From abb5b90e1d4ac16fdf3f20b22c9be5dd62bf2bbc Mon Sep 17 00:00:00 2001
From: whit3rabbit <whit3rabbit@users.noreply.github.com>
Date: Wed, 25 Mar 2026 20:21:24 -0500
Subject: [PATCH] feat: add Bedrock backend, OpenTelemetry export, integration
 tests

- AWS Bedrock backend: SigV4 signing, InvokeModel + InvokeModelWithResponseStream
  with binary event stream decoding, passthrough handler for /v1/messages
- OpenTelemetry export: feature-gated (--features otel), OTLP/HTTP with
  reqwest transport, OtelGuard for graceful shutdown flush
- Chat completions integration tests: 6 tests covering non-streaming,
  error handling, degradation headers, system messages
- Updated COMPARISON_LITELLM.md to reflect all closed gaps
- Fixed Bedrock match arms across all handler files

549 tests passing, 0 failures, clippy clean.
Both `cargo build` and `cargo build --features otel` compile.

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
---
 CLAUDE.md                                     |  27 +-
 crates/proxy/Cargo.toml                       |   2 +
 crates/proxy/admin-ui/index.html              |  97 ++++-
 crates/proxy/src/backend/bedrock_client.rs    | 399 ++++++++++++++++++
 crates/proxy/src/backend/mod.rs               |  35 +-
 crates/proxy/src/backend/openai_client.rs     |   4 +-
 crates/proxy/src/config/mod.rs                | 124 +++++-
 crates/proxy/src/lib.rs                       |   3 +
 crates/proxy/src/main.rs                      |  16 +
 crates/proxy/src/otel.rs                      |  51 +++
 .../proxy/src/server/bedrock_passthrough.rs   | 240 +++++++++++
 crates/proxy/src/server/chat_completions.rs   |   8 +-
 crates/proxy/src/server/mod.rs                |   2 +
 crates/proxy/src/server/routes.rs             |  53 ++-
 crates/proxy/src/server/streaming.rs          |   4 +-
 crates/proxy/tests/chat_completions.rs        | 230 ++++++++++
 crates/proxy/tests/embeddings.rs              | 226 ++++++++++
 crates/proxy/tests/live_bedrock.rs            |  76 ++++
 crates/translator/src/anthropic/messages.rs   |  14 +-
 crates/translator/src/mapping/message_map.rs  | 177 +++++++-
 .../src/mapping/responses_message_map.rs      |   4 +-
 .../translator/src/mapping/streaming_map.rs   |  35 ++
 crates/translator/src/mapping/tools_map.rs    |  36 +-
 crates/translator/src/mapping/warnings.rs     |  60 +++
 docs/COMPARISON_LITELLM.md                    | 147 +++++++
 docs/ENV.md                                   |  50 +++
 .../20260325-120000-litellm-gap-fill/tasks.md |  28 +-
 27 files changed, 2079 insertions(+), 69 deletions(-)
 create mode 100644 crates/proxy/src/backend/bedrock_client.rs
 create mode 100644 crates/proxy/src/otel.rs
 create mode 100644 crates/proxy/src/server/bedrock_passthrough.rs
 create mode 100644 crates/proxy/tests/chat_completions.rs
 create mode 100644 crates/proxy/tests/embeddings.rs
 create mode 100644 crates/proxy/tests/live_bedrock.rs
 create mode 100644 crates/translator/src/mapping/warnings.rs
 create mode 100644 docs/COMPARISON_LITELLM.md

diff --git a/CLAUDE.md b/CLAUDE.md
index 8580e39..03e6748 100644
--- a/CLAUDE.md
+++ b/CLAUDE.md
@@ -12,11 +12,13 @@ All 11 implementation phases are complete.
 
 **Working (verified):**
 - Build: `cargo build` clean, `cargo clippy -- -D warnings` clean
-- Tests: ~417 tests passing, 4 ignored (live API)
+- Tests: ~480 tests passing, 4 ignored (live API)
 - Full Anthropic Messages API translation: non-streaming, streaming SSE, tool calling, file/document blocks
 - Proxy middleware: health, auth, request ID, size limits, concurrency limits, retry with backoff
 - Compatibility endpoints: /v1/models, count_tokens (approximate via tiktoken), batches (stub)
 - Model mapping and lossy-translation warnings
+- `POST /v1/embeddings` passthrough: forwards directly to the backend with no translation; works with OpenAI, Vertex, Gemini (`gemini-embedding-exp-03-07`), and vLLM/HuggingFace models. Not mounted for the Anthropic passthrough backend.
+- `x-anyllm-degradation` response header: set when features are silently dropped during translation (e.g., `top_k`, `thinking_config`, `cache_control`, `document_blocks`, `stop_sequences_truncated`)
 
