diff --git a/CHANGELOG.md b/CHANGELOG.md
index 8a125ff26..6ddf2b4e1 100644
--- a/CHANGELOG.md
+++ b/CHANGELOG.md
@@ -9,14 +9,33 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0
 
 ## Recent Releases
 
+**v0.1.72 (April 3, 2026)** - Grok Backend Update & Circuit Breaker Phase 2
+Grok backend update with latest improvements. LLM API circuit breaker extended to ChatCompletions, Response API, and Gemini backends (was Claude-only). Config plumbing smoke tests for all backends.
+
 **v0.1.71 (April 1, 2026)** - Trace Memory & Evaluation Polish
 Trace analyzer subagents now launch in the background after each round to write insights from execution traces into memory. Improved evaluation criteria generation and system prompt tuning. Fixes for final injection, eval criteria GPT pre-collab, trace analyzer launch, and trace memory.
 
 **v0.1.70 (March 30, 2026)** - Evaluation Criteria Redesign
 Redesigned three-tier evaluation criteria with anti-pattern definitions and aspiration statements. Improved checklist-gated evaluation with tighter iterative submission cycles. Fast iteration mode, WebUI review modal, and background trace analysis from round 2.
 
-**v0.1.69 (March 27, 2026)** - WebUI Automation & Improved Skill
-WebUI automation now auto-starts without browser interaction — open the URL at any point mid-run to monitor progress. MassGen skill redesign for increased usability and WebUI integration. Quickstart Wizard rework, Workspace Browser expansion, and flexible evaluation criteria field names.
+---
+
+## [0.1.72] - 2026-04-03
+
+### Changed
+- **Grok Backend Update** ([#1044](https://github.com/massgen/MassGen/pull/1044)): Updated Grok backend with latest improvements
+
+### Added
+- **Circuit Breaker Phase 2** ([#1038](https://github.com/massgen/MassGen/pull/1038)): LLM API circuit breaker extended to ChatCompletions, Response API, and Gemini backends (was Claude-only in v0.1.68); Gemini also handles 503 errors
+- **Config Plumbing Smoke Tests** ([#1038](https://github.com/massgen/MassGen/pull/1038)): Smoke tests verify circuit breaker wiring and API call timing for all backends
+
+### Fixed
+- **Response API Timing** ([#1038](https://github.com/massgen/MassGen/pull/1038)): Added start/end API call timing to ResponseBackend non-MCP path
+
+### Technical Details
+- **Major Focus**: Circuit Breaker Phase 2 — rate limit protection across all major backends
+- **PRs Merged**: [#1038](https://github.com/massgen/MassGen/pull/1038), [#1044](https://github.com/massgen/MassGen/pull/1044)
+- **Contributors**: @amabito, @HenryQi, @ncrispino and the MassGen team
 
 ---
 
diff --git a/CONTRIBUTING.md b/CONTRIBUTING.md
index 5ab07f6d8..b861e9251 100644
--- a/CONTRIBUTING.md
+++ b/CONTRIBUTING.md
@@ -359,7 +359,7 @@ Create a `.env` file in the `massgen` directory as described in [README](README.
 
 ## 🔧 Development Workflow
 
-> **Important**: Our next version is v0.1.72. If you want to contribute, please contribute to the `dev/v0.1.72` branch (or `main` if dev/v0.1.72 doesn't exist yet).
+> **Important**: Our next version is v0.1.73. If you want to contribute, please contribute to the `dev/v0.1.73` branch (or `main` if dev/v0.1.73 doesn't exist yet).
 
 ### 1. Create Feature Branch
 
@@ -368,7 +368,7 @@ Create a `.env` file in the `massgen` directory as described in [README](README.
 git fetch upstream
 
 # Create feature branch from dev/v0.1.60 (or main if dev branch doesn't exist yet)
-git checkout -b feature/your-feature-name upstream/dev/v0.1.72
+git checkout -b feature/your-feature-name upstream/dev/v0.1.73
 ```
 
 ### 2. Make Your Changes
@@ -507,7 +507,7 @@ git push origin feature/your-feature-name
 ```
 
