         maxConcurrent: MAX_CONCURRENT,
         sessionTTL: SESSION_TTL,
         allowedTools: SKIP_PERMISSIONS ? "all (skip-permissions)" : ALLOWED_TOOLS,
@@ -535,7 +629,7 @@ server.listen(PORT, "0.0.0.0", () => {
   console.log(`Architecture: on-demand spawning (no pool)`);
   console.log(`Models: ${MODELS.map((m) => m.id).join(", ")}`);
   console.log(`Claude binary: ${CLAUDE}`);
-  console.log(`Timeout: ${TIMEOUT}ms (first-byte: ${FIRST_BYTE_TIMEOUT}ms) | Max concurrent: ${MAX_CONCURRENT}`);
+  console.log(`Timeout: ${TIMEOUT}ms (base first-byte: ${BASE_FIRST_BYTE_TIMEOUT}ms, adaptive by model/prompt) | Max concurrent: ${MAX_CONCURRENT}`);
   console.log(`Tools: ${SKIP_PERMISSIONS ? "all (skip-permissions)" : ALLOWED_TOOLS.join(", ")}`);
   console.log(`Sessions: TTL=${SESSION_TTL / 1000}s`);
   if (SYSTEM_PROMPT) console.log(`System prompt: "${SYSTEM_PROMPT.slice(0, 80)}..."`);