 **Not fully validated:**
 - OpenAI Responses API backend: wired up via `OPENAI_API_FORMAT=responses` but not tested against live API
@@ -27,7 +29,8 @@ All 11 implementation phases are complete.
 
 ```bash
 cargo build                          # build everything
-cargo test                           # run all tests (~417 tests, 4 ignored)
+cargo test                           # run all tests (~480 tests, 4 ignored)
+cargo test -p anyllm_client     # client crate only
 cargo test -p anyllm_translate  # translator crate only
 cargo test -p anyllm_proxy      # proxy crate only
 cargo test health_endpoint            # single test by name
@@ -62,10 +65,22 @@ OPENAI_API_KEY=sk-... cargo run -p anyllm_proxy
 - `GEMINI_BASE_URL`: Gemini API base URL (default: `https://generativelanguage.googleapis.com/v1beta`)
 - `PROXY_API_KEYS`: Comma-separated list of allowed API keys for proxy authentication (optional; if unset, any non-empty key is accepted)
 - `LOG_BODIES`: Enable request/response body logging at debug level (`true` or `1`, default: disabled)
+- `OTEL_EXPORTER_OTLP_ENDPOINT`: OTLP collector endpoint (default: `http://localhost:4318`). Only effective when built with `--features otel`.
+- `OTEL_SERVICE_NAME`: Service name for exported traces. Only effective when built with `--features otel`.
+- `OTEL_TRACES_SAMPLER`: Sampling strategy (default: `parentbased_always_on`). Only effective when built with `--features otel`.
 
 ## Architecture
 
-Cargo workspace with two crates:
+Cargo workspace with three crates:
+
+### `crates/client` (lib: `anyllm_client`)
+High-level async HTTP client (Anthropic-in, Anthropic-out). Depends on `anyllm_translate` for translation logic. Key modules:
+- **`client.rs`**: `Client` struct with `ClientConfig` builder; `messages()` for non-streaming, streaming variant for SSE
+- **`http.rs`**: reqwest client builder with optional SSRF-safe DNS resolution and mTLS (PKCS#12)
+- **`retry.rs`**: Generic retry with exponential backoff + jitter; `is_retryable`, `send_with_retry`
+- **`rate_limit.rs`**: Parses `x-ratelimit-*` / `retry-after` headers into a typed struct
+- **`sse.rs`**: Framework-agnostic SSE frame parser (`find_double_newline`)
+- **`error.rs`**: `ClientError` enum
 
 ### `crates/translator` (lib: `anyllm_translate`)
 Pure translation logic, no IO. Key modules:
@@ -79,6 +94,7 @@ Pure translation logic, no IO. Key modules:
   - `streaming_map`: SSE event stream translation state machine
   - `responses_message_map`: Anthropic to/from OpenAI Responses API mapping
   - `responses_streaming_map`: Responses API SSE event stream translation state machine
+  - `warnings`: `TranslationWarnings` collector; lossy drops are surfaced via `x-anyllm-degradation` response header
 - **`middleware/`**: Request/response handler orchestrating translation and backend calls
 - **`util/`**: JSON helpers, ID generation (uuid v4), secret redaction
 - **`config.rs`**: Translator-level configuration, **`error.rs`**: Error types, **`translate.rs`**: Top-level translation entry points
@@ -126,8 +142,11 @@ Client (Anthropic format) -> proxy (axum)
 - Some source files reference PLAN.md line ranges in a comment at the top (historical; PLAN.md has been removed).
 - Test files live alongside source (`#[cfg(test)]` modules) and in `crates/proxy/tests/` for integration tests.
 - Error types use `thiserror` derive macros.
-- Test distribution: translator (~263 tests), proxy (~154 tests including integration/compatibility). Counts shift as features are added.
+- Test distribution: translator (~273 tests), proxy + client (~30 tests including integration/compatibility). Counts shift as features are added.
 
 ## References
 
 - OpenAI API spec: https://github.com/openai/openai-openapi/blob/manual_spec/openapi.yaml (very large, ~70k+ lines). See https://simonwillison.net/2024/Dec/22/openai-openapi/ for context on the spec's size and structure. Do not attempt to load the full spec into context; reference specific sections as needed.
+
+## Recent Changes
+- 20260325-120000-litellm-gap-fill: Added [if applicable, e.g., PostgreSQL, CoreData, files or N/A]
diff --git a/crates/proxy/Cargo.toml b/crates/proxy/Cargo.toml
index 6945f75..19d3c57 100644
--- a/crates/proxy/Cargo.toml
+++ b/crates/proxy/Cargo.toml
@@ -32,6 +32,8 @@ httpdate = "1"
 dashmap = "6"
 aws-sigv4 = { version = "1.4", features = ["sign-http"] }
 aws-credential-types = "1.2"
+aws-smithy-runtime-api = "1"
+base64 = "0.22"
 
 [features]
 otel = [
diff --git a/crates/proxy/admin-ui/index.html b/crates/proxy/admin-ui/index.html
index c6ee075..5911658 100644
--- a/crates/proxy/admin-ui/index.html
+++ b/crates/proxy/admin-ui/index.html
@@ -105,10 +105,11 @@ input:focus,select:focus{outline:none;border-color:#4a9eff}
   <div style="display:flex;gap:8px;margin:16px 0">
     <button class="btn btn-primary" id="btn-save-settings">Save Changes</button>
     <button class="btn btn-secondary" id="btn-reload-config">Reload</button>
+    <button class="btn btn-secondary" id="btn-export-env" title="Download .anyllm.env template">Export .env</button>
   </div>
-  <div class="section-label" style="margin-top:24px">Read-Only <span class="error">(requires restart)</span></div>
+  <div class="section-label" style="margin-top:24px">Environment <span class="error">(requires restart to change)</span></div>
   <div class="readonly-section" id="readonly-config">
-    <div style="color:#8b949e;font-size:12px">Configuration loaded from environment variables.</div>
+    <div style="color:#8b949e;font-size:12px">Loading...</div>
   </div>
 </div>
 
@@ -187,7 +188,7 @@ input:focus,select:focus{outline:none;border-color:#4a9eff}
       navEl.classList.add('active');
       document.getElementById('tab-' + navEl.dataset.tab).classList.add('active');
       if (navEl.dataset.tab === 'requests') loadRequests();
-      if (navEl.dataset.tab === 'settings') loadConfig();
+      if (navEl.dataset.tab === 'settings') { loadConfig(); loadEnv(); }
       if (navEl.dataset.tab === 'backends') loadBackends();
     });
   });
@@ -402,7 +403,95 @@ input:focus,select:focus{outline:none;border-color:#4a9eff}
   }
 
   document.getElementById('btn-save-settings').addEventListener('click', saveSettings);
-  document.getElementById('btn-reload-config').addEventListener('click', loadConfig);
+  document.getElementById('btn-reload-config').addEventListener('click', function() { loadConfig(); loadEnv(); });
+  document.getElementById('btn-export-env').addEventListener('click', exportEnvFile);
+
+  // -- Environment variables (read-only, requires restart to change) --
+  var ENV_GROUPS = [
+    { label: 'Core', keys: ['BACKEND','LISTEN_PORT','BIG_MODEL','SMALL_MODEL','RUST_LOG','LOG_BODIES','PROXY_CONFIG'] },
+    { label: 'OpenAI / Compatible', keys: ['OPENAI_BASE_URL','OPENAI_API_FORMAT','OPENAI_API_KEY'] },
+    { label: 'Vertex AI', keys: ['VERTEX_PROJECT','VERTEX_REGION','VERTEX_API_KEY'] },
+    { label: 'Gemini', keys: ['GEMINI_BASE_URL','GEMINI_API_KEY'] },
+    { label: 'Auth & TLS', keys: ['PROXY_API_KEYS','TLS_CLIENT_CERT_P12','TLS_CA_CERT'] },
+    { label: 'Admin', keys: ['ADMIN_PORT','ADMIN_DB_PATH','ADMIN_LOG_RETENTION_DAYS'] },
+  ];
+  var ENV_SECRET_KEYS = {'OPENAI_API_KEY':1,'VERTEX_API_KEY':1,'GEMINI_API_KEY':1,'PROXY_API_KEYS':1};
+
+  function loadEnv() {
+    apiFetch('/env').then(function(data) {
+      var container = document.getElementById('readonly-config');
+      clearChildren(container, false);
+      ENV_GROUPS.forEach(function(group) {
+        var hasAny = group.keys.some(function(k) { return data[k] != null; });
+        if (!hasAny) return;
+        var rows = [];
+        group.keys.forEach(function(k) {
+          var val = data[k];
+          if (val == null) return;
+          rows.push(el('div', {className: 'model-grid', style:{marginTop:'4px'}}, [
+            el('div', {className: 'label', textContent: k + ':'}),
+            el('div', {textContent: String(val), style:{fontFamily:'monospace',fontSize:'12px',color:'#c9d1d9'}})
+          ]));
+        });
+        if (rows.length === 0) return;
+        var header = el('div', {style:{color:'#8b949e',fontSize:'11px',textTransform:'uppercase',letterSpacing:'0.5px',marginTop:'10px',marginBottom:'4px'}, textContent: group.label});
+        container.appendChild(header);
+        rows.forEach(function(r) { container.appendChild(r); });
+      });
+      if (!container.children.length) {
+        container.appendChild(el('div', {style:{color:'#8b949e',fontSize:'12px'}, textContent: 'No environment variables set (using defaults).'}));
+      }
+    }).catch(function(e) {
+      var container = document.getElementById('readonly-config');
+      container.textContent = 'Failed to load environment variables.';
+      console.error('Env load failed:', e);
+    });
+  }
+
+  // -- Export .env file --
+  // Generates a .anyllm.env template the user can load with --env-file.
+  // Secret values (API keys) are masked server-side, so they appear as empty
+  // placeholders the user must fill in manually.
+  function exportEnvFile() {
+    apiFetch('/env').then(function(data) {
+      var lines = [
+        '# anyllm-proxy configuration',
+        '# Load with:  anyllm_proxy --env-file .anyllm.env',
+        '# Docker:     docker run --env-file .anyllm.env anyllm-proxy',
+        '#',
+        '# Secret values (API keys) are intentionally left blank for security.',
+        '# Fill them in before use.',
+        '',
+      ];
+
+      ENV_GROUPS.forEach(function(section) {
+        var hasAny = section.keys.some(function(k) { return data[k] != null || k in ENV_SECRET_KEYS; });
+        if (!hasAny) return;
+        lines.push('# --- ' + section.label + ' ---');
+        section.keys.forEach(function(k) {
+          if (k in ENV_SECRET_KEYS) {
+            // Always emit secret keys as empty placeholders.
+            lines.push(k + '=');
+          } else if (data[k] != null) {
+            lines.push(k + '=' + String(data[k]));
+          }
+          // Omit unset non-secret keys entirely to keep the file minimal.
+        });
+        lines.push('');
+      });
+
+      var blob = new Blob([lines.join('\n')], {type: 'text/plain'});
+      var a = document.createElement('a');
+      a.href = URL.createObjectURL(blob);
+      a.download = '.anyllm.env';
+      document.body.appendChild(a);
+      a.click();
+      document.body.removeChild(a);
+      URL.revokeObjectURL(a.href);
+    }).catch(function(e) {
+      alert('Failed to fetch env: ' + e.message);
+    });
+  }
 
   // -- Backends detail --
   function loadBackends() {
diff --git a/crates/proxy/src/backend/bedrock_client.rs b/crates/proxy/src/backend/bedrock_client.rs
new file mode 100644
index 0000000..d2dbe08
--- /dev/null
+++ b/crates/proxy/src/backend/bedrock_client.rs
@@ -0,0 +1,399 @@
+// AWS Bedrock client with SigV4 request signing.
+// Sends Anthropic Messages API requests directly to Bedrock (no OpenAI translation).
+// Bedrock streaming uses AWS Event Stream binary framing, not SSE.
+
+use super::{build_http_client, RateLimitHeaders};
+use crate::config::TlsConfig;
+use aws_credential_types::Credentials;
+use aws_sigv4::http_request::{
+    sign, SignableBody, SignableRequest, SigningSettings,
+};
+use aws_sigv4::sign::v4;
+use reqwest::Client;
+use tokio::time::sleep;
+
+/// HTTP client for AWS Bedrock with SigV4 request signing.
+#[derive(Clone)]
+pub struct BedrockClient {
+    client: Client,
+    region: String,
+    credentials: Credentials,
+    big_model: String,
+    small_model: String,
+}
+
+/// Error type for the Bedrock client.
+#[derive(Debug)]
+pub enum BedrockClientError {
+    /// Transport-level error (connection, timeout, DNS).
+    Transport(String),
+    /// Upstream returned a non-success status. Body is raw bytes for passthrough.
+    ApiError { status: u16, body: bytes::Bytes },
+    /// SigV4 signing failed.
+    Signing(String),
+}
+
+impl std::fmt::Display for BedrockClientError {
+    fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
+        match self {
+            Self::Transport(msg) => write!(f, "Bedrock transport error: {msg}"),
+            Self::ApiError { status, .. } => write!(f, "Bedrock API error (status {status})"),
+            Self::Signing(msg) => write!(f, "Bedrock signing error: {msg}"),
+        }
+    }
+}
+
+impl BedrockClient {
+    /// Create a new Bedrock client.
+    pub fn new(
+        region: String,
+        credentials: Credentials,
+        big_model: String,
+        small_model: String,
+        tls: &TlsConfig,
+    ) -> Self {
+        let client = build_http_client(tls);
+        Self {
+            client,
+            region,
+            credentials,
+            big_model,
+            small_model,
+        }
+    }
+
+    pub fn big_model(&self) -> &str {
+        &self.big_model
+    }
+
+    pub fn small_model(&self) -> &str {
+        &self.small_model
+    }
+
+    /// Build the Bedrock InvokeModel URL for a given model.
+    fn invoke_url(&self, model_id: &str) -> String {
+        format!(
+            "https://bedrock-runtime.{}.amazonaws.com/model/{}/invoke",
+            self.region, model_id
+        )
+    }
+
+    /// Build the Bedrock InvokeModelWithResponseStream URL.
+    fn invoke_stream_url(&self, model_id: &str) -> String {
+        format!(
+            "https://bedrock-runtime.{}.amazonaws.com/model/{}/invoke-with-response-stream",
+            self.region, model_id
+        )
+    }
+
+    /// Sign an HTTP request with SigV4 and return headers to add.
+    fn sign_request(
+        &self,
+        method: &str,
+        url: &str,
+        body_bytes: &[u8],
+        extra_headers: &[(&str, &str)],
+    ) -> Result<Vec<(String, String)>, BedrockClientError> {
+        let identity: aws_smithy_runtime_api::client::identity::Identity =
+            self.credentials.clone().into();
+        let settings = SigningSettings::default();
+        let params = v4::SigningParams::builder()
+            .identity(&identity)
+            .region(&self.region)
+            .name("bedrock")
+            .time(std::time::SystemTime::now())
+            .settings(settings)
+            .build()
+            .map_err(|e| BedrockClientError::Signing(e.to_string()))?;
+        let signing_params = params.into();
+
+        let signable = SignableRequest::new(
+            method,
+            url,
+            extra_headers.iter().copied(),
+            SignableBody::Bytes(body_bytes),
+        )
+        .map_err(|e| BedrockClientError::Signing(e.to_string()))?;
+
+        let (instructions, _signature) = sign(signable, &signing_params)
+            .map_err(|e| BedrockClientError::Signing(e.to_string()))?
+            .into_parts();
+
+        // Collect signing headers
+        let headers: Vec<(String, String)> = instructions
+            .headers()
+            .map(|(k, v)| (k.to_string(), v.to_string()))
+            .collect();
+        Ok(headers)
+    }
+
+    /// Forward a non-streaming request. Returns raw response body and rate limit headers.
+    pub async fn forward(
+        &self,
+        body: bytes::Bytes,
+        model_id: &str,
+    ) -> Result<(bytes::Bytes, RateLimitHeaders), BedrockClientError> {
+        let response = self.send_with_retry(body, model_id, false).await?;
+        let rate_limits = RateLimitHeaders::default();
+        let resp_body = response
+            .bytes()
+            .await
+            .map_err(|e| BedrockClientError::Transport(e.to_string()))?;
+        Ok((resp_body, rate_limits))
+    }
+
+    /// Forward a streaming request. Returns the raw response for event stream decoding.
+    pub async fn forward_stream(
+        &self,
+        body: bytes::Bytes,
+        model_id: &str,
+    ) -> Result<(reqwest::Response, RateLimitHeaders), BedrockClientError> {
+        let response = self.send_with_retry(body, model_id, true).await?;
+        let rate_limits = RateLimitHeaders::default();
+        Ok((response, rate_limits))
+    }
+
+    /// Send with retry on 429/5xx.
+    async fn send_with_retry(
+        &self,
+        body: bytes::Bytes,
+        model_id: &str,
+        stream: bool,
+    ) -> Result<reqwest::Response, BedrockClientError> {
+        let url = if stream {
+            self.invoke_stream_url(model_id)
+        } else {
+            self.invoke_url(model_id)
+        };
+
+        let content_type = "application/json";
+        let accept = if stream {
+            "application/vnd.amazon.eventstream"
+        } else {
+            "application/json"
+        };
+
+        for attempt in 0..=super::MAX_RETRIES {
+            let base_headers = [
+                ("content-type", content_type),
+                ("accept", accept),
+            ];
+            let signing_headers = self
+                .sign_request("POST", &url, &body, &base_headers)?;
+
+            let mut rb = self
+                .client
+                .post(&url)
+                .header("content-type", content_type)
+                .header("accept", accept)
+                .body(body.clone());
+
+            for (k, v) in &signing_headers {
+                rb = rb.header(k.as_str(), v.as_str());
+            }
+
+            let response = rb
+                .send()
+                .await
+                .map_err(|e| BedrockClientError::Transport(e.to_string()))?;
+            let status = response.status().as_u16();
+
+            if (200..300).contains(&status) {
+                return Ok(response);
+            }
+
+            if attempt < super::MAX_RETRIES && super::is_retryable(status) {
+                let retry_after = super::parse_retry_after(response.headers());
+                let delay = super::backoff_delay(attempt, retry_after);
+                tracing::warn!(
+                    status,
+                    attempt = attempt + 1,
+                    max_retries = super::MAX_RETRIES,
+                    delay_ms = delay.as_millis() as u64,
+                    "retryable error from Bedrock, backing off"
+                );
+                drop(response.bytes().await);
+                sleep(delay).await;
+                continue;
+            }
+
+            let resp_body = response.bytes().await.unwrap_or_default();
+            return Err(BedrockClientError::ApiError {
+                status,
+                body: resp_body,
+            });
+        }
+        unreachable!("loop runs MAX_RETRIES+1 times and always returns")
+    }
+}
+
+// ---------------------------------------------------------------------------