 Then create a pull request on GitHub:
-- Base branch: `dev/v0.1.72` (or `main` if dev branch doesn't exist yet)
+- Base branch: `dev/v0.1.73` (or `main` if dev branch doesn't exist yet)
 - Compare branch: `feature/your-feature-name`
 - Add clear description of changes
 - Link any related issues
@@ -617,7 +617,7 @@ Have a significant feature idea not covered by existing tracks?
 - [ ] Tests pass locally
 - [ ] Documentation is updated if needed
 - [ ] Commit messages follow convention
-- [ ] PR targets `dev/v0.1.72` branch (or `main` if dev branch doesn't exist yet)
+- [ ] PR targets `dev/v0.1.73` branch (or `main` if dev branch doesn't exist yet)
 
 ### PR Description Should Include
 
diff --git a/README.md b/README.md
index 90c62458c..a8eeb3089 100644
--- a/README.md
+++ b/README.md
@@ -69,7 +69,7 @@ This project started with the "threads of thought" and "iterative refinement" id
 <details open>
 <summary><h3>🆕 Latest Features</h3></summary>
 
-- [v0.1.71 Features](#-latest-features-v0171)
+- [v0.1.72 Features](#-latest-features-v0172)
 </details>
 
 <details open>
@@ -122,15 +122,15 @@ This project started with the "threads of thought" and "iterative refinement" id
 <details open>
 <summary><h3>🗺️ Roadmap</h3></summary>
 
-- [Recent Achievements (v0.1.71)](#recent-achievements-v0171)
-- [Previous Achievements (v0.0.3 - v0.1.70)](#previous-achievements-v003---v0170)
+- [Recent Achievements (v0.1.72)](#recent-achievements-v0172)
+- [Previous Achievements (v0.0.3 - v0.1.71)](#previous-achievements-v003---v0171)
 - [Key Future Enhancements](#key-future-enhancements)
   - Bug Fixes & Backend Improvements
   - Advanced Agent Collaboration
   - Expanded Model, Tool & Agent Integrations
   - Improved Performance & Scalability
   - Enhanced Developer Experience
-- [v0.1.72 Roadmap](#v0172-roadmap)
+- [v0.1.73 Roadmap](#v0173-roadmap)
 </details>
 
 <details open>
@@ -155,20 +155,19 @@ This project started with the "threads of thought" and "iterative refinement" id
 
 ---
 
-## 🆕 Latest Features (v0.1.71)
+## 🆕 Latest Features (v0.1.72)
 
-**🎉 Released: April 1, 2026**
+**🎉 Released: April 3, 2026**
 
-**What's New in v0.1.71:**
-- **🔍 Trace Analyzer Subagents** - Launch in the background after each round to write insights from execution traces into memory.
-- **📋 Better Evaluation Criteria** - Improved criteria generation for higher-quality, more opinionated output.
-- **🧠 System Prompt Tuning** - Adjusted system prompts for better agent performance across coordination rounds.
-- **🔧 Stability Fixes** - Fixed final injection, eval criteria GPT pre-collab, trace analyzer launch, and memory handling.
+**What's New in v0.1.72:**
+- **🦎 Grok Backend Update** - Updated Grok backend with latest improvements.
+- **⚡ Circuit Breaker Phase 2** - LLM API circuit breaker extended to ChatCompletions, Response API, and Gemini backends (was Claude-only).
+- **🧪 Config Plumbing Smoke Tests** - Verify circuit breaker wiring for all backends.
 
-**Try v0.1.71 Features:**
+**Try v0.1.72 Features:**
 ```bash
-pip install massgen==0.1.71
-uv run massgen --config @examples/features/trace_analyzer_background.yaml "Create an svg of an AI agent coding."
+pip install massgen==0.1.72
+uv run massgen --config @examples/providers/others/grok_x_search.yaml "Research the latest posts and news about AI agents in the last week, and summarize the key trends and insights."
 ```
 
 → [See full release history and examples](massgen/configs/README.md#release-history--examples)
@@ -1240,17 +1239,18 @@ MassGen is currently in its foundational stage, with a focus on parallel, asynch
 
 ⚠️ **Early Stage Notice:** As MassGen is in active development, please expect upcoming breaking architecture changes as we continue to refine and improve the system.
 
-### Recent Achievements (v0.1.71)
+### Recent Achievements (v0.1.72)
 
-**🎉 Released: April 1, 2026**
+**🎉 Released: April 3, 2026**
 
-#### Trace Memory & Evaluation Polish
-- **Trace Analyzer Subagents**: Background trace analysis after each round — writes insights from execution traces into memory for next-round continuity
-- **Better Evaluation Criteria**: Improved criteria generation for higher-quality, more opinionated output
-- **System Prompt Tuning**: Adjusted system prompts for better agent performance across coordination rounds
-- **Stability Fixes**: Fixed final injection, eval criteria GPT pre-collab, trace analyzer launch, trace memory, and auto round memory
+#### Grok Backend Update & Circuit Breaker Phase 2
+- **Grok Backend Update** ([#1044](https://github.com/massgen/MassGen/pull/1044)): Updated Grok backend with latest improvements
+- **Circuit Breaker Phase 2** ([#1038](https://github.com/massgen/MassGen/pull/1038)): LLM API circuit breaker extended to ChatCompletions, Response API, and Gemini backends (was Claude-only); Gemini also handles 503
+- **Config Plumbing Smoke Tests** ([#1038](https://github.com/massgen/MassGen/pull/1038)): Verify circuit breaker wiring for all backends
 
-### Previous Achievements (v0.0.3 - v0.1.70)
+### Previous Achievements (v0.0.3 - v0.1.71)
+
+✅ **Trace Memory & Evaluation Polish (v0.1.71)**: Trace analyzer subagents launch in background after each round to write insights from execution traces into memory. Improved evaluation criteria generation and system prompt tuning.
 
 ✅ **Evaluation Criteria Redesign (v0.1.70)**: Redesigned three-tier evaluation criteria with anti-pattern definitions and aspiration statements. Improved checklist-gated evaluation. Fast iteration mode, WebUI review modal, and background trace analysis.
 
@@ -1537,9 +1537,9 @@ MassGen is currently in its foundational stage, with a focus on parallel, asynch
 
 We welcome community contributions to achieve these goals.
 
-### v0.1.72 Roadmap
+### v0.1.73 Roadmap
 
-Version 0.1.72 focuses on cloud execution:
+Version 0.1.73 focuses on cloud execution:
 
 #### Planned Features
 - **Cloud Modal MVP** ([#982](https://github.com/massgen/MassGen/issues/982)): Run MassGen as a cloud job on Modal — progress streams to terminal, results saved locally under `.massgen/cloud_jobs/`
diff --git a/README_PYPI.md b/README_PYPI.md
index edb4a1220..5b342430f 100644
--- a/README_PYPI.md
+++ b/README_PYPI.md
@@ -68,7 +68,7 @@ This project started with the "threads of thought" and "iterative refinement" id
 <details open>
 <summary><h3>🆕 Latest Features</h3></summary>
 
-- [v0.1.71 Features](#-latest-features-v0171)
+- [v0.1.72 Features](#-latest-features-v0172)
 </details>
 
 <details open>
@@ -121,15 +121,15 @@ This project started with the "threads of thought" and "iterative refinement" id
 <details open>
 <summary><h3>🗺️ Roadmap</h3></summary>
 
-- [Recent Achievements (v0.1.71)](#recent-achievements-v0171)
-- [Previous Achievements (v0.0.3 - v0.1.70)](#previous-achievements-v003---v0170)
+- [Recent Achievements (v0.1.72)](#recent-achievements-v0172)
+- [Previous Achievements (v0.0.3 - v0.1.71)](#previous-achievements-v003---v0171)
 - [Key Future Enhancements](#key-future-enhancements)
   - Bug Fixes & Backend Improvements
   - Advanced Agent Collaboration
   - Expanded Model, Tool & Agent Integrations
   - Improved Performance & Scalability
   - Enhanced Developer Experience
-- [v0.1.72 Roadmap](#v0172-roadmap)
+- [v0.1.73 Roadmap](#v0173-roadmap)
 </details>
 
 <details open>
@@ -154,20 +154,19 @@ This project started with the "threads of thought" and "iterative refinement" id
 
 ---
 
-## 🆕 Latest Features (v0.1.71)
+## 🆕 Latest Features (v0.1.72)
 
-**🎉 Released: April 1, 2026**
+**🎉 Released: April 3, 2026**
 
-**What's New in v0.1.71:**
-- **🔍 Trace Analyzer Subagents** - Launch in the background after each round to write insights from execution traces into memory.
-- **📋 Better Evaluation Criteria** - Improved criteria generation for higher-quality, more opinionated output.
-- **🧠 System Prompt Tuning** - Adjusted system prompts for better agent performance across coordination rounds.
-- **🔧 Stability Fixes** - Fixed final injection, eval criteria GPT pre-collab, trace analyzer launch, and memory handling.
+**What's New in v0.1.72:**
+- **🦎 Grok Backend Update** - Updated Grok backend with latest improvements.
+- **⚡ Circuit Breaker Phase 2** - LLM API circuit breaker extended to ChatCompletions, Response API, and Gemini backends (was Claude-only).
+- **🧪 Config Plumbing Smoke Tests** - Verify circuit breaker wiring for all backends.
 