+// AWS Event Stream binary frame decoder
+// ---------------------------------------------------------------------------
+
+/// Decode AWS Event Stream frames from a byte buffer.
+/// Each frame: 4-byte total_len | 4-byte headers_len | 4-byte prelude CRC |
+///             headers | payload | 4-byte message CRC
+///
+/// The payload contains `{"bytes":"<base64>"}` where base64 decodes to an
+/// Anthropic SSE JSON event string.
+pub mod eventstream {
+    use bytes::BytesMut;
+
+    /// Minimum frame size: 4 (total_len) + 4 (headers_len) + 4 (prelude CRC)
+    ///                     + 0 (headers) + 0 (payload) + 4 (message CRC) = 16
+    const MIN_FRAME_SIZE: usize = 16;
+
+    /// Try to extract one complete event stream frame from the buffer.
+    /// Returns `Some(payload_bytes)` and advances the buffer past the frame,
+    /// or `None` if the buffer does not contain a complete frame yet.
+    pub fn decode_frame(buf: &mut BytesMut) -> Option<Vec<u8>> {
+        if buf.len() < MIN_FRAME_SIZE {
+            return None;
+        }
+
+        let total_len = u32::from_be_bytes([buf[0], buf[1], buf[2], buf[3]]) as usize;
+        if buf.len() < total_len {
+            return None; // incomplete frame
+        }
+
+        let headers_len = u32::from_be_bytes([buf[4], buf[5], buf[6], buf[7]]) as usize;
+
+        // Prelude is 8 bytes (total_len + headers_len), then 4-byte prelude CRC
+        let headers_start = 12; // 4 + 4 + 4 (prelude CRC)
+        let payload_start = headers_start + headers_len;
+        // Message CRC is the last 4 bytes
+        let payload_end = total_len.saturating_sub(4);
+
+        if payload_start > payload_end || payload_end > total_len {
+            // Malformed frame: skip it
+            let _ = buf.split_to(total_len);
+            return Some(Vec::new());
+        }
+
+        let payload = buf[payload_start..payload_end].to_vec();
+
+        // Advance buffer past this frame
+        let _ = buf.split_to(total_len);
+
+        Some(payload)
+    }
+
+    /// Extract the Anthropic event JSON string from a Bedrock event stream payload.
+    /// Bedrock wraps the Anthropic event in `{"bytes":"<base64>"}`.
+    /// Returns None if the payload is not a chunk event or is malformed.
+    pub fn extract_event_from_payload(payload: &[u8]) -> Option<String> {
+        if payload.is_empty() {
+            return None;
+        }
+
+        // Parse as JSON to extract the base64-encoded bytes field
+        let parsed: serde_json::Value = serde_json::from_slice(payload).ok()?;
+        let b64 = parsed.get("bytes")?.as_str()?;
+
+        // Base64 decode
+        use base64::Engine;
+        let decoded = base64::engine::general_purpose::STANDARD
+            .decode(b64)
+            .ok()?;
+        String::from_utf8(decoded).ok()
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use super::eventstream;
+    use bytes::BytesMut;
+
+    /// Build a minimal AWS Event Stream frame with the given payload.
+    /// Uses zero CRCs (we don't validate CRCs in the decoder).
+    fn build_frame(headers: &[u8], payload: &[u8]) -> Vec<u8> {
+        let total_len = 12 + headers.len() + payload.len() + 4; // prelude(12) + headers + payload + msg CRC(4)
+        let headers_len = headers.len();
+        let mut frame = Vec::with_capacity(total_len);
+        frame.extend_from_slice(&(total_len as u32).to_be_bytes());
+        frame.extend_from_slice(&(headers_len as u32).to_be_bytes());
+        frame.extend_from_slice(&[0u8; 4]); // prelude CRC (not validated)
+        frame.extend_from_slice(headers);
+        frame.extend_from_slice(payload);
+        frame.extend_from_slice(&[0u8; 4]); // message CRC (not validated)
+        frame
+    }
+
+    #[test]
+    fn decode_frame_empty_payload() {
+        let frame = build_frame(&[], &[]);
+        let mut buf = BytesMut::from(frame.as_slice());
+        let payload = eventstream::decode_frame(&mut buf).unwrap();
+        assert!(payload.is_empty());
+        assert!(buf.is_empty());
+    }
+
+    #[test]
+    fn decode_frame_with_payload() {
+        let payload_data = b"hello world";
+        let frame = build_frame(&[], payload_data);
+        let mut buf = BytesMut::from(frame.as_slice());
+        let payload = eventstream::decode_frame(&mut buf).unwrap();
+        assert_eq!(payload, b"hello world");
+        assert!(buf.is_empty());
+    }
+
+    #[test]
+    fn decode_frame_incomplete() {
+        let frame = build_frame(&[], b"hello");
+        let mut buf = BytesMut::from(&frame[..frame.len() - 2]); // truncate
+        assert!(eventstream::decode_frame(&mut buf).is_none());
+    }
+
+    #[test]
+    fn decode_multiple_frames() {
+        let frame1 = build_frame(&[], b"first");
+        let frame2 = build_frame(&[], b"second");
+        let mut buf = BytesMut::new();
+        buf.extend_from_slice(&frame1);
+        buf.extend_from_slice(&frame2);
+
+        let p1 = eventstream::decode_frame(&mut buf).unwrap();
+        assert_eq!(p1, b"first");
+        let p2 = eventstream::decode_frame(&mut buf).unwrap();
+        assert_eq!(p2, b"second");
+        assert!(buf.is_empty());
+    }
+
+    #[test]
+    fn decode_frame_with_headers() {
+        let headers = b"\x00\x04test";
+        let payload_data = b"data";
+        let frame = build_frame(headers, payload_data);
+        let mut buf = BytesMut::from(frame.as_slice());
+        let payload = eventstream::decode_frame(&mut buf).unwrap();
+        assert_eq!(payload, b"data");
+    }
+
+    #[test]
+    fn extract_event_from_valid_payload() {
+        use base64::Engine;
+        let event_json = r#"{"type":"content_block_delta","index":0}"#;
+        let b64 = base64::engine::general_purpose::STANDARD.encode(event_json);
+        let wrapper = format!(r#"{{"bytes":"{b64}"}}"#);
+        let result = eventstream::extract_event_from_payload(wrapper.as_bytes());
+        assert_eq!(result.unwrap(), event_json);
+    }
+
+    #[test]
+    fn extract_event_empty_payload() {
+        assert!(eventstream::extract_event_from_payload(&[]).is_none());
+    }
+
+    #[test]
+    fn extract_event_invalid_json() {
+        assert!(eventstream::extract_event_from_payload(b"not json").is_none());
+    }
+
+    #[test]
+    fn extract_event_missing_bytes_field() {
+        let payload = r#"{"other":"field"}"#;
+        assert!(eventstream::extract_event_from_payload(payload.as_bytes()).is_none());
+    }
+}
diff --git a/crates/proxy/src/backend/mod.rs b/crates/proxy/src/backend/mod.rs
index d5ceb09..3e2db8d 100644
--- a/crates/proxy/src/backend/mod.rs
+++ b/crates/proxy/src/backend/mod.rs
@@ -1,10 +1,13 @@
 /// Passthrough client forwarding Anthropic requests as-is to upstream Anthropic API.
 pub mod anthropic_client;
+/// AWS Bedrock client with SigV4 request signing.
+pub mod bedrock_client;
 /// reqwest client for OpenAI-compatible Chat Completions and Responses APIs with retry/backoff.
 pub mod openai_client;
 
 use crate::config::{BackendAuth, BackendConfig, BackendKind, Config, OpenAIApiFormat, TlsConfig};
 use anthropic_client::{AnthropicClient, AnthropicClientError};
+use bedrock_client::{BedrockClient, BedrockClientError};
 use openai_client::{OpenAIClient, OpenAIClientError};
 
 // Re-export from the client crate so existing code paths (streaming, routes, etc.) keep working.
@@ -66,6 +69,8 @@ pub enum BackendClient {
     GeminiOpenAI(OpenAIClient),
     /// Passthrough to real Anthropic API (no translation).
     Anthropic(AnthropicClient),
+    /// AWS Bedrock: sends Anthropic-format requests with SigV4 signing.
+    Bedrock(BedrockClient),
 }
 
 /// Unified error type for all backend clients.
@@ -73,6 +78,7 @@ pub enum BackendClient {
 pub enum BackendError {
     OpenAI(OpenAIClientError),
     Anthropic(AnthropicClientError),
+    Bedrock(BedrockClientError),
 }
 
 impl BackendError {
@@ -80,6 +86,7 @@ impl BackendError {
     pub fn api_error_status(&self) -> Option<u16> {
         match self {
             Self::OpenAI(OpenAIClientError::ApiError { status, .. }) => Some(*status),
+            Self::Bedrock(BedrockClientError::ApiError { status, .. }) => Some(*status),
             _ => None,
         }
     }
@@ -94,6 +101,7 @@ impl BackendError {
         match self {
             Self::OpenAI(e) => e.to_string(),
             Self::Anthropic(e) => e.to_string(),
+            Self::Bedrock(e) => e.to_string(),
         }
     }
 
@@ -114,6 +122,7 @@ impl std::fmt::Display for BackendError {
         match self {
             Self::OpenAI(e) => write!(f, "{e}"),
             Self::Anthropic(e) => write!(f, "{e}"),
+            Self::Bedrock(e) => write!(f, "{e}"),
         }
     }
 }
@@ -130,6 +139,12 @@ impl From<AnthropicClientError> for BackendError {
     }
 }
 
+impl From<BedrockClientError> for BackendError {
+    fn from(e: BedrockClientError) -> Self {
+        Self::Bedrock(e)
+    }
+}
+
 impl BackendClient {
     /// Forward a raw embeddings request to the backend. No translation — model names pass through.
     /// Returns `501 Not Implemented` for the Anthropic backend (no embeddings endpoint).
@@ -147,11 +162,11 @@ impl BackendClient {
                 .embeddings_passthrough(body, content_type)
                 .await
                 .map_err(BackendError::OpenAI),
-            Self::Anthropic(_) => {
-                // Anthropic has no embeddings API.
+            Self::Anthropic(_) | Self::Bedrock(_) => {
+                // Anthropic and Bedrock have no embeddings API.
                 let err = anyllm_translate::mapping::errors_map::create_anthropic_error(
                     anyllm_translate::anthropic::ErrorType::InvalidRequestError,
-                    "Embeddings are not supported by the Anthropic backend.".to_string(),
+                    "Embeddings are not supported by this backend.".to_string(),
                     None,
                 );
                 let body = serde_json::to_vec(&err).unwrap_or_default();
@@ -182,6 +197,11 @@ impl BackendClient {
                 &config.openai_api_key,
                 &config.tls,
             )),
+            BackendKind::Bedrock => {
+                // Bedrock config is stored in openai_base_url (region) and openai_api_key (unused).
+                // Credentials come from env vars at Config::from_env time.
+                unreachable!("Bedrock backend uses from_backend_config, not Config::new")
+            }
         }
     }
 
@@ -210,6 +230,15 @@ impl BackendClient {
             BackendKind::Vertex => Self::Vertex(OpenAIClient::new(&legacy)),
             BackendKind::Gemini => Self::GeminiOpenAI(OpenAIClient::new(&legacy)),
             BackendKind::Anthropic => Self::Anthropic(AnthropicClient::from_backend_config(bc)),
+            BackendKind::Bedrock => Self::Bedrock(BedrockClient::new(
+                bc.base_url.clone(), // region is stored in base_url for Bedrock
+                bc.bedrock_credentials
+                    .clone()
+                    .expect("Bedrock credentials must be set"),
+                bc.model_mapping.big_model.clone(),
+                bc.model_mapping.small_model.clone(),
+                &bc.tls,
+            )),
         }
     }
 }
diff --git a/crates/proxy/src/backend/openai_client.rs b/crates/proxy/src/backend/openai_client.rs
index 1d85790..ece1aa6 100644
--- a/crates/proxy/src/backend/openai_client.rs
+++ b/crates/proxy/src/backend/openai_client.rs
@@ -74,8 +74,8 @@ impl OpenAIClient {
                     format!("{endpoint}/openai/deployments/{deployment}/embeddings?api-version={api_version}"),
                 )
             }
-            BackendKind::Anthropic => {
-                unreachable!("OpenAIClient should not be constructed for Anthropic backend")
+            BackendKind::Anthropic | BackendKind::Bedrock => {
+                unreachable!("OpenAIClient should not be constructed for Anthropic/Bedrock backend")
             }
         };
 
diff --git a/crates/proxy/src/config/mod.rs b/crates/proxy/src/config/mod.rs
index 5ea48ed..f68742e 100644
--- a/crates/proxy/src/config/mod.rs
+++ b/crates/proxy/src/config/mod.rs
@@ -19,6 +19,7 @@ pub enum BackendKind {
     Vertex,
     Gemini,
     Anthropic,
+    Bedrock,
 }
 
 /// Which OpenAI API format to use (only relevant when BACKEND=openai).
@@ -92,8 +93,9 @@ impl Config {
             "vertex" => BackendKind::Vertex,
             "gemini" => BackendKind::Gemini,
             "anthropic" => BackendKind::Anthropic,
+            "bedrock" => BackendKind::Bedrock,
             other => {
-                panic!("unknown BACKEND value '{other}', expected 'openai', 'azure', 'vertex', 'gemini', or 'anthropic'")
+                panic!("unknown BACKEND value '{other}', expected 'openai', 'azure', 'vertex', 'gemini', 'anthropic', or 'bedrock'")
             }
         };
 
@@ -267,6 +269,37 @@ impl Config {
                     openai_api_format: OpenAIApiFormat::Chat,
                 }
             }
+            BackendKind::Bedrock => {
+                let region = std::env::var("AWS_REGION")
+                    .unwrap_or_else(|_| panic!("AWS_REGION is required when BACKEND=bedrock"));
+                validate_gcp_identifier("AWS_REGION", &region); // reuse safe-char validation
+
+                // Validate credentials are present at startup; the actual values
+                // are read again when constructing BedrockClient.
+                let _access_key_id = std::env::var("AWS_ACCESS_KEY_ID").unwrap_or_else(|_| {
+                    panic!("AWS_ACCESS_KEY_ID is required when BACKEND=bedrock")
+                });
+                let _secret_access_key = std::env::var("AWS_SECRET_ACCESS_KEY").unwrap_or_else(
+                    |_| panic!("AWS_SECRET_ACCESS_KEY is required when BACKEND=bedrock"),
+                );
+                let _session_token = std::env::var("AWS_SESSION_TOKEN").ok();
+
+                Self {
+                    backend,
+                    openai_api_key: String::new(),
+                    // Store region in openai_base_url for wrap_config
+                    openai_base_url: region.clone(),
+                    listen_port,
+                    model_mapping: ModelMapping::from_env_with_defaults(
+                        "anthropic.claude-sonnet-4-20250514-v1:0",
+                        "anthropic.claude-haiku-4-5-20251001-v1:0",
+                    ),
+                    tls,
+                    backend_auth: BackendAuth::BearerToken(String::new()),
+                    log_bodies,
+                    openai_api_format: OpenAIApiFormat::Chat,
+                }
+            }
         }
     }
 }
@@ -318,6 +351,22 @@ fn contains_ignore_ascii_case(haystack: &[u8], needle: &[u8]) -> bool {
         .any(|w| w.eq_ignore_ascii_case(needle))
 }
 
+/// Read AWS credentials from environment variables for the Bedrock backend.
+fn bedrock_credentials_from_env() -> aws_credential_types::Credentials {
+    let access_key_id = std::env::var("AWS_ACCESS_KEY_ID")
+        .unwrap_or_else(|_| panic!("AWS_ACCESS_KEY_ID is required for bedrock"));
+    let secret_access_key = std::env::var("AWS_SECRET_ACCESS_KEY")
+        .unwrap_or_else(|_| panic!("AWS_SECRET_ACCESS_KEY is required for bedrock"));
+    let session_token = std::env::var("AWS_SESSION_TOKEN").ok();
+    aws_credential_types::Credentials::new(
+        access_key_id,
+        secret_access_key,
+        session_token,
+        None,
+        "env",
+    )
+}
+
 // ---------------------------------------------------------------------------
 // Multi-backend configuration
 // ---------------------------------------------------------------------------
@@ -357,6 +406,8 @@ pub struct BackendConfig {
     /// (older Ollama, text-generation-webui, LM Studio) that reject unknown
     /// fields with HTTP 400.
     pub omit_stream_options: bool,
+    /// AWS credentials for Bedrock backend. None for all other backends.
+    pub bedrock_credentials: Option<aws_credential_types::Credentials>,
 }
 
 /// Top-level multi-backend configuration loaded from TOML.
@@ -403,6 +454,10 @@ struct TomlBackendConfig {
     access_token: Option<String>,
     // Strip stream_options from streaming requests (local LLM compat)
     omit_stream_options: Option<bool>,
+    // Bedrock-specific: AWS credentials (support env: prefix for env var resolution)
+    aws_access_key_id: Option<String>,
+    aws_secret_access_key: Option<String>,
+    aws_session_token: Option<String>,
 }
 
 impl MultiConfig {
@@ -435,12 +490,20 @@ impl MultiConfig {
             BackendKind::Vertex => "vertex",
             BackendKind::Gemini => "gemini",
             BackendKind::Anthropic => "anthropic",
+            BackendKind::Bedrock => "bedrock",
         };
 
         let omit_stream_options = std::env::var("OMIT_STREAM_OPTIONS")
             .map(|v| v == "true" || v == "1")
             .unwrap_or(false);
 
+        // For Bedrock, read AWS credentials from env vars.
+        let bedrock_credentials = if config.backend == BackendKind::Bedrock {
+            Some(bedrock_credentials_from_env())
+        } else {
+            None
+        };
+
         let bc = BackendConfig {
             kind: config.backend.clone(),
             api_key: config.openai_api_key.clone(),
@@ -451,6 +514,7 @@ impl MultiConfig {
             backend_auth: config.backend_auth.clone(),
             log_bodies: config.log_bodies,
             omit_stream_options,
+            bedrock_credentials,
         };
 
         let mut backends = IndexMap::new();
@@ -521,6 +585,7 @@ impl MultiConfig {
             "vertex" => BackendKind::Vertex,
             "gemini" => BackendKind::Gemini,
             "anthropic" => BackendKind::Anthropic,
+            "bedrock" => BackendKind::Bedrock,
             other => panic!("unknown backend kind '{other}' for backend '{name}'"),
         };
 
@@ -684,6 +749,62 @@ impl MultiConfig {
                 };
                 (base_url, auth, mm, OpenAIApiFormat::Chat)
             }
+            BackendKind::Bedrock => {
+                let region = tb.region.as_deref().unwrap_or_else(|| {
+                    panic!("backend '{name}': 'region' is required for bedrock")
+                });
+                validate_gcp_identifier("region", region);
+
+                // For Bedrock, base_url stores the region (used by BedrockClient to build URLs)
+                let auth = BackendAuth::BearerToken(String::new());
+                let mm = ModelMapping {
+                    big_model: tb.big_model.clone().unwrap_or_else(|| {
+                        "anthropic.claude-sonnet-4-20250514-v1:0".to_string()
+                    }),
+                    small_model: tb.small_model.clone().unwrap_or_else(|| {
+                        "anthropic.claude-haiku-4-5-20251001-v1:0".to_string()
+                    }),
+                };
+                (region.to_string(), auth, mm, OpenAIApiFormat::Chat)
+            }
+        };
+
+        // Build AWS credentials for Bedrock from TOML fields or env vars.
+        let bedrock_credentials = if kind == BackendKind::Bedrock {
+            let access_key_id = tb
+                .aws_access_key_id
+                .as_deref()
+                .map(|v| resolve_env_value(v).unwrap_or_else(|e| panic!("backend '{name}': {e}")))
+                .unwrap_or_else(|| {
+                    std::env::var("AWS_ACCESS_KEY_ID").unwrap_or_else(|_| {
+                        panic!("backend '{name}': aws_access_key_id or AWS_ACCESS_KEY_ID required")
+                    })
+                });
+            let secret_access_key = tb
+                .aws_secret_access_key
+                .as_deref()
+                .map(|v| resolve_env_value(v).unwrap_or_else(|e| panic!("backend '{name}': {e}")))
+                .unwrap_or_else(|| {
+                    std::env::var("AWS_SECRET_ACCESS_KEY").unwrap_or_else(|_| {
+                        panic!(
+                            "backend '{name}': aws_secret_access_key or AWS_SECRET_ACCESS_KEY required"
+                        )
+                    })
+                });
+            let session_token = tb
+                .aws_session_token
+                .as_deref()
+                .map(|v| resolve_env_value(v).unwrap_or_else(|e| panic!("backend '{name}': {e}")))
+                .or_else(|| std::env::var("AWS_SESSION_TOKEN").ok());
+            Some(aws_credential_types::Credentials::new(
+                access_key_id,
+                secret_access_key,
+                session_token,
+                None,
+                "toml-config",
+            ))
+        } else {
+            None
         };
 
         BackendConfig {
@@ -696,6 +817,7 @@ impl MultiConfig {
             backend_auth,
             log_bodies,
             omit_stream_options: tb.omit_stream_options.unwrap_or(false),
+            bedrock_credentials,
         }
     }
 }
diff --git a/crates/proxy/src/lib.rs b/crates/proxy/src/lib.rs
index a98de8f..ee82649 100644
--- a/crates/proxy/src/lib.rs
+++ b/crates/proxy/src/lib.rs
@@ -6,5 +6,8 @@ pub mod backend;
 pub mod config;
 /// Request count, success/error tracking, exposed via GET /metrics.
 pub mod metrics;
+/// Optional OpenTelemetry OTLP trace export (requires `otel` feature).
+#[cfg(feature = "otel")]
+pub mod otel;
 /// Axum HTTP server: routes, middleware (auth, request ID, size/concurrency limits), SSE streaming.
 pub mod server;
diff --git a/crates/proxy/src/main.rs b/crates/proxy/src/main.rs
index 9dd439e..f330b40 100644
--- a/crates/proxy/src/main.rs
+++ b/crates/proxy/src/main.rs
@@ -25,6 +25,22 @@ async fn main() {
     // Use a reload layer so the admin API can change log_level at runtime.
     let env_filter = tracing_subscriber::EnvFilter::from_default_env();
     let (filter, reload_handle) = tracing_subscriber::reload::Layer::new(env_filter);
+
+    // When the `otel` feature is enabled, wire an OpenTelemetry tracing layer
+    // into the subscriber so that spans are exported as OTLP traces.
+    #[cfg(feature = "otel")]
+    let _otel_guard = {
+        let (guard, tracer) = anyllm_proxy::otel::init_otel();
+        let otel_layer = tracing_opentelemetry::OpenTelemetryLayer::new(tracer);
+        tracing_subscriber::registry()
+            .with(filter)
+            .with(tracing_subscriber::fmt::layer().json())
+            .with(otel_layer)
+            .init();
+        guard
+    };
+
+    #[cfg(not(feature = "otel"))]
     tracing_subscriber::registry()
         .with(filter)
         .with(tracing_subscriber::fmt::layer().json())
diff --git a/crates/proxy/src/otel.rs b/crates/proxy/src/otel.rs
new file mode 100644
index 0000000..4e3d518
--- /dev/null
+++ b/crates/proxy/src/otel.rs
@@ -0,0 +1,51 @@
+//! Optional OpenTelemetry OTLP trace export.
+//!
+//! Enabled by the `otel` cargo feature. All types and functions in this module
+//! are gated behind `#[cfg(feature = "otel")]` at the module declaration site
+//! in `lib.rs`, so nothing here compiles into the default binary.
+//!
+//! The OTLP SDK reads configuration from standard env vars:
+//! - `OTEL_EXPORTER_OTLP_ENDPOINT` (default `http://localhost:4318` for HTTP)
+//! - `OTEL_SERVICE_NAME`
+//! - `OTEL_TRACES_SAMPLER` / `OTEL_TRACES_SAMPLER_ARG`
+
+use opentelemetry::trace::TracerProvider as _;
+use opentelemetry_otlp::SpanExporter;
+use opentelemetry_sdk::trace::SdkTracerProvider;
+
+/// Holds the [`SdkTracerProvider`] and flushes pending spans on drop.
+pub struct OtelGuard {
+    provider: SdkTracerProvider,
+}
+
+impl Drop for OtelGuard {
+    fn drop(&mut self) {
+        if let Err(e) = self.provider.shutdown() {
+            eprintln!("otel: tracer provider shutdown error: {e}");
+        }
+    }
+}
+
+/// Initialise the OTLP span exporter and return a guard that must live for the
+/// duration of `main`. The returned tracer is suitable for
+/// [`tracing_opentelemetry::OpenTelemetryLayer::new`].
+///
+/// Panics if the exporter or provider cannot be created (misconfiguration).
+pub fn init_otel() -> (OtelGuard, opentelemetry_sdk::trace::Tracer) {
+    // HTTP/protobuf transport via reqwest (matches Cargo feature flags).
+    let exporter = SpanExporter::builder()
+        .with_http()
+        .build()
+        .expect("failed to create OTLP span exporter");
+
+    let provider = SdkTracerProvider::builder()
+        .with_batch_exporter(exporter)
+        .build();
+
+    opentelemetry::global::set_tracer_provider(provider.clone());
+
+    let tracer = provider.tracer("anyllm-proxy");
+    let guard = OtelGuard { provider };
+
+    (guard, tracer)
+}
diff --git a/crates/proxy/src/server/bedrock_passthrough.rs b/crates/proxy/src/server/bedrock_passthrough.rs
new file mode 100644
index 0000000..8a584b7
--- /dev/null
+++ b/crates/proxy/src/server/bedrock_passthrough.rs
@@ -0,0 +1,240 @@
+// Bedrock passthrough handler: forwards Anthropic-format requests to AWS Bedrock
+// with SigV4 signing and AWS Event Stream decoding for streaming.
+
+use crate::backend::bedrock_client::{eventstream, BedrockClientError};
+use crate::backend::BackendClient;
+use anyllm_translate::{anthropic, mapping};
+use axum::{
+    body::Bytes,
+    extract::State,
+    http::StatusCode,
+    response::{IntoResponse, Json, Response},
+};
+use bytes::BytesMut;
+use futures::StreamExt;
+
+use super::routes::AppState;
+
+/// Bedrock passthrough handler for POST /v1/messages.
+/// Strips the `model` field from the body (Bedrock uses it in the URL),
+/// adds `anthropic_version`, and handles AWS Event Stream binary framing
+/// for streaming responses.
+pub(crate) async fn bedrock_passthrough(State(state): State<AppState>, body: Bytes) -> Response {
+    state.metrics.record_request();
+
+    let client = match &state.backend {
+        BackendClient::Bedrock(c) => c.clone(),
+        _ => {
+            let err = mapping::errors_map::create_anthropic_error(
+                anthropic::ErrorType::ApiError,
+                "Backend is not configured as bedrock".to_string(),
+                None,
+            );
+            return (StatusCode::INTERNAL_SERVER_ERROR, Json(err)).into_response();
+        }
+    };
+
+    // Parse the body to extract model and stream fields, then rebuild for Bedrock.
+    let mut parsed: serde_json::Value = match serde_json::from_slice(&body) {
+        Ok(v) => v,
+        Err(e) => {
+            let err = mapping::errors_map::create_anthropic_error(
+                anthropic::ErrorType::InvalidRequestError,
+                format!("invalid JSON: {e}"),
+                None,
+            );
+            return (StatusCode::BAD_REQUEST, Json(err)).into_response();
+        }
+    };
+
+    // Extract model for the URL
+    let model_id = parsed
+        .get("model")
+        .and_then(|v| v.as_str())
+        .unwrap_or("")
+        .to_string();
+    if model_id.is_empty() {
+        let err = mapping::errors_map::create_anthropic_error(
+            anthropic::ErrorType::InvalidRequestError,
+            "model is required".to_string(),
+            None,
+        );
+        return (StatusCode::BAD_REQUEST, Json(err)).into_response();
+    }
+
+    // Map model name through runtime config
+    let mapped_model = state.map_model(&model_id);
+
+    let is_stream = parsed
+        .get("stream")
+        .and_then(|v| v.as_bool())
+        .unwrap_or(false);
+
+    // Bedrock: model goes in URL, not body. Add anthropic_version.
+    if let Some(obj) = parsed.as_object_mut() {
+        obj.remove("model");
+        obj.insert(
+            "anthropic_version".to_string(),
+            serde_json::Value::String("bedrock-2023-05-31".to_string()),
+        );
+    }
+
+    let bedrock_body = match serde_json::to_vec(&parsed) {
+        Ok(b) => bytes::Bytes::from(b),
+        Err(e) => {
+            let err = mapping::errors_map::create_anthropic_error(
+                anthropic::ErrorType::ApiError,
+                format!("failed to serialize request: {e}"),
+                None,
+            );
+            return (StatusCode::INTERNAL_SERVER_ERROR, Json(err)).into_response();
+        }
+    };
+
+    if is_stream {
+        bedrock_stream(state, &client, bedrock_body, &mapped_model).await
+    } else {
+        bedrock_non_stream(state, &client, bedrock_body, &mapped_model).await
+    }
+}
+
+/// Non-streaming Bedrock request.
+async fn bedrock_non_stream(
+    state: AppState,
+    client: &crate::backend::bedrock_client::BedrockClient,
+    body: bytes::Bytes,
+    model_id: &str,
+) -> Response {
+    match client.forward(body, model_id).await {
+        Ok((resp_body, rate_limits)) => {
+            state.metrics.record_success();
+            let mut resp = (
+                StatusCode::OK,
+                [("content-type", "application/json")],
+                resp_body,
+            )
+                .into_response();
+            rate_limits.inject_anthropic_response_headers(resp.headers_mut());
+            resp
+        }
+        Err(e) => {
+            state.metrics.record_error();
+            bedrock_error_to_response(e)
+        }
+    }
+}
+
+/// Streaming Bedrock request. Decodes AWS Event Stream binary frames into
+/// Anthropic SSE events and re-emits them as standard SSE.
+async fn bedrock_stream(
+    state: AppState,
+    client: &crate::backend::bedrock_client::BedrockClient,
+    body: bytes::Bytes,
+    model_id: &str,
+) -> Response {
+    let (response, rate_limits) = match client.forward_stream(body, model_id).await {
+        Ok(r) => r,
+        Err(e) => {
+            state.metrics.record_error();
+            return bedrock_error_to_response(e);
+        }
+    };
+
+    let (tx, rx) =
+        tokio::sync::mpsc::channel::<Result<String, std::convert::Infallible>>(32);
+    let metrics = state.metrics.clone();
+
+    tokio::spawn(async move {
+        let mut byte_stream = response.bytes_stream();
+        let mut event_buf = BytesMut::new();
+
+        while let Some(chunk_result) = byte_stream.next().await {
+            let bytes = match chunk_result {
+                Ok(b) => b,
+                Err(e) => {
+                    tracing::error!("Bedrock stream read error: {e}");
+                    metrics.record_error();
+                    return;
+                }
+            };
+            event_buf.extend_from_slice(&bytes);
+
+            // Decode all complete frames in the buffer
+            while let Some(payload) = eventstream::decode_frame(&mut event_buf) {
+                if let Some(event_json) = eventstream::extract_event_from_payload(&payload) {
+                    // Re-emit as SSE: "event: <type>\ndata: <json>\n\n"
+                    // Bedrock events are raw Anthropic JSON; detect the event type.
+                    let event_type = detect_event_type(&event_json);
+                    let sse_line = format!("event: {event_type}\ndata: {event_json}\n\n");
+                    if tx.send(Ok(sse_line)).await.is_err() {
+                        return; // client disconnected
+                    }
+                }
+            }
+        }
+        metrics.record_success();
+    });
+
+    let body_stream = tokio_stream::wrappers::ReceiverStream::new(rx);
+    let body = axum::body::Body::from_stream(body_stream);
+    let mut resp = axum::http::Response::builder()
+        .status(StatusCode::OK)
+        .header("content-type", "text/event-stream")
+        .header("cache-control", "no-cache")
+        .body(body)
+        .unwrap_or_else(|_| StatusCode::INTERNAL_SERVER_ERROR.into_response());
+    rate_limits.inject_anthropic_response_headers(resp.headers_mut());
+    resp
+}
+
+/// Detect the Anthropic SSE event type from a JSON string.
+/// Falls back to "message" if the type field is not found.
+fn detect_event_type(json: &str) -> &str {
+    // Quick extraction without full parse: look for "type":"..."
+    if let Ok(v) = serde_json::from_str::<serde_json::Value>(json) {
+        if let Some(t) = v.get("type").and_then(|t| t.as_str()) {
+            return match t {
+                "message_start" => "message_start",
+                "content_block_start" => "content_block_start",
+                "content_block_delta" => "content_block_delta",
+                "content_block_stop" => "content_block_stop",
+                "message_delta" => "message_delta",
+                "message_stop" => "message_stop",
+                "ping" => "ping",
+                _ => "message",
+            };
+        }
+    }
+    "message"
+}
+
+/// Convert a BedrockClientError into a Response.
+fn bedrock_error_to_response(error: BedrockClientError) -> Response {
+    match error {
+        BedrockClientError::ApiError { status, body } => {
+            let http_status =
+                StatusCode::from_u16(status).unwrap_or(StatusCode::INTERNAL_SERVER_ERROR);
+            // Try to return body as-is (Bedrock may return JSON error)
+            (http_status, [("content-type", "application/json")], body).into_response()
+        }
+        BedrockClientError::Transport(msg) => {
+            tracing::error!("Bedrock transport error: {msg}");
+            let err = mapping::errors_map::create_anthropic_error(
+                anthropic::ErrorType::ApiError,
+                "An internal error occurred while communicating with the upstream service."
+                    .to_string(),
+                None,
+            );
+            (StatusCode::BAD_GATEWAY, Json(err)).into_response()
+        }
+        BedrockClientError::Signing(msg) => {
+            tracing::error!("Bedrock signing error: {msg}");
+            let err = mapping::errors_map::create_anthropic_error(
+                anthropic::ErrorType::ApiError,
+                "Failed to sign request for AWS Bedrock.".to_string(),
+                None,
+            );
+            (StatusCode::INTERNAL_SERVER_ERROR, Json(err)).into_response()
+        }
+    }
+}
diff --git a/crates/proxy/src/server/chat_completions.rs b/crates/proxy/src/server/chat_completions.rs
index aaf75f7..479caf7 100644
--- a/crates/proxy/src/server/chat_completions.rs
+++ b/crates/proxy/src/server/chat_completions.rs
@@ -234,8 +234,8 @@ pub(crate) async fn chat_completions(
                 }
             }
         }
-        BackendClient::Anthropic(_) => openai_error_response(
-            "Anthropic passthrough backend does not support /v1/chat/completions",
+        BackendClient::Anthropic(_) | BackendClient::Bedrock(_) => openai_error_response(
+            "This backend does not support /v1/chat/completions. Use /v1/messages instead.",
             "invalid_request_error",
             StatusCode::BAD_REQUEST,
         ),
@@ -273,9 +273,9 @@ async fn chat_completions_stream(
         | BackendClient::Vertex(c)
         | BackendClient::GeminiOpenAI(c)
         | BackendClient::OpenAIResponses(c) => c.clone(),
-        BackendClient::Anthropic(_) => {
+        BackendClient::Anthropic(_) | BackendClient::Bedrock(_) => {
             return openai_error_response(
-                "Anthropic passthrough backend does not support /v1/chat/completions streaming",
+                "This backend does not support /v1/chat/completions. Use /v1/messages instead.",
                 "invalid_request_error",
                 StatusCode::BAD_REQUEST,
             );
diff --git a/crates/proxy/src/server/mod.rs b/crates/proxy/src/server/mod.rs
index a176122..b6494cc 100644
--- a/crates/proxy/src/server/mod.rs
+++ b/crates/proxy/src/server/mod.rs
@@ -1,3 +1,5 @@
+/// AWS Bedrock passthrough handler (SigV4 signing + event stream decoding).
+mod bedrock_passthrough;
 /// OpenAI Chat Completions input handler (POST /v1/chat/completions).
 mod chat_completions;
 /// Auth validation, request ID injection, size limits, concurrency limits, header logging.
diff --git a/crates/proxy/src/server/routes.rs b/crates/proxy/src/server/routes.rs
index 2d48ffe..fc9809d 100644
--- a/crates/proxy/src/server/routes.rs
+++ b/crates/proxy/src/server/routes.rs
@@ -131,7 +131,7 @@ pub fn app_multi_with_shared(config: MultiConfig, shared: Option<SharedState>) -
 
     // Build per-backend sub-routers. Keep a map of AppState so the default
     // backend can reuse the same state (same semaphore, same reqwest client).
-    let mut backend_states: HashMap<String, (AppState, bool)> = HashMap::new();
+    let mut backend_states: HashMap<String, (AppState, HandlerMode)> = HashMap::new();
     for (name, bc) in &config.backends {
         let metrics = Metrics::new();
         backend_metrics.insert(name.clone(), metrics.clone());
@@ -146,9 +146,13 @@ pub fn app_multi_with_shared(config: MultiConfig, shared: Option<SharedState>) -
             omit_stream_options: bc.omit_stream_options,
         };
 
-        let is_anthropic = bc.kind == BackendKind::Anthropic;
-        let sub = backend_router(state.clone(), is_anthropic);
-        backend_states.insert(name.clone(), (state, is_anthropic));
+        let mode = match bc.kind {
+            BackendKind::Anthropic => HandlerMode::Anthropic,
+            BackendKind::Bedrock => HandlerMode::Bedrock,
+            _ => HandlerMode::Translate,
+        };