-**Try v0.1.71 Features:**
+**Try v0.1.72 Features:**
 ```bash
-pip install massgen==0.1.71
-uv run massgen --config @examples/features/trace_analyzer_background.yaml "Create an svg of an AI agent coding."
+pip install massgen==0.1.72
+uv run massgen --config @examples/providers/others/grok_x_search.yaml "Research the latest posts and news about AI agents in the last week, and summarize the key trends and insights."
 ```
 
 → [See full release history and examples](massgen/configs/README.md#release-history--examples)
@@ -1239,17 +1238,18 @@ MassGen is currently in its foundational stage, with a focus on parallel, asynch
 
 ⚠️ **Early Stage Notice:** As MassGen is in active development, please expect upcoming breaking architecture changes as we continue to refine and improve the system.
 
-### Recent Achievements (v0.1.71)
+### Recent Achievements (v0.1.72)
 
-**🎉 Released: April 1, 2026**
+**🎉 Released: April 3, 2026**
 
-#### Trace Memory & Evaluation Polish
-- **Trace Analyzer Subagents**: Background trace analysis after each round — writes insights from execution traces into memory for next-round continuity
-- **Better Evaluation Criteria**: Improved criteria generation for higher-quality, more opinionated output
-- **System Prompt Tuning**: Adjusted system prompts for better agent performance across coordination rounds
-- **Stability Fixes**: Fixed final injection, eval criteria GPT pre-collab, trace analyzer launch, trace memory, and auto round memory
+#### Grok Backend Update & Circuit Breaker Phase 2
+- **Grok Backend Update** ([#1044](https://github.com/massgen/MassGen/pull/1044)): Updated Grok backend with latest improvements
+- **Circuit Breaker Phase 2** ([#1038](https://github.com/massgen/MassGen/pull/1038)): LLM API circuit breaker extended to ChatCompletions, Response API, and Gemini backends (was Claude-only); Gemini also handles 503
+- **Config Plumbing Smoke Tests** ([#1038](https://github.com/massgen/MassGen/pull/1038)): Verify circuit breaker wiring for all backends
 
-### Previous Achievements (v0.0.3 - v0.1.70)
+### Previous Achievements (v0.0.3 - v0.1.71)
+
+✅ **Trace Memory & Evaluation Polish (v0.1.71)**: Trace analyzer subagents launch in background after each round to write insights from execution traces into memory. Improved evaluation criteria generation and system prompt tuning.
 
 ✅ **Evaluation Criteria Redesign (v0.1.70)**: Redesigned three-tier evaluation criteria with anti-pattern definitions and aspiration statements. Improved checklist-gated evaluation. Fast iteration mode, WebUI review modal, and background trace analysis.
 
@@ -1536,9 +1536,9 @@ MassGen is currently in its foundational stage, with a focus on parallel, asynch
 
 We welcome community contributions to achieve these goals.
 
-### v0.1.72 Roadmap
+### v0.1.73 Roadmap
 
-Version 0.1.72 focuses on cloud execution:
+Version 0.1.73 focuses on cloud execution:
 
 #### Planned Features
 - **Cloud Modal MVP** ([#982](https://github.com/massgen/MassGen/issues/982)): Run MassGen as a cloud job on Modal — progress streams to terminal, results saved locally under `.massgen/cloud_jobs/`
diff --git a/ROADMAP.md b/ROADMAP.md
index c93ae1ba2..ee2bc4571 100644
--- a/ROADMAP.md
+++ b/ROADMAP.md
@@ -1,10 +1,10 @@
 # MassGen Roadmap
 
-**Current Version:** v0.1.71
+**Current Version:** v0.1.72
 
 **Release Schedule:** Mondays, Wednesdays, Fridays @ 9am PT
 
-**Last Updated:** April 1, 2026
+**Last Updated:** April 3, 2026
 
 This roadmap outlines MassGen's development priorities for upcoming releases. Each release focuses on specific capabilities with real-world use cases.
 
@@ -42,40 +42,26 @@ Want to contribute or collaborate on a specific track? Reach out to the track ow
 
 | Release | Target | Feature | Owner | Use Case |
 |---------|--------|---------|-------|----------|
-| **v0.1.72** | 04/04/26 | Cloud Modal MVP | @ncrispino | Run MassGen as a cloud job on Modal ([#982](https://github.com/massgen/MassGen/issues/982)) |
-| **v0.1.73** | 04/07/26 | OpenAI Audio API | @ncrispino | Support OpenAI audio API for audio understanding ([#960](https://github.com/massgen/MassGen/issues/960)) |
-| **v0.1.74** | 04/09/26 | Image/Video Edit Capabilities | @ncrispino | Check and support img/video editing capabilities ([#959](https://github.com/massgen/MassGen/issues/959)) |
+| **v0.1.73** | 04/07/26 | Cloud Modal MVP | @ncrispino | Run MassGen as a cloud job on Modal ([#982](https://github.com/massgen/MassGen/issues/982)) |
+| **v0.1.74** | 04/09/26 | OpenAI Audio API | @ncrispino | Support OpenAI audio API for audio understanding ([#960](https://github.com/massgen/MassGen/issues/960)) |
+| **v0.1.75** | 04/11/26 | Image/Video Edit Capabilities | @ncrispino | Check and support img/video editing capabilities ([#959](https://github.com/massgen/MassGen/issues/959)) |
 
 *All releases ship on MWF @ 9am PT when ready*
 
 ---
 
-## ✅ v0.1.71 - Trace Memory & Evaluation Polish (Completed)
+## ✅ v0.1.72 - Grok Backend Update & Circuit Breaker Phase 2 (Completed)
 
-**Released:** April 1, 2026
+**Released:** April 3, 2026 | PRs: [#1038](https://github.com/massgen/MassGen/pull/1038), [#1044](https://github.com/massgen/MassGen/pull/1044)
 
 ### Features
-- **Trace Analyzer Subagents**: Background trace analysis after each round — writes insights from execution traces into memory for next-round continuity
-- **Better Evaluation Criteria**: Improved criteria generation for higher-quality, more opinionated output
-- **System Prompt Tuning**: Adjusted system prompts for better agent performance across coordination rounds
-- **Stability Fixes**: Fixed final injection, eval criteria GPT pre-collab, trace analyzer launch, trace memory, and auto round memory
+- **Grok Backend Update**: Updated Grok backend with latest improvements
+- **Circuit Breaker Phase 2**: LLM API circuit breaker extended to ChatCompletions, Response API, and Gemini backends (was Claude-only); Gemini also handles 503
+- **Config Plumbing Smoke Tests**: Verify circuit breaker wiring for all backends
 
 ---
 
-## ✅ v0.1.70 - Evaluation Criteria Redesign (Completed)
-
-**Released:** March 30, 2026 | PRs: [#1035](https://github.com/massgen/MassGen/pull/1035)
-
-### Features
-- **Evaluation Criteria Redesign**: Three-tier categorization (`primary`, `standard`, `stretch`) with anti-pattern definitions and aspiration statements
-- **Improved Checklist-Gated Evaluation**: Tighter iterative submission cycles with improved scoring and improvement proposals before final voting
-- **Fast Iteration Mode**: Streamlined multi-round submission phases via `fast_iteration.yaml`
-- **WebUI Review Modal**: Approve and comment on outputs in the browser when working in git
-- **Background Trace Analysis**: Execution trace analyzer starts automatically from round 2
-
----
-
-## 📋 v0.1.72 - Cloud Modal MVP
+## 📋 v0.1.73 - Cloud Modal MVP
 
 ### Features
 
@@ -91,7 +77,7 @@ Want to contribute or collaborate on a specific track? Reach out to the track ow
 
 ---
 
-## 📋 v0.1.73 - OpenAI Audio API
+## 📋 v0.1.74 - OpenAI Audio API
 
 ### Features
 
@@ -107,7 +93,7 @@ Want to contribute or collaborate on a specific track? Reach out to the track ow
 