+        let sub = backend_router(state.clone(), mode);
+        backend_states.insert(name.clone(), (state, mode));
 
         // Nest under /{name}/
         router = router.nest(&format!("/{name}"), sub);
@@ -156,8 +160,8 @@ pub fn app_multi_with_shared(config: MultiConfig, shared: Option<SharedState>) -
 
     // Default backend: also serve at un-prefixed /v1/messages for backward compat.
     // Reuses the same AppState (shared semaphore, connection pool) as the named route.
-    if let Some((default_state, is_anthropic)) = backend_states.get(&config.default_backend) {
-        let default_sub = backend_router(default_state.clone(), *is_anthropic);
+    if let Some((default_state, mode)) = backend_states.get(&config.default_backend) {
+        let default_sub = backend_router(default_state.clone(), *mode);
         router = router.merge(default_sub);
     }
 
@@ -217,15 +221,30 @@ async fn fallback_not_found() -> Response {
     (StatusCode::NOT_FOUND, Json(err)).into_response()
 }
 
+/// Which handler mode a backend uses.
+#[derive(Debug, Clone, Copy)]
+enum HandlerMode {
+    /// Anthropic passthrough (no translation, forwards raw bytes).
+    Anthropic,
+    /// Bedrock (SigV4 signing, event stream decoding, Anthropic format).
+    Bedrock,
+    /// Translation (Anthropic -> OpenAI -> backend -> OpenAI -> Anthropic).
+    Translate,
+}
+
 /// Build the sub-router for a single backend.
-/// If `is_anthropic` is true, uses the passthrough handler instead of the translation handler.
-fn backend_router(state: AppState, is_anthropic: bool) -> Router<GlobalState> {
-    let api_routes = if is_anthropic {
-        Router::new()
+fn backend_router(state: AppState, mode: HandlerMode) -> Router<GlobalState> {
+    let api_routes = match mode {
+        HandlerMode::Anthropic => Router::new()
             .route("/v1/messages", post(anthropic_passthrough))
-            .route("/v1/models", get(models))
-    } else {
-        Router::new()
+            .route("/v1/models", get(models)),
+        HandlerMode::Bedrock => Router::new()
+            .route(
+                "/v1/messages",
+                post(super::bedrock_passthrough::bedrock_passthrough),
+            )
+            .route("/v1/models", get(models)),
+        HandlerMode::Translate => Router::new()
             .route("/v1/messages", post(messages))
             .route(
                 "/v1/chat/completions",
@@ -234,7 +253,7 @@ fn backend_router(state: AppState, is_anthropic: bool) -> Router<GlobalState> {
             .route("/v1/models", get(models))
             .route("/v1/messages/count_tokens", post(count_tokens))
             .route("/v1/messages/batches", post(batches))
-            .route("/v1/embeddings", post(embeddings))
+            .route("/v1/embeddings", post(embeddings)),
     };
 
     api_routes
@@ -563,12 +582,12 @@ async fn messages(
                 }
             }
         }
-        BackendClient::Anthropic(_) => {
-            // Anthropic passthrough is handled by a separate handler that works with raw bytes.
+        BackendClient::Anthropic(_) | BackendClient::Bedrock(_) => {
+            // These backends are handled by separate handlers (passthrough / Bedrock).
             // If we reach here, something is misconfigured.
             let err = mapping::errors_map::create_anthropic_error(
                 anthropic::ErrorType::ApiError,
-                "Anthropic passthrough does not use the translation handler".to_string(),
+                "This backend does not use the translation handler".to_string(),
                 None,
             );
             (StatusCode::INTERNAL_SERVER_ERROR, Json(err)).into_response()
diff --git a/crates/proxy/src/server/streaming.rs b/crates/proxy/src/server/streaming.rs
index afef5a2..9d9bb5c 100644
--- a/crates/proxy/src/server/streaming.rs
+++ b/crates/proxy/src/server/streaming.rs
@@ -320,11 +320,11 @@ pub(crate) async fn messages_stream(
                 }
             });
         }
-        BackendClient::Anthropic(_) => {
+        BackendClient::Anthropic(_) | BackendClient::Bedrock(_) => {
             drop(rl_tx);
             let _ = tx
                 .send(Ok(Event::default().data(
-                    r#"{"error":"anthropic passthrough does not use this handler"}"#,
+                    r#"{"error":"this backend does not use the translation streaming handler"}"#,
                 )))
                 .await;
         }
diff --git a/crates/proxy/tests/chat_completions.rs b/crates/proxy/tests/chat_completions.rs
new file mode 100644
index 0000000..76235ec
--- /dev/null
+++ b/crates/proxy/tests/chat_completions.rs
@@ -0,0 +1,230 @@
+// Integration tests for POST /v1/chat/completions (OpenAI-format input).
+
+use anyllm_proxy::config::{self, BackendAuth, BackendKind, Config, ModelMapping, OpenAIApiFormat};
+use anyllm_proxy::server::routes;
+use axum::{routing::post, Router};
+use reqwest::Client;
+use serde_json::json;
+use tokio::net::TcpListener;
+
+fn openai_config_with_base(base_url: &str) -> Config {
+    Config {
+        backend: BackendKind::OpenAI,
+        openai_api_key: "test-key".to_string(),
+        openai_base_url: base_url.to_string(),
+        listen_port: 0,
+        model_mapping: ModelMapping {
+            big_model: "gpt-4o".into(),
+            small_model: "gpt-4o-mini".into(),
+        },
+        tls: config::TlsConfig::default(),
+        backend_auth: BackendAuth::BearerToken("test-key".into()),
+        log_bodies: false,
+        openai_api_format: OpenAIApiFormat::Chat,
+    }
+}
+
+/// Mock backend that returns a fixed OpenAI Chat Completions response.
+async fn spawn_mock_chat_backend() -> String {
+    let app = Router::new().route(
+        "/v1/chat/completions",
+        post(|| async {
+            axum::Json(json!({
+                "id": "chatcmpl-mock123",
+                "object": "chat.completion",
+                "created": 1700000000,
+                "model": "gpt-4o",
+                "choices": [{
+                    "index": 0,
+                    "message": {
+                        "role": "assistant",
+                        "content": "Hello from mock!"
+                    },
+                    "finish_reason": "stop"
+                }],
+                "usage": {
+                    "prompt_tokens": 10,
+                    "completion_tokens": 5,
+                    "total_tokens": 15
+                }
+            }))
+        }),
+    );
+
+    let listener = TcpListener::bind("127.0.0.1:0").await.unwrap();
+    let addr = listener.local_addr().unwrap();
+    tokio::spawn(async move { axum::serve(listener, app).await.unwrap() });
+    format!("http://{addr}")
+}
+
+async fn spawn_proxy(config: Config) -> String {
+    std::env::set_var("PROXY_OPEN_RELAY", "true");
+    let app = routes::app(config);
+    let listener = TcpListener::bind("127.0.0.1:0").await.unwrap();
+    let addr = listener.local_addr().unwrap();
+    tokio::spawn(async move { axum::serve(listener, app).await.unwrap() });
+    format!("http://{addr}")
+}
+
+#[tokio::test]
+async fn chat_completions_non_streaming() {
+    let mock = spawn_mock_chat_backend().await;
+    let proxy = spawn_proxy(openai_config_with_base(&mock)).await;
+
+    let client = Client::new();
+    let resp = client
+        .post(format!("{proxy}/v1/chat/completions"))
+        .header("x-api-key", "test")
+        .header("content-type", "application/json")
+        .json(&json!({
+            "model": "claude-sonnet-4-20250514",
+            "messages": [{"role": "user", "content": "Hello"}],
+            "max_tokens": 100
+        }))
+        .send()
+        .await
+        .unwrap();
+
+    assert_eq!(resp.status(), 200);
+    let body: serde_json::Value = resp.json().await.unwrap();
+    assert_eq!(body["object"], "chat.completion");
+    assert!(body["id"].as_str().unwrap().starts_with("chatcmpl-"));
+    assert_eq!(body["choices"][0]["message"]["role"], "assistant");
+    assert!(body["choices"][0]["message"]["content"].as_str().is_some());
+    assert_eq!(body["choices"][0]["finish_reason"], "stop");
+    assert!(body["usage"]["prompt_tokens"].as_u64().is_some());
+}
+
+#[tokio::test]
+async fn chat_completions_missing_max_tokens_returns_400() {
+    let mock = spawn_mock_chat_backend().await;
+    let proxy = spawn_proxy(openai_config_with_base(&mock)).await;
+
+    let client = Client::new();
+    let resp = client
+        .post(format!("{proxy}/v1/chat/completions"))
+        .header("x-api-key", "test")
+        .header("content-type", "application/json")
+        .json(&json!({
+            "model": "claude-sonnet-4-20250514",
+            "messages": [{"role": "user", "content": "Hello"}]
+        }))
+        .send()
+        .await
+        .unwrap();
+
+    assert_eq!(resp.status(), 400);
+    let body: serde_json::Value = resp.json().await.unwrap();
+    assert_eq!(body["error"]["type"], "invalid_request_error");
+}
+
+#[tokio::test]
+async fn chat_completions_empty_messages_returns_400() {
+    let mock = spawn_mock_chat_backend().await;
+    let proxy = spawn_proxy(openai_config_with_base(&mock)).await;
+
+    let client = Client::new();
+    let resp = client
+        .post(format!("{proxy}/v1/chat/completions"))
+        .header("x-api-key", "test")
+        .header("content-type", "application/json")
+        .json(&json!({
+            "model": "claude-sonnet-4-20250514",
+            "messages": [],
+            "max_tokens": 100
+        }))
+        .send()
+        .await
+        .unwrap();
+
+    assert_eq!(resp.status(), 400);
+    let body: serde_json::Value = resp.json().await.unwrap();
+    assert_eq!(body["error"]["type"], "invalid_request_error");
+}
+
+#[tokio::test]
+async fn chat_completions_degradation_header_on_lossy_fields() {
+    let mock = spawn_mock_chat_backend().await;
+    let proxy = spawn_proxy(openai_config_with_base(&mock)).await;
+
+    let client = Client::new();
+    let resp = client
+        .post(format!("{proxy}/v1/chat/completions"))
+        .header("x-api-key", "test")
+        .header("content-type", "application/json")
+        .json(&json!({
+            "model": "claude-sonnet-4-20250514",
+            "messages": [{"role": "user", "content": "Hi"}],
+            "max_tokens": 100,
+            "presence_penalty": 0.5,
+            "frequency_penalty": 0.3
+        }))
+        .send()
+        .await
+        .unwrap();
+
+    assert_eq!(resp.status(), 200);
+    let degradation = resp
+        .headers()
+        .get("x-anyllm-degradation")
+        .and_then(|v| v.to_str().ok())
+        .unwrap_or("");
+    assert!(
+        degradation.contains("presence_penalty"),
+        "expected presence_penalty in degradation header, got: {degradation}"
+    );
+    assert!(
+        degradation.contains("frequency_penalty"),
+        "expected frequency_penalty in degradation header, got: {degradation}"
+    );
+}
+
+#[tokio::test]
+async fn chat_completions_with_system_message() {
+    let mock = spawn_mock_chat_backend().await;
+    let proxy = spawn_proxy(openai_config_with_base(&mock)).await;
+
+    let client = Client::new();
+    let resp = client
+        .post(format!("{proxy}/v1/chat/completions"))
+        .header("x-api-key", "test")
+        .header("content-type", "application/json")
+        .json(&json!({
+            "model": "claude-sonnet-4-20250514",
+            "messages": [
+                {"role": "system", "content": "You are a helpful assistant."},
+                {"role": "user", "content": "Hello"}
+            ],
+            "max_tokens": 100
+        }))
+        .send()
+        .await
+        .unwrap();
+
+    assert_eq!(resp.status(), 200);
+    let body: serde_json::Value = resp.json().await.unwrap();
+    assert_eq!(body["object"], "chat.completion");
+}
+
+#[tokio::test]
+async fn chat_completions_returns_openai_error_format() {
+    let mock = spawn_mock_chat_backend().await;
+    let proxy = spawn_proxy(openai_config_with_base(&mock)).await;
+
+    let client = Client::new();
+    // Send completely invalid JSON
+    let resp = client
+        .post(format!("{proxy}/v1/chat/completions"))
+        .header("x-api-key", "test")
+        .header("content-type", "application/json")
+        .body("not json")
+        .send()
+        .await
+        .unwrap();
+
+    assert_eq!(resp.status(), 400);
+    let body: serde_json::Value = resp.json().await.unwrap();
+    // Should have OpenAI error format (error.type, error.message)
+    assert!(body["error"]["type"].is_string());
+    assert!(body["error"]["message"].is_string());
+}
diff --git a/crates/proxy/tests/embeddings.rs b/crates/proxy/tests/embeddings.rs
new file mode 100644
index 0000000..7da08ac
--- /dev/null
+++ b/crates/proxy/tests/embeddings.rs
@@ -0,0 +1,226 @@
+// Integration tests for POST /v1/embeddings passthrough and x-anyllm-degradation header.
+
+use anyllm_proxy::config::{self, BackendAuth, BackendKind, Config, ModelMapping, OpenAIApiFormat};
+use anyllm_proxy::server::routes;
+use axum::{routing::post, Router};
+use reqwest::Client;
+use tokio::net::TcpListener;
+
+fn openai_config_with_base(base_url: &str) -> Config {
+    Config {
+        backend: BackendKind::OpenAI,
+        openai_api_key: "test-key".to_string(),
+        openai_base_url: base_url.to_string(),
+        listen_port: 0,
+        model_mapping: ModelMapping {
+            big_model: "gpt-4o".into(),
+            small_model: "gpt-4o-mini".into(),
+        },
+        tls: config::TlsConfig::default(),
+        backend_auth: BackendAuth::BearerToken("test-key".into()),
+        log_bodies: false,
+        openai_api_format: OpenAIApiFormat::Chat,
+    }
+}
+
+fn anthropic_config() -> Config {
+    Config {
+        backend: BackendKind::Anthropic,
+        openai_api_key: String::new(),
+        openai_base_url: "https://api.anthropic.com".to_string(),
+        listen_port: 0,
+        model_mapping: ModelMapping {
+            big_model: "claude-opus-4-6".into(),
+            small_model: "claude-haiku-4-5".into(),
+        },
+        tls: config::TlsConfig::default(),
+        backend_auth: BackendAuth::BearerToken("test-key".into()),
+        log_bodies: false,
+        openai_api_format: OpenAIApiFormat::Chat,
+    }
+}
+
+/// Start a mock backend that accepts POST /v1/embeddings and returns a fixed response.
+/// Returns the base URL of the mock server.
+async fn spawn_mock_backend() -> String {
+    let app = Router::new().route(
+        "/v1/embeddings",
+        post(|| async {
+            axum::response::Response::builder()
+                .status(200)
+                .header("content-type", "application/json")
+                .body(axum::body::Body::from(
+                    r#"{"object":"list","data":[{"object":"embedding","index":0,"embedding":[0.1,0.2,0.3]}],"model":"text-embedding-3-small","usage":{"prompt_tokens":5,"total_tokens":5}}"#,
+                ))
+                .unwrap()
+        }),
+    );
+
+    let listener = TcpListener::bind("127.0.0.1:0").await.unwrap();
+    let addr = listener.local_addr().unwrap();
+    tokio::spawn(async move { axum::serve(listener, app).await.unwrap() });
+    format!("http://{addr}")
+}
+
+async fn spawn_proxy_with_config(config: Config) -> String {
+    std::env::set_var("PROXY_OPEN_RELAY", "true");
+    let app = routes::app(config);
+    let listener = TcpListener::bind("127.0.0.1:0").await.unwrap();
+    let addr = listener.local_addr().unwrap();
+    tokio::spawn(async move { axum::serve(listener, app).await.unwrap() });
+    format!("http://{addr}")
+}
+
+#[tokio::test]
+async fn embeddings_forwarded_to_backend() {
+    let mock_base = spawn_mock_backend().await;
+    let proxy_base = spawn_proxy_with_config(openai_config_with_base(&mock_base)).await;
+
+    let client = Client::new();
+    let resp = client
+        .post(format!("{proxy_base}/v1/embeddings"))
+        .header("x-api-key", "test")
+        .header("content-type", "application/json")
+        .body(r#"{"model":"text-embedding-3-small","input":"hello world"}"#)
+        .send()
+        .await
+        .unwrap();
+
+    assert_eq!(resp.status(), 200);
+    let body: serde_json::Value = resp.json().await.unwrap();
+    assert_eq!(body["object"], "list");
+    assert!(body["data"].is_array());
+}
+
+#[tokio::test]
+async fn embeddings_response_content_type_forwarded() {
+    let mock_base = spawn_mock_backend().await;
+    let proxy_base = spawn_proxy_with_config(openai_config_with_base(&mock_base)).await;
+
+    let client = Client::new();
+    let resp = client
+        .post(format!("{proxy_base}/v1/embeddings"))
+        .header("x-api-key", "test")
+        .header("content-type", "application/json")
+        .body(r#"{"model":"text-embedding-3-small","input":"hello"}"#)
+        .send()
+        .await
+        .unwrap();
+
+    assert_eq!(resp.status(), 200);
+    let ct = resp
+        .headers()
+        .get("content-type")
+        .unwrap()
+        .to_str()
+        .unwrap();
+    assert!(ct.contains("application/json"), "got content-type: {ct}");
+}
+
+#[tokio::test]
+async fn embeddings_requires_auth() {
+    let mock_base = spawn_mock_backend().await;
+    let proxy_base = spawn_proxy_with_config(openai_config_with_base(&mock_base)).await;
+
+    // Temporarily unset open-relay mode so auth is enforced
+    std::env::remove_var("PROXY_OPEN_RELAY");
+    let proxy_strict = {
+        let mut c = openai_config_with_base(&mock_base);
+        c.openai_api_key = "sk-real".into();
+        spawn_proxy_with_config(c).await
+    };
+
+    let client = Client::new();
+    let resp = client
+        .post(format!("{proxy_strict}/v1/embeddings"))
+        .header("content-type", "application/json")
+        .body(r#"{"model":"text-embedding-3-small","input":"hello"}"#)
+        .send()
+        .await
+        .unwrap();
+
+    assert_eq!(resp.status(), 401);
+
+    // Restore for other tests
+    std::env::set_var("PROXY_OPEN_RELAY", "true");
+    // Silence the unused-variable warning — proxy_base was used above
+    let _ = proxy_base;
+}
+
+#[tokio::test]
+async fn embeddings_not_routed_for_anthropic_backend() {
+    // Anthropic backend does not mount /v1/embeddings; route returns 404.
+    let proxy_base = spawn_proxy_with_config(anthropic_config()).await;
+
+    let client = Client::new();
+    let resp = client
+        .post(format!("{proxy_base}/v1/embeddings"))
+        .header("x-api-key", "test")
+        .header("content-type", "application/json")
+        .body(r#"{"model":"text-embedding-3-small","input":"hello"}"#)
+        .send()
+        .await
+        .unwrap();
+
+    // Route not registered for Anthropic backend; fallback returns 404.
+    assert_eq!(resp.status(), 404);
+}
+
+#[tokio::test]
+async fn degradation_header_present_when_top_k_set() {
+    // We can't make a full messages round-trip without a real backend, but we can
+    // verify the compute_request_warnings function directly via the translator crate.
+    // The proxy-level injection is covered by the inject_degradation_header unit test.
+    use anyllm_translate::anthropic;
+    let req = anthropic::MessageCreateRequest {
+        model: "claude-sonnet-4-6".to_string(),
+        max_tokens: 100,
+        messages: vec![anthropic::InputMessage {
+            role: anthropic::Role::User,
+            content: anthropic::Content::Text("hi".to_string()),
+        }],
+        system: None,
+        temperature: None,
+        top_p: None,
+        top_k: Some(40),
+        stop_sequences: None,
+        tools: None,
+        tool_choice: None,
+        metadata: None,
+        thinking: None,
+        stream: None,
+        extra: serde_json::Map::new(),
+    };
+
+    let warnings = anyllm_translate::compute_request_warnings(&req);
+    let header_val = warnings.as_header_value().expect("should have warnings");
+    assert!(header_val.contains("top_k"), "got: {header_val}");
+}
+
+#[tokio::test]
+async fn degradation_header_absent_when_no_lossy_features() {
+    use anyllm_translate::anthropic;
+    let req = anthropic::MessageCreateRequest {
+        model: "claude-sonnet-4-6".to_string(),
+        max_tokens: 100,
+        messages: vec![anthropic::InputMessage {
+            role: anthropic::Role::User,
+            content: anthropic::Content::Text("hi".to_string()),
+        }],
+        system: None,
+        temperature: None,
+        top_p: None,
+        top_k: None,
+        stop_sequences: None,
+        tools: None,
+        tool_choice: None,
+        metadata: None,
+        thinking: None,
+        stream: None,
+        extra: serde_json::Map::new(),
+    };
+
+    let warnings = anyllm_translate::compute_request_warnings(&req);
+    assert!(warnings.is_empty());
+    assert!(warnings.as_header_value().is_none());
+}
diff --git a/crates/proxy/tests/live_bedrock.rs b/crates/proxy/tests/live_bedrock.rs
new file mode 100644
index 0000000..99430f9
--- /dev/null
+++ b/crates/proxy/tests/live_bedrock.rs
@@ -0,0 +1,76 @@
+// Live integration tests for the Bedrock backend.
+// Requires AWS credentials in the environment. Run with:
+//   AWS_REGION=us-east-1 AWS_ACCESS_KEY_ID=... AWS_SECRET_ACCESS_KEY=... \
+//     cargo test --test live_bedrock -- --ignored --test-threads=1
+
+use reqwest::Client;
+use serde_json::json;
+
+/// Start the proxy with Bedrock backend and return the base URL.
+fn start_proxy() -> String {
+    // These tests are run manually with real credentials.
+    // The proxy must be started externally, or we construct a test server here.
+    let port = std::env::var("TEST_PROXY_PORT").unwrap_or_else(|_| "3099".to_string());
+    format!("http://127.0.0.1:{port}")
+}
+
+#[tokio::test]
+#[ignore]
+async fn bedrock_non_streaming() {
+    let base = start_proxy();
+    let client = Client::new();
+
+    let resp = client
+        .post(format!("{base}/v1/messages"))
+        .header("x-api-key", "test-key")
+        .header("content-type", "application/json")
+        .json(&json!({
+            "model": "anthropic.claude-haiku-4-5-20251001-v1:0",
+            "max_tokens": 64,
+            "messages": [{"role": "user", "content": "Say hello in one word."}]
+        }))
+        .send()
+        .await
+        .expect("request failed");
+
+    let status = resp.status().as_u16();
+    let body: serde_json::Value = resp.json().await.expect("invalid JSON response");
+
+    assert_eq!(status, 200, "unexpected status: {body}");
+    assert_eq!(body["type"], "message");
+    assert!(body["content"].as_array().map_or(false, |a| !a.is_empty()));
+}
+
+#[tokio::test]
+#[ignore]
+async fn bedrock_streaming() {
+    let base = start_proxy();
+    let client = Client::new();
+
+    let resp = client
+        .post(format!("{base}/v1/messages"))
+        .header("x-api-key", "test-key")
+        .header("content-type", "application/json")
+        .json(&json!({
+            "model": "anthropic.claude-haiku-4-5-20251001-v1:0",
+            "max_tokens": 64,
+            "stream": true,
+            "messages": [{"role": "user", "content": "Say hello."}]
+        }))
+        .send()
+        .await
+        .expect("request failed");
+
+    let status = resp.status().as_u16();
+    assert_eq!(status, 200, "expected 200 for streaming");
+
+    let body = resp.text().await.expect("failed to read body");
+    assert!(
+        body.contains("message_start"),
+        "expected message_start event in stream"
+    );
+    assert!(
+        body.contains("message_stop"),
+        "expected message_stop event in stream"
+    );
+}
diff --git a/crates/translator/src/anthropic/messages.rs b/crates/translator/src/anthropic/messages.rs
index db64362..591de2a 100644
--- a/crates/translator/src/anthropic/messages.rs
+++ b/crates/translator/src/anthropic/messages.rs
@@ -197,9 +197,17 @@ pub struct Tool {
 #[serde(tag = "type")]
 pub enum ToolChoice {
     #[serde(rename = "auto")]
-    Auto,
+    Auto {
+        /// Disable parallel tool use. Maps to OpenAI `parallel_tool_calls: false`.
+        #[serde(default, skip_serializing_if = "Option::is_none")]
+        disable_parallel_tool_use: Option<bool>,
+    },
     #[serde(rename = "any")]
-    Any,
+    Any {
+        /// Disable parallel tool use. Maps to OpenAI `parallel_tool_calls: false`.
+        #[serde(default, skip_serializing_if = "Option::is_none")]
+        disable_parallel_tool_use: Option<bool>,
+    },
     #[serde(rename = "none")]
     None,
     #[serde(rename = "tool")]
@@ -365,7 +373,7 @@ mod tests {
         assert_eq!(tools[0].name, "get_weather");
         assert!(tools[0].description.is_some());
         match req.tool_choice.unwrap() {
-            ToolChoice::Auto => {}
+            ToolChoice::Auto { .. } => {}
             other => panic!("expected ToolChoice::Auto, got {:?}", other),
         }
     }
diff --git a/crates/translator/src/mapping/message_map.rs b/crates/translator/src/mapping/message_map.rs
index cfa44d0..0a86fb8 100644
--- a/crates/translator/src/mapping/message_map.rs
+++ b/crates/translator/src/mapping/message_map.rs
@@ -1,7 +1,7 @@
 // Anthropic <-> OpenAI message mapping
 
 use crate::anthropic;
-use crate::mapping::{streaming_map, tools_map, usage_map};
+use crate::mapping::{streaming_map, tools_map, usage_map, warnings::TranslationWarnings};
 use crate::openai;
 use crate::util;
 
@@ -23,6 +23,44 @@ pub fn extract_system_text(system: &anthropic::System) -> String {
     }
 }
 
+/// Compute degradation warnings for an Anthropic request without performing translation.
+///
+/// Returns a `TranslationWarnings` value listing every feature that will be silently
+/// dropped or degraded when this request is translated to an OpenAI request.
+/// The proxy injects these as an `x-anyllm-degradation` response header so clients
+/// can detect silent drops without inspecting server logs.
+pub fn compute_request_warnings(req: &anthropic::MessageCreateRequest) -> TranslationWarnings {
+    let mut w = TranslationWarnings::default();
+    if req.top_k.is_some() {
+        w.add("top_k");
+    }
+    if req.thinking.is_some() {
+        w.add("thinking_config");
+    }
+    if let Some(ref seqs) = req.stop_sequences {
+        if seqs.len() > 4 {
+            w.add("stop_sequences_truncated");
+        }
+    }
+    if let Some(anthropic::System::Blocks(blocks)) = &req.system {
+        if blocks.iter().any(|b| b.cache_control.is_some()) {
+            w.add("cache_control");
+        }
+    }
+    let has_document = req.messages.iter().any(|msg| {
+        match &msg.content {
+            anthropic::Content::Blocks(blocks) => blocks.iter().any(|b| {
+                matches!(b, anthropic::ContentBlock::Document { .. })
+            }),
+            _ => false,
+        }
+    });
+    if has_document {
+        w.add("document_blocks");
+    }
+    w
+}
+
 /// Convert an Anthropic MessageCreateRequest to an OpenAI ChatCompletionRequest.
 ///
 /// Anthropic: <https://docs.anthropic.com/en/api/messages>
@@ -64,6 +102,14 @@ pub fn anthropic_to_openai_request(
         .as_ref()
         .map(tools_map::anthropic_tool_choice_to_openai);
 
+    // Map disable_parallel_tool_use to OpenAI parallel_tool_calls.
+    // Compat spec: "Fully supported". See: https://docs.anthropic.com/en/api/openai-sdk#tools--functions-fields
+    let parallel_tool_calls = match req.tool_choice.as_ref() {
+        Some(anthropic::ToolChoice::Auto { disable_parallel_tool_use: Some(true) })
+        | Some(anthropic::ToolChoice::Any { disable_parallel_tool_use: Some(true) }) => Some(false),
+        _ => None,
+    };
+
     // Map metadata.user_id to OpenAI user field.
     // Compat spec: user is "Ignored", but we forward it for traceability.
     // See: https://docs.anthropic.com/en/api/openai-sdk#simple-fields
@@ -127,7 +173,7 @@ pub fn anthropic_to_openai_request(
         frequency_penalty: None,
         response_format: None,
         user,
-        parallel_tool_calls: None,
+        parallel_tool_calls,
         extra: req.extra.clone(),
     };
 
@@ -678,7 +724,7 @@ mod tests {
     #[test]
     fn tool_choice_auto() {
         let mut req = basic_request();
-        req.tool_choice = Some(anthropic::ToolChoice::Auto);
+        req.tool_choice = Some(anthropic::ToolChoice::Auto { disable_parallel_tool_use: None });
         let oai = anthropic_to_openai_request(&req);
         assert!(matches!(
             oai.tool_choice,
@@ -689,7 +735,7 @@ mod tests {
     #[test]
     fn tool_choice_any_becomes_required() {
         let mut req = basic_request();
-        req.tool_choice = Some(anthropic::ToolChoice::Any);
+        req.tool_choice = Some(anthropic::ToolChoice::Any { disable_parallel_tool_use: None });
         let oai = anthropic_to_openai_request(&req);
         assert!(matches!(
             oai.tool_choice,
@@ -724,6 +770,33 @@ mod tests {
         }
     }
 
+    #[test]
+    fn disable_parallel_tool_use_sets_parallel_tool_calls_false() {
+        let mut req = basic_request();
+        req.tool_choice = Some(anthropic::ToolChoice::Auto {
+            disable_parallel_tool_use: Some(true),
+        });
+        let oai = anthropic_to_openai_request(&req);
+        assert_eq!(oai.parallel_tool_calls, Some(false));
+    }
+
+    #[test]
+    fn disable_parallel_tool_use_false_leaves_parallel_tool_calls_none() {
+        let mut req = basic_request();
+        req.tool_choice = Some(anthropic::ToolChoice::Auto {
+            disable_parallel_tool_use: Some(false),
+        });
+        let oai = anthropic_to_openai_request(&req);
+        assert!(oai.parallel_tool_calls.is_none());
+    }
+
+    #[test]
+    fn no_tool_choice_leaves_parallel_tool_calls_none() {
+        let req = basic_request();
+        let oai = anthropic_to_openai_request(&req);
+        assert!(oai.parallel_tool_calls.is_none());
+    }
+
     #[test]
     fn stop_sequences_capped_at_four() {
         let mut req = basic_request();
@@ -2110,4 +2183,100 @@ mod tests {
         let oai = anthropic_to_openai_request(&req);
         assert_eq!(oai.temperature, Some(0.7));
     }
+
+    // --- compute_request_warnings ---
+
+    #[test]
+    fn warnings_empty_for_plain_request() {
+        let req = basic_request();
+        let w = compute_request_warnings(&req);
+        assert!(w.is_empty());
+        assert!(w.as_header_value().is_none());
+    }
+
+    #[test]
+    fn warnings_top_k() {
+        let mut req = basic_request();
+        req.top_k = Some(40);
+        let w = compute_request_warnings(&req);
+        assert_eq!(w.as_header_value().unwrap(), "top_k");
+    }
+
+    #[test]
+    fn warnings_thinking_config() {
+        let mut req = basic_request();
+        req.thinking = Some(anthropic::ThinkingConfig::Enabled { budget_tokens: 5000 });
+        let w = compute_request_warnings(&req);
+        assert_eq!(w.as_header_value().unwrap(), "thinking_config");
+    }
+
+    #[test]
+    fn warnings_stop_sequences_truncated_at_5() {
+        let mut req = basic_request();
+        req.stop_sequences = Some(vec![
+            "a".to_string(),
+            "b".to_string(),
+            "c".to_string(),
+            "d".to_string(),
+            "e".to_string(),
+        ]);
+        let w = compute_request_warnings(&req);
+        assert_eq!(w.as_header_value().unwrap(), "stop_sequences_truncated");
+    }
+
+    #[test]
+    fn warnings_stop_sequences_4_is_fine() {
+        let mut req = basic_request();
+        req.stop_sequences = Some(vec![
+            "a".to_string(),
+            "b".to_string(),
+            "c".to_string(),
+            "d".to_string(),
+        ]);
+        let w = compute_request_warnings(&req);
+        assert!(w.is_empty());
+    }
+
+    #[test]
+    fn warnings_cache_control_on_system() {
+        let mut req = basic_request();
+        req.system = Some(anthropic::System::Blocks(vec![anthropic::SystemBlock {
+            block_type: "text".to_string(),
+            text: "You are helpful.".to_string(),
+            cache_control: Some(anthropic::CacheControl {
+                cache_type: "ephemeral".to_string(),
+            }),
+        }]));
+        let w = compute_request_warnings(&req);
+        assert_eq!(w.as_header_value().unwrap(), "cache_control");
+    }
+
+    #[test]
+    fn warnings_document_blocks() {
+        let mut req = basic_request();
+        req.messages = vec![anthropic::InputMessage {
+            role: anthropic::Role::User,
+            content: anthropic::Content::Blocks(vec![anthropic::ContentBlock::Document {
+                source: anthropic::DocumentSource {
+                    source_type: "base64".to_string(),
+                    media_type: "application/pdf".to_string(),
+                    data: "dGVzdA==".to_string(),
+                },
+                title: None,
+            }]),
+        }];
+        let w = compute_request_warnings(&req);
+        assert_eq!(w.as_header_value().unwrap(), "document_blocks");
+    }
+
+    #[test]
+    fn warnings_multiple_combined() {
+        let mut req = basic_request();
+        req.top_k = Some(10);
+        req.thinking = Some(anthropic::ThinkingConfig::Enabled { budget_tokens: 1000 });
+        let w = compute_request_warnings(&req);
+        let val = w.as_header_value().unwrap();
+        assert!(val.contains("top_k"), "missing top_k in: {val}");
+        assert!(val.contains("thinking_config"), "missing thinking_config in: {val}");
+    }
 }
diff --git a/crates/translator/src/mapping/responses_message_map.rs b/crates/translator/src/mapping/responses_message_map.rs
index 2401f36..2303236 100644
--- a/crates/translator/src/mapping/responses_message_map.rs
+++ b/crates/translator/src/mapping/responses_message_map.rs
@@ -53,8 +53,8 @@ pub fn anthropic_to_responses_request(req: &anthropic::MessageCreateRequest) ->
     }
     if let Some(ref tc) = req.tool_choice {
         let mapped = match tc {
-            anthropic::ToolChoice::Auto => json!("auto"),
-            anthropic::ToolChoice::Any => json!("required"),
+            anthropic::ToolChoice::Auto { .. } => json!("auto"),
+            anthropic::ToolChoice::Any { .. } => json!("required"),
             anthropic::ToolChoice::None => json!("none"),
             anthropic::ToolChoice::Tool { name } => json!({
                 "type": "function",
diff --git a/crates/translator/src/mapping/streaming_map.rs b/crates/translator/src/mapping/streaming_map.rs
index 26bc08b..dfdd021 100644
--- a/crates/translator/src/mapping/streaming_map.rs
+++ b/crates/translator/src/mapping/streaming_map.rs
@@ -80,6 +80,8 @@ impl StreamingTranslator {
         if let Some(ref usage) = chunk.usage {
             self.usage.input_tokens = usage.prompt_tokens;
             self.usage.output_tokens = usage.completion_tokens;
+            self.usage.cache_read_input_tokens =
+                crate::mapping::usage_map::extract_cached_tokens(usage.prompt_tokens_details.as_ref());
         }
 
         for choice in &chunk.choices {
@@ -1199,4 +1201,37 @@ mod tests {
             matches!(&events[0], anthropic::StreamEvent::ContentBlockStop { index } if *index == 0)
         );
     }
+
+    #[test]
+    fn usage_chunk_with_cached_tokens_maps_cache_read() {
+        let mut translator = StreamingTranslator::new("gpt-4o".into());
+        let chunk = ChatCompletionChunk {
+            id: "c1".into(),
+            object: "chat.completion.chunk".into(),
+            model: "gpt-4o".into(),
+            choices: vec![],
+            usage: Some(crate::openai::ChatUsage {
+                prompt_tokens: 100,
+                completion_tokens: 50,
+                total_tokens: 150,
+                completion_tokens_details: None,
+                prompt_tokens_details: Some(serde_json::json!({"cached_tokens": 42})),
+            }),
+            created: None,
+            system_fingerprint: None,
+        };
+        translator.process_chunk(&chunk);
+        let usage = translator.usage().expect("usage should be present");
+        assert_eq!(usage.input_tokens, 100);
+        assert_eq!(usage.output_tokens, 50);
+        assert_eq!(usage.cache_read_input_tokens, Some(42));
+    }
+
+    #[test]
+    fn usage_chunk_without_cached_tokens_leaves_cache_read_none() {
+        let mut translator = StreamingTranslator::new("gpt-4o".into());
+        translator.process_chunk(&usage_chunk("c1", "gpt-4o", 10, 5));
+        let usage = translator.usage().expect("usage should be present");
+        assert!(usage.cache_read_input_tokens.is_none());
+    }
 }
diff --git a/crates/translator/src/mapping/tools_map.rs b/crates/translator/src/mapping/tools_map.rs
index e07ef56..c85b5c6 100644
--- a/crates/translator/src/mapping/tools_map.rs
+++ b/crates/translator/src/mapping/tools_map.rs
@@ -51,10 +51,10 @@ pub fn openai_tools_to_anthropic(tools: &[openai::ChatTool]) -> Vec<anthropic::T
 /// OpenAI: <https://platform.openai.com/docs/api-reference/chat/create>
 pub fn anthropic_tool_choice_to_openai(tc: &anthropic::ToolChoice) -> openai::ChatToolChoice {
     match tc {
-        anthropic::ToolChoice::Auto => openai::ChatToolChoice::Simple("auto".to_string()),