 ---
 
-## 📋 v0.1.74 - Image/Video Edit Capabilities
+## 📋 v0.1.75 - Image/Video Edit Capabilities
 
 ### Features
 
diff --git a/ROADMAP_v0.1.72.md b/ROADMAP_v0.1.73.md
similarity index 54%
rename from ROADMAP_v0.1.72.md
rename to ROADMAP_v0.1.73.md
index 66022ad01..99bf62fca 100644
--- a/ROADMAP_v0.1.72.md
+++ b/ROADMAP_v0.1.73.md
@@ -1,10 +1,10 @@
-# MassGen v0.1.72 Roadmap
+# MassGen v0.1.73 Roadmap
 
-**Target Release:** April 4, 2026
+**Target Release:** April 7, 2026
 
 ## Overview
 
-Version 0.1.72 focuses on running MassGen as a cloud job on Modal.
+Version 0.1.73 focuses on running MassGen as a cloud job on Modal.
 
 ---
 
@@ -27,5 +27,5 @@ Version 0.1.72 focuses on running MassGen as a cloud job on Modal.
 
 ## Related Tracks
 
-- **v0.1.71**: Trace Memory & Evaluation Polish — better eval criteria, system prompt tuning, stability fixes
-- **v0.1.73**: OpenAI Audio API ([#960](https://github.com/massgen/MassGen/issues/960))
+- **v0.1.72**: Grok Backend Update & Circuit Breaker Phase 2 — circuit breaker across all backends, Grok improvements ([#1038](https://github.com/massgen/MassGen/pull/1038), [#1044](https://github.com/massgen/MassGen/pull/1044))
+- **v0.1.74**: OpenAI Audio API ([#960](https://github.com/massgen/MassGen/issues/960))
diff --git a/docs/announcements/archive/v0.1.71.md b/docs/announcements/archive/v0.1.71.md
new file mode 100644
index 000000000..5464a8ac5
--- /dev/null
+++ b/docs/announcements/archive/v0.1.71.md
@@ -0,0 +1,61 @@
+# MassGen v0.1.71 Release Announcement
+
+<!--
+This is the current release announcement. Copy this + feature-highlights.md to LinkedIn/X.
+After posting, update the social links below.
+-->
+
+## Release Summary
+
+We're excited to release MassGen v0.1.71 — Trace Memory & Evaluation Polish! 🚀 Trace analyzer subagents now launch in the background after each round to write insights from the previous round's execution trace into memory. Plus: improved evaluation criteria generation, system prompt tuning, and stability fixes.
+
+## Install
+
+```bash
+pip install massgen==0.1.71
+```
+
+## Links
+
+- **Release notes:** https://github.com/massgen/MassGen/releases/tag/v0.1.71
+- **X post:** [TO BE ADDED AFTER POSTING]
+- **LinkedIn post:** [TO BE ADDED AFTER POSTING]
+
+---
+
+## Full Announcement (for LinkedIn)
+
+Copy everything below this line, then append content from `feature-highlights.md`:
+
+---
+
+We're excited to release MassGen v0.1.71 — Trace Memory & Evaluation Polish! 🚀 Trace analyzer subagents now launch in the background after each round to write insights from the previous round's execution trace into memory. Plus: improved evaluation criteria generation, system prompt tuning, and stability fixes.
+
+**Key Improvements:**
+
+🔍 **Trace Analyzer Subagents** — Background intelligence from execution traces:
+- Automatically launches after each round to analyze the previous round's execution trace
+- Writes insights into memory for next-round continuity
+- Fixes for trace memory handling and analyzer launch issues
+
+📋 **Better Evaluation Criteria** — Improved criteria generation for higher-quality, more opinionated output
+
+🧠 **System Prompt Tuning** — Adjusted system prompts for better agent performance across coordination rounds
+
+**Plus:**
+- 🔧 **Fix final injection** — Corrected injection behavior at the final stage
+- 🔧 **Fix eval criteria GPT pre-collab** — Resolved evaluation criteria issues with GPT models during pre-collaboration phase
+- 🔧 **Auto round fix for memory** — Fixed automatic round handling for memory
+
+**Getting Started:**
+
+```bash
+pip install massgen==0.1.71
+uv run massgen --config @examples/features/trace_analyzer_background.yaml "Create an svg of an AI agent coding."
+```
+
+Release notes: https://github.com/massgen/MassGen/releases/tag/v0.1.71
+
+Feature highlights:
+
+<!-- Paste feature-highlights.md content here -->
diff --git a/docs/announcements/current-release.md b/docs/announcements/current-release.md
index 5464a8ac5..7f4c0abf9 100644
--- a/docs/announcements/current-release.md
+++ b/docs/announcements/current-release.md
@@ -1,4 +1,4 @@
-# MassGen v0.1.71 Release Announcement
+# MassGen v0.1.72 Release Announcement
 
 <!--
 This is the current release announcement. Copy this + feature-highlights.md to LinkedIn/X.
@@ -7,17 +7,17 @@ After posting, update the social links below.
 