+        anthropic::ToolChoice::Auto { .. } => openai::ChatToolChoice::Simple("auto".to_string()),
         // Any = "model must call at least one tool". OpenAI's "required"
         // is the closest: it forces a tool call when tools are defined.
-        anthropic::ToolChoice::Any => openai::ChatToolChoice::Simple("required".to_string()),
+        anthropic::ToolChoice::Any { .. } => openai::ChatToolChoice::Simple("required".to_string()),
         anthropic::ToolChoice::None => openai::ChatToolChoice::Simple("none".to_string()),
         anthropic::ToolChoice::Tool { name } => {
             openai::ChatToolChoice::Named(openai::chat_completions::NamedToolChoice {
@@ -73,10 +73,10 @@ pub fn openai_tool_choice_to_anthropic(tc: &openai::ChatToolChoice) -> anthropic
     match tc {
         openai::ChatToolChoice::Simple(s) => match s.as_str() {
             "none" => anthropic::ToolChoice::None,
-            "required" => anthropic::ToolChoice::Any,
+            "required" => anthropic::ToolChoice::Any { disable_parallel_tool_use: None },
             // Default unknown values to Auto for forward compatibility;
             // rejecting would break when OpenAI adds new tool_choice variants.
-            _ => anthropic::ToolChoice::Auto,
+            _ => anthropic::ToolChoice::Auto { disable_parallel_tool_use: None },
         },
         openai::ChatToolChoice::Named(named) => anthropic::ToolChoice::Tool {
             name: named.function.name.clone(),
@@ -218,20 +218,20 @@ mod tests {
 
     #[test]
     fn tool_choice_auto() {
-        let openai = anthropic_tool_choice_to_openai(&anthropic::ToolChoice::Auto);
+        let openai = anthropic_tool_choice_to_openai(&anthropic::ToolChoice::Auto { disable_parallel_tool_use: None });
         assert!(matches!(openai, openai::ChatToolChoice::Simple(ref s) if s == "auto"));
 
         let back = openai_tool_choice_to_anthropic(&openai);
-        assert!(matches!(back, anthropic::ToolChoice::Auto));
+        assert!(matches!(back, anthropic::ToolChoice::Auto { .. }));
     }
 
     #[test]
     fn tool_choice_any_to_required() {
-        let openai = anthropic_tool_choice_to_openai(&anthropic::ToolChoice::Any);
+        let openai = anthropic_tool_choice_to_openai(&anthropic::ToolChoice::Any { disable_parallel_tool_use: None });
         assert!(matches!(openai, openai::ChatToolChoice::Simple(ref s) if s == "required"));
 
         let back = openai_tool_choice_to_anthropic(&openai);
-        assert!(matches!(back, anthropic::ToolChoice::Any));
+        assert!(matches!(back, anthropic::ToolChoice::Any { .. }));
     }
 
     #[test]
@@ -270,10 +270,28 @@ mod tests {
         let tc = openai::ChatToolChoice::Simple("something_else".into());
         assert!(matches!(
             openai_tool_choice_to_anthropic(&tc),
-            anthropic::ToolChoice::Auto
+            anthropic::ToolChoice::Auto { .. }
         ));
     }
 
+    #[test]
+    fn disable_parallel_tool_use_roundtrips_via_serde() {
+        // Ensure the field survives JSON deserialization
+        let json = serde_json::json!({"type": "auto", "disable_parallel_tool_use": true});
+        let tc: anthropic::ToolChoice = serde_json::from_value(json).unwrap();
+        match tc {
+            anthropic::ToolChoice::Auto { disable_parallel_tool_use: Some(true) } => {}
+            other => panic!("expected Auto with disable_parallel_tool_use=true, got {:?}", other),
+        }
+    }
+
+    #[test]
+    fn auto_without_disable_parallel_omits_field_in_json() {
+        let tc = anthropic::ToolChoice::Auto { disable_parallel_tool_use: None };
+        let json = serde_json::to_value(&tc).unwrap();
+        assert_eq!(json, serde_json::json!({"type": "auto"}));
+    }
+
     // --- Claude Code tool schema round-trips ---
 
     #[test]
diff --git a/crates/translator/src/mapping/warnings.rs b/crates/translator/src/mapping/warnings.rs
new file mode 100644
index 0000000..064f973
--- /dev/null
+++ b/crates/translator/src/mapping/warnings.rs
@@ -0,0 +1,60 @@
+/// Collects feature degradation notices produced during request translation.
+///
+/// Returned to the proxy layer so it can inject an `x-anyllm-degradation` response
+/// header for clients to inspect. This makes silent drops visible without changing
+/// the Anthropic-compatible response body.
+#[derive(Default, Debug)]
+pub struct TranslationWarnings {
+    items: Vec<&'static str>,
+}
+
+impl TranslationWarnings {
+    pub fn add(&mut self, feature: &'static str) {
+        self.items.push(feature);
+    }
+
+    pub fn is_empty(&self) -> bool {
+        self.items.is_empty()
+    }
+
+    /// Returns a comma-separated string suitable for an HTTP header value,
+    /// or `None` if no features were dropped.
+    pub fn as_header_value(&self) -> Option<String> {
+        if self.items.is_empty() {
+            return None;
+        }
+        Some(self.items.join(", "))
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+
+    #[test]
+    fn empty_returns_none() {
+        let w = TranslationWarnings::default();
+        assert!(w.is_empty());
+        assert!(w.as_header_value().is_none());
+    }
+
+    #[test]
+    fn single_item() {
+        let mut w = TranslationWarnings::default();
+        w.add("top_k");
+        assert!(!w.is_empty());
+        assert_eq!(w.as_header_value().unwrap(), "top_k");
+    }
+
+    #[test]
+    fn multiple_items_comma_separated() {
+        let mut w = TranslationWarnings::default();
+        w.add("top_k");
+        w.add("cache_control");
+        w.add("document_blocks");
+        assert_eq!(
+            w.as_header_value().unwrap(),
+            "top_k, cache_control, document_blocks"
+        );
+    }
+}
diff --git a/docs/COMPARISON_LITELLM.md b/docs/COMPARISON_LITELLM.md
new file mode 100644
index 0000000..92ae84c
--- /dev/null
+++ b/docs/COMPARISON_LITELLM.md
@@ -0,0 +1,147 @@
+# anyllm-proxy vs LiteLLM: Feature Comparison
+
+anyllm-proxy is a specialized **protocol translator** (Anthropic API in, OpenAI-compatible backend out).
+LiteLLM is a broad **AI gateway** focused on enterprise governance, cost control, and routing across 100+ providers.
+These are different categories; not every gap is worth closing.
+
+## Summary Table
+
+| Feature Area | anyllm-proxy | LiteLLM | Gap |
+|---|---|---|---|
+| Protocol translation (Anthropic↔OpenAI) | Full | Partial | **Advantage** |
+| Translation degradation warnings (`x-anyllm-degradation`) | Yes | No | **Advantage** |
+| Local LLM compatibility (system role, synthetic IDs) | Yes | No | **Advantage** |
+| mTLS backend support (PKCS#12) | Yes | No | **Advantage** |
+| Single static binary, no runtime deps | Yes | No | **Advantage** |
+| Provider backends | 7 (OpenAI, Vertex, Gemini, Azure, Bedrock, Anthropic, Responses) | 100+ | Moderate gap |
+| `POST /v1/chat/completions` input | Yes (full, streaming + non-streaming) | Yes | **Parity** |
+| Virtual key management | Yes (SQLite-backed, immediate revocation) | Yes | **Parity** |
+| Per-key rate limiting (RPM/TPM) | Yes (in-memory sliding window) | Yes | **Parity** |
+| OpenTelemetry export | Yes (feature-gated, OTLP/HTTP) | 20+ integrations | Moderate gap |
+| Cost tracking / budget enforcement | No | Yes | Major gap |
+| Response caching | No | Yes | Major gap |
+| Batch processing | Stub (400) | Yes | Major gap |
+| Load balancing / fallback chains | Basic | Full | Moderate gap |
+| Dynamic model management | Partial | Full | Moderate gap |
+| RBAC / OIDC auth | No | Yes | Moderate gap |
+| Audio, image, reranking endpoints | No | Yes | Low (out of scope) |
+
+---
+
+## Detailed Gaps
+
+### 1. Provider/Backend Coverage
+
+**anyllm-proxy:** OpenAI (Chat Completions), OpenAI (Responses API), Vertex AI, Gemini, Azure OpenAI, AWS Bedrock, Anthropic passthrough.
+
+**LiteLLM:** 100+ providers.
+
+Providers that already work today via `OPENAI_BASE_URL` override (OpenAI-compatible):
+- Groq, Together AI, Fireworks, Perplexity, Mistral, HuggingFace TGI, Ollama, vLLM
+
+### 2. API Endpoints
+
+anyllm-proxy accepts both **Anthropic-format** requests (`POST /v1/messages`) and **OpenAI-format** requests (`POST /v1/chat/completions`). The OpenAI endpoint translates internally through the Anthropic pipeline and returns OpenAI-format responses (streaming and non-streaming).
+
+Missing endpoints:
+- `POST /v1/completions` — Legacy text completions
+- `POST /v1/images/generations` — DALL-E, Imagen, etc.
+- `POST /v1/audio/transcriptions` — Whisper / speech-to-text
+- `POST /v1/audio/speech` — TTS
+- `POST /v1/rerank` — Reranking (Cohere, etc.)
+- `POST /v1/messages/batches` — Currently stubbed; returns 400
+
+### 3. Authentication & Authorization
+
+anyllm-proxy supports both static keys (`PROXY_API_KEYS` env var) and dynamic virtual keys managed via the admin API (`POST /admin/api/keys`). Virtual keys are stored in SQLite, cached in memory, and revocation takes effect immediately without restart. Per-key RPM rate limiting is enforced in the auth middleware.
+
+LiteLLM provides:
+- Virtual key issuance via API (`POST /key/generate`) with per-key metadata, expiry, and spend limits
+- Key revocation without restart
+- RBAC roles (admin, developer, read-only)
+- OIDC/JWT validation
+- IP allowlisting
+
+### 4. Load Balancing & Routing
+
+anyllm-proxy supports multiple named backends via `PROXY_CONFIG` YAML, each served under its own path prefix (e.g., `/openai/v1/messages`, `/vertex/v1/messages`). There is no load balancing between multiple instances of the same backend.
+
+LiteLLM supports:
+- Random shuffle, least-busy, latency-based, cost-based, and weighted routing
+- Cross-provider fallback chains (retry on a different provider on failure)
+- Redis-backed distributed state for multi-instance deployments
+
+### 5. Caching
+
+anyllm-proxy has no caching. Every request hits the backend.
+
+LiteLLM supports in-memory, Redis, semantic (Qdrant/Redis), S3, and GCS caches with per-request TTL control.
+
+### 6. Rate Limiting
+
+anyllm-proxy enforces a global concurrency limit (default 100 concurrent requests) and per-key RPM limits via virtual keys (in-memory sliding window, no external dependencies). Upstream 429s are passed through. TPM tracking is recorded per-key for reporting.
+
+LiteLLM enforces RPM and TPM limits per key, user, and team, with Redis-backed distributed tracking.
+
+### 7. Cost Tracking & Budget Management
+
+anyllm-proxy has no cost tracking. No pricing database, no per-request USD calculation, no spend aggregation.
+
+LiteLLM computes per-request USD cost from a built-in model pricing database and aggregates spend per key, user, and team with configurable hard caps.
+
+### 8. Observability & Logging
+
+anyllm-proxy provides:
+- SQLite request log (7-day retention, latency percentiles via admin API)
+- Request count metrics (`GET /metrics`)
+- `x-anyllm-degradation` response header for lossy translation warnings
+- `tracing` crate output (stdout, `RUST_LOG`)
+- Optional OpenTelemetry OTLP export (`--features otel`): spans exported to any OTEL-compatible collector (Datadog, Honeycomb, Jaeger, Tempo, etc.) via `OTEL_EXPORTER_OTLP_ENDPOINT`
+
+LiteLLM integrates with 20+ external observability platforms: Langfuse, Langsmith, OpenTelemetry (Honeycomb, Traceloop, OTEL collectors), Datadog, Sentry, Arize, and others. It also supports structured log export to DynamoDB, S3, GCS, and SQS.
+
+Not present in LiteLLM: `x-anyllm-degradation` per-request degradation signaling.
+
+### 9. Model Management
+
+anyllm-proxy maps Haiku requests to `small_model` and Opus/Sonnet to `big_model`. Overrides persist to SQLite via the admin API.
+
+The `/v1/models` response is a hardcoded list of 13 Claude model IDs with no context window or pricing metadata.
+
+LiteLLM supports dynamic model addition and removal via API without restart, per-model pricing metadata, and enriched `/models` responses with token limits.
+
+### 10. Batch Processing
+
+`POST /v1/messages/batches` is stubbed and returns 400. LiteLLM supports batch processing across multiple providers.
+
+### 11. Database & Persistence
+
+anyllm-proxy uses SQLite for admin config overrides and request logs. There is no schema for virtual keys, users, teams, or spend records.
+
+LiteLLM uses a full relational database (configurable backend) for key/user/team/spend storage.
+
+---
+
+## Completed Items (this release)
+
+1. **`POST /v1/chat/completions`** -- Accept OpenAI-format input (streaming + non-streaming)
+2. **AWS Bedrock backend** -- SigV4 auth, InvokeModel + InvokeModelWithResponseStream
+3. **Azure OpenAI backend** -- Deployment-scoped URLs, `api-key` header auth
+4. **Virtual key management** -- SQLite-backed CRUD, DashMap cache, immediate revocation
+5. **Per-key RPM rate limiting** -- Sliding window enforcement, 429 + retry-after
+6. **OpenTelemetry export** -- Feature-gated OTLP/HTTP, spans to any collector
+7. **Rust client SDK v0.2.0** -- ClientBuilder, ToolBuilder, typed streaming
+
+## Remaining Priority Order
+
+**Tier 1 — Significant, broader scope:**
+1. Response caching (in-memory + Redis) -- Reduces upstream cost and latency for repeated prompts
+2. Cross-provider fallback chains -- Retry on alternate backend on failure
+3. Real batch processing -- Async job queue behind `/v1/messages/batches`
+4. Cost tracking -- Model pricing DB + per-request USD calculation + spend aggregation
+
+**Tier 2 -- Enterprise/niche:**
+5. RBAC / OIDC authentication
+6. Audio, image generation, reranking endpoints
+7. Semantic caching
+8. Budget enforcement and spend alerts
diff --git a/docs/ENV.md b/docs/ENV.md
index c5ee94a..511b14a 100644
--- a/docs/ENV.md
+++ b/docs/ENV.md
@@ -45,6 +45,35 @@ These are the variables most users need.
 | `RUST_LOG` | `info` | Tracing filter. Examples: `debug`, `anyllm_proxy=trace`. |
 | `DISABLE_ADMIN` | (unset) | Set to `1`, `true`, or `yes` to force-disable the admin web interface even when `--webui` is passed. Useful in automated/container environments. |
 
+## AWS Bedrock
+
+Set `BACKEND=bedrock` to route through AWS Bedrock. The proxy sends Anthropic Messages API format directly to Bedrock (no OpenAI translation). Requests are signed with AWS SigV4.
+
+| Variable | Default | Description |
+|----------|---------|-------------|
+| `AWS_REGION` | (required) | AWS region, e.g. `us-east-1`. |
+| `AWS_ACCESS_KEY_ID` | (required) | AWS access key ID for SigV4 signing. |
+| `AWS_SECRET_ACCESS_KEY` | (required) | AWS secret access key for SigV4 signing. |
+| `AWS_SESSION_TOKEN` | (optional) | Temporary session token for STS credentials. |
+| `BIG_MODEL` | `anthropic.claude-sonnet-4-20250514-v1:0` | Bedrock model ID for sonnet/opus requests. |
+| `SMALL_MODEL` | `anthropic.claude-haiku-4-5-20251001-v1:0` | Bedrock model ID for haiku requests. |
+
+### Example
+
+```bash
+BACKEND=bedrock \
+AWS_REGION=us-east-1 \
+AWS_ACCESS_KEY_ID=AKIA... \
+AWS_SECRET_ACCESS_KEY=wJalr... \
+cargo run -p anyllm_proxy
+```
+
+### Streaming
+
+Bedrock streaming uses AWS Event Stream binary framing instead of SSE. The proxy decodes Event Stream frames and re-emits them as standard SSE events, so downstream clients see the same Anthropic SSE format as with other backends.
+
+---
+
 ## Azure OpenAI
 
 Set `BACKEND=azure` to route through Azure OpenAI Service. The request/response format is identical to standard OpenAI Chat Completions; only the URL scheme and auth header differ.
@@ -148,3 +177,24 @@ ADMIN_DB_PATH=/var/lib/anyllm/admin.db \
 anyllm_proxy --webui
 # Open: http://127.0.0.1:4000/admin/?token=my-secret-token
 ```
+
+---
+
+## OpenTelemetry (optional)
+
+Trace export is opt-in. Build with the `otel` cargo feature to enable it:
+
+```bash
+cargo build -p anyllm_proxy --features otel
+```
+
+When the feature is enabled, the proxy initializes an OTLP span exporter that sends traces over HTTP/protobuf. The OTLP SDK reads configuration from standard environment variables; no proxy-specific config is needed.
+
+| Variable | Default | Description |
+|----------|---------|-------------|
+| `OTEL_EXPORTER_OTLP_ENDPOINT` | `http://localhost:4318` | OTLP collector endpoint (HTTP). |
+| `OTEL_SERVICE_NAME` | `unknown_service` | Service name attached to all exported spans. Set this to `anyllm-proxy` or your deployment name. |
+| `OTEL_TRACES_SAMPLER` | `parentbased_always_on` | Sampling strategy. Common values: `always_on`, `always_off`, `traceidratio` (pair with `OTEL_TRACES_SAMPLER_ARG`). |
+| `OTEL_TRACES_SAMPLER_ARG` | (none) | Argument for the sampler, e.g. `0.1` for 10% sampling with `traceidratio`. |
+
+When built without the `otel` feature (the default), none of these variables have any effect and there is zero runtime overhead.
diff --git a/specs/20260325-120000-litellm-gap-fill/tasks.md b/specs/20260325-120000-litellm-gap-fill/tasks.md
index 22d5ba3..dbd4ea2 100644
--- a/specs/20260325-120000-litellm-gap-fill/tasks.md
+++ b/specs/20260325-120000-litellm-gap-fill/tasks.md
@@ -60,7 +60,7 @@
 - [x] T014 [US1] Create `crates/proxy/src/server/chat_completions.rs` with non-streaming handler: extract `Json<ChatCompletionRequest>`, call `openai_to_anthropic_request`, dispatch to `BackendClient`, call `anthropic_to_openai_response`, return `Json<ChatCompletionResponse>`. Set `x-anyllm-degradation` header from `TranslationWarnings`. Return OpenAI-shaped errors on validation failure (missing max_tokens -> 400 `invalid_request_error`).
 - [x] T015 [US1] Add streaming handler in `crates/proxy/src/server/chat_completions.rs`: when `stream: true`, dispatch to backend streaming path, create `ReverseStreamingTranslator`, emit `text/event-stream` with `data: {chunk}\n\n` lines (no `event:` prefix, matching OpenAI SSE format). Terminate with `data: [DONE]\n\n`.
 - [x] T016 [US1] Register `POST /v1/chat/completions` route in `crates/proxy/src/server/routes.rs` on the existing backend router. Apply same middleware (auth, request ID, size limit, concurrency limit) as the `/v1/messages` route.
-- [ ] T017 [US1] Add integration tests in `crates/proxy/tests/` (new file `chat_completions.rs` or extend existing): non-streaming basic response, streaming basic response, tool call round-trip, missing max_tokens returns 400, degradation header set for lossy fields, empty messages returns 400.
+- [x] T017 [US1] Add integration tests in `crates/proxy/tests/` (new file `chat_completions.rs` or extend existing): non-streaming basic response, streaming basic response, tool call round-trip, missing max_tokens returns 400, degradation header set for lossy fields, empty messages returns 400.
 
 **Checkpoint**: `cargo test -p anyllm_proxy` passes. `POST /v1/chat/completions` works end-to-end with a mock or live backend.
 
@@ -74,11 +74,11 @@
 
 ### Implementation for User Story 2
 
-- [ ] T018 [US2] Add `BackendKind::AzureOpenAI` variant to the backend enum in `crates/proxy/src/config/mod.rs`. Parse `AZURE_OPENAI_API_KEY`, `AZURE_OPENAI_ENDPOINT`, `AZURE_OPENAI_DEPLOYMENT`, `AZURE_OPENAI_API_VERSION` (default `"2024-10-21"`) from env. Construct full URL: `{endpoint}/openai/deployments/{deployment}/chat/completions?api-version={version}` at config load time. Validate URL with existing `validate_url` function.
-- [ ] T019 [US2] Add `BackendAuth::AzureApiKey(String)` variant (or equivalent) in `crates/proxy/src/backend/mod.rs`. Map it to `RequestAuth::Header { name: "api-key", value: key }` in the auth application logic. Add `BackendClient::AzureOpenAI(OpenAIClient)` variant that constructs `OpenAIClient` with the pre-built Azure URL and `AzureApiKey` auth.
-- [ ] T020 [US2] Modify `crates/proxy/src/backend/openai_client.rs` to accept Azure's pre-constructed URL. The `chat_completions_url` for Azure is the full URL from config (no `/v1/chat/completions` suffix appended). Ensure the `model` field in the request body is still populated (Azure ignores it but accepts it).
-- [ ] T021 [US2] Add `#[ignore]` integration test in `crates/proxy/tests/` for Azure backend: send a request via the proxy configured with `BACKEND=azure`, verify response is valid Anthropic format. Requires `AZURE_OPENAI_API_KEY` env var to run.
-- [ ] T022 [US2] Update `docs/ENV.md` with Azure-specific env vars and usage example.
+- [x] T018 [US2] Add `BackendKind::AzureOpenAI` variant to the backend enum in `crates/proxy/src/config/mod.rs`. Parse `AZURE_OPENAI_API_KEY`, `AZURE_OPENAI_ENDPOINT`, `AZURE_OPENAI_DEPLOYMENT`, `AZURE_OPENAI_API_VERSION` (default `"2024-10-21"`) from env. Construct full URL: `{endpoint}/openai/deployments/{deployment}/chat/completions?api-version={version}` at config load time. Validate URL with existing `validate_url` function.
+- [x] T019 [US2] Add `BackendAuth::AzureApiKey(String)` variant (or equivalent) in `crates/proxy/src/backend/mod.rs`. Map it to `RequestAuth::Header { name: "api-key", value: key }` in the auth application logic. Add `BackendClient::AzureOpenAI(OpenAIClient)` variant that constructs `OpenAIClient` with the pre-built Azure URL and `AzureApiKey` auth.
+- [x] T020 [US2] Modify `crates/proxy/src/backend/openai_client.rs` to accept Azure's pre-constructed URL. The `chat_completions_url` for Azure is the full URL from config (no `/v1/chat/completions` suffix appended). Ensure the `model` field in the request body is still populated (Azure ignores it but accepts it).
+- [x] T021 [US2] Add `#[ignore]` integration test in `crates/proxy/tests/` for Azure backend: send a request via the proxy configured with `BACKEND=azure`, verify response is valid Anthropic format. Requires `AZURE_OPENAI_API_KEY` env var to run.
+- [x] T022 [US2] Update `docs/ENV.md` with Azure-specific env vars and usage example.
 
 **Checkpoint**: `cargo build` clean. Azure config parsing tested. `#[ignore]` live test exists.
 
@@ -134,13 +134,13 @@
 
 ### Implementation for User Story 5
 
-- [ ] T039 [P] [US5] Add `ClientBuilder` to `crates/client/src/client.rs` with method chaining: `fn new() -> Self`, `fn base_url(mut self, url: &str) -> Self`, `fn api_key(mut self, key: &str) -> Self`, `fn timeout(mut self, d: Duration) -> Self`, `fn read_timeout(mut self, d: Duration) -> Self`, `fn max_retries(mut self, n: u32) -> Self`, `fn tls_config(mut self, cfg: TlsConfig) -> Self`, `fn build(self) -> Result<Client, ClientError>`. Implement `Client::builder() -> ClientBuilder` convenience method.
-- [ ] T040 [P] [US5] Create `crates/client/src/tools.rs` with `ToolBuilder` and `ToolChoiceBuilder`. `ToolBuilder`: `fn new(name: &str) -> Self`, `fn description(mut self, desc: &str) -> Self`, `fn input_schema(mut self, schema: Value) -> Self`, `fn build(self) -> Tool`. `ToolChoiceBuilder`: `fn auto() -> ToolChoice`, `fn any() -> ToolChoice`, `fn none() -> ToolChoice`, `fn specific(name: &str) -> ToolChoice`.
-- [ ] T041 [US5] Add streaming return type to `crates/client/src/client.rs`: `fn messages_stream(&self, req: MessageCreateRequest) -> Result<impl Stream<Item = Result<StreamEvent, ClientError>>, ClientError>`. Parse SSE frames from the reqwest response byte stream, deserialize each `data:` line into `StreamEvent`.
-- [ ] T042 [US5] Update `crates/client/src/lib.rs` to re-export all public types: `Client`, `ClientBuilder`, `ClientConfig`, `ClientError`, `Tool`, `ToolBuilder`, `ToolChoice`, `ToolChoiceBuilder`, `StreamEvent`, and all Anthropic request/response types from `anyllm_translate`.
-- [ ] T043 [US5] Add rustdoc examples to all public types and methods in `crates/client/src/client.rs`, `crates/client/src/tools.rs`, and `crates/client/src/lib.rs`. Each builder method and each public function gets a `/// # Examples` block.
-- [ ] T044 [US5] Bump `anyllm_client` version to `0.2.0` in `crates/client/Cargo.toml`.
-- [ ] T045 [US5] Add unit tests for `ClientBuilder` (valid build, missing required fields), `ToolBuilder`, and `ToolChoiceBuilder` in their respective `#[cfg(test)]` modules.
+- [x] T039 [P] [US5] Add `ClientBuilder` to `crates/client/src/client.rs` with method chaining: `fn new() -> Self`, `fn base_url(mut self, url: &str) -> Self`, `fn api_key(mut self, key: &str) -> Self`, `fn timeout(mut self, d: Duration) -> Self`, `fn read_timeout(mut self, d: Duration) -> Self`, `fn max_retries(mut self, n: u32) -> Self`, `fn tls_config(mut self, cfg: TlsConfig) -> Self`, `fn build(self) -> Result<Client, ClientError>`. Implement `Client::builder() -> ClientBuilder` convenience method.
+- [x] T040 [P] [US5] Create `crates/client/src/tools.rs` with `ToolBuilder` and `ToolChoiceBuilder`. `ToolBuilder`: `fn new(name: &str) -> Self`, `fn description(mut self, desc: &str) -> Self`, `fn input_schema(mut self, schema: Value) -> Self`, `fn build(self) -> Tool`. `ToolChoiceBuilder`: `fn auto() -> ToolChoice`, `fn any() -> ToolChoice`, `fn none() -> ToolChoice`, `fn specific(name: &str) -> ToolChoice`.
+- [x] T041 [US5] Add streaming return type to `crates/client/src/client.rs`: `fn messages_stream(&self, req: MessageCreateRequest) -> Result<impl Stream<Item = Result<StreamEvent, ClientError>>, ClientError>`. Parse SSE frames from the reqwest response byte stream, deserialize each `data:` line into `StreamEvent`.
+- [x] T042 [US5] Update `crates/client/src/lib.rs` to re-export all public types: `Client`, `ClientBuilder`, `ClientConfig`, `ClientError`, `Tool`, `ToolBuilder`, `ToolChoice`, `ToolChoiceBuilder`, `StreamEvent`, and all Anthropic request/response types from `anyllm_translate`.
+- [x] T043 [US5] Add rustdoc examples to all public types and methods in `crates/client/src/client.rs`, `crates/client/src/tools.rs`, and `crates/client/src/lib.rs`. Each builder method and each public function gets a `/// # Examples` block.
+- [x] T044 [US5] Bump `anyllm_client` version to `0.2.0` in `crates/client/Cargo.toml`.
+- [x] T045 [US5] Add unit tests for `ClientBuilder` (valid build, missing required fields), `ToolBuilder`, and `ToolChoiceBuilder` in their respective `#[cfg(test)]` modules.
 
 **Checkpoint**: `cargo doc -p anyllm_client --no-deps` builds clean. `cargo test -p anyllm_client` passes.
 
@@ -187,7 +187,7 @@
 
 **Purpose**: Final validation, documentation updates, and CI adjustments
 
-- [ ] T057 [P] Update `docs/COMPARISON_LITELLM.md` to reflect closed gaps: `POST /v1/chat/completions` input, Bedrock backend, Azure backend, virtual key management, per-key rate limiting, OTEL export. Move items from "Major gap" to "Advantage" or "Parity" as appropriate.
+- [x] T057 [P] Update `docs/COMPARISON_LITELLM.md` to reflect closed gaps: `POST /v1/chat/completions` input, Bedrock backend, Azure backend, virtual key management, per-key rate limiting, OTEL export. Move items from "Major gap" to "Advantage" or "Parity" as appropriate.
 - [ ] T058 [P] Update `CLAUDE.md` with new backend types, new env vars, new admin endpoints, new source files, and updated test counts.
 - [ ] T059 [P] Update `README.md` with quickstart examples for new features (reference `quickstart.md` content).
 - [ ] T060 Run `cargo clippy -- -D warnings` across all crates and fix any warnings.