 ## Release Summary
 
-We're excited to release MassGen v0.1.71 — Trace Memory & Evaluation Polish! 🚀 Trace analyzer subagents now launch in the background after each round to write insights from the previous round's execution trace into memory. Plus: improved evaluation criteria generation, system prompt tuning, and stability fixes.
+We're excited to release MassGen v0.1.72 — Grok Backend Update & Circuit Breaker Phase 2! 🚀 Updated Grok backend with latest improvements. LLM API circuit breaker extended to ChatCompletions, Response API, and Gemini backends (was Claude-only).
 
 ## Install
 
 ```bash
-pip install massgen==0.1.71
+pip install massgen==0.1.72
 ```
 
 ## Links
 
-- **Release notes:** https://github.com/massgen/MassGen/releases/tag/v0.1.71
+- **Release notes:** https://github.com/massgen/MassGen/releases/tag/v0.1.72
 - **X post:** [TO BE ADDED AFTER POSTING]
 - **LinkedIn post:** [TO BE ADDED AFTER POSTING]
 
@@ -29,32 +29,24 @@ Copy everything below this line, then append content from `feature-highlights.md
 
 ---
 
-We're excited to release MassGen v0.1.71 — Trace Memory & Evaluation Polish! 🚀 Trace analyzer subagents now launch in the background after each round to write insights from the previous round's execution trace into memory. Plus: improved evaluation criteria generation, system prompt tuning, and stability fixes.
+We're excited to release MassGen v0.1.72 — Grok Backend Update & Circuit Breaker Phase 2! 🚀 Updated Grok backend with latest improvements. LLM API circuit breaker extended to ChatCompletions, Response API, and Gemini backends (was Claude-only).
 
 **Key Improvements:**
 
-🔍 **Trace Analyzer Subagents** — Background intelligence from execution traces:
-- Automatically launches after each round to analyze the previous round's execution trace
-- Writes insights into memory for next-round continuity
-- Fixes for trace memory handling and analyzer launch issues
+🦎 **Grok Backend Update** — Updated Grok backend with latest improvements
 
-📋 **Better Evaluation Criteria** — Improved criteria generation for higher-quality, more opinionated output
-
-🧠 **System Prompt Tuning** — Adjusted system prompts for better agent performance across coordination rounds
-
-**Plus:**
-- 🔧 **Fix final injection** — Corrected injection behavior at the final stage
-- 🔧 **Fix eval criteria GPT pre-collab** — Resolved evaluation criteria issues with GPT models during pre-collaboration phase
-- 🔧 **Auto round fix for memory** — Fixed automatic round handling for memory
+⚡ **Circuit Breaker Phase 2** — Rate limit protection across all major backends:
+- Extended to ChatCompletions, Response API, and Gemini backends (previously Claude-only)
+- Gemini backend also handles 503 errors
 
 **Getting Started:**
 
 ```bash
-pip install massgen==0.1.71
-uv run massgen --config @examples/features/trace_analyzer_background.yaml "Create an svg of an AI agent coding."
+pip install massgen==0.1.72
+uv run massgen --config @examples/providers/others/grok_x_search.yaml "Research the latest posts and news about AI agents in the last week, and summarize the key trends and insights."
 ```
 
-Release notes: https://github.com/massgen/MassGen/releases/tag/v0.1.71
+Release notes: https://github.com/massgen/MassGen/releases/tag/v0.1.72
 
 Feature highlights:
 
diff --git a/docs/announcements/github-release-v0.1.71.md b/docs/announcements/github-release-v0.1.71.md
deleted file mode 100644
index d8a5506c2..000000000
--- a/docs/announcements/github-release-v0.1.71.md
+++ /dev/null
@@ -1,27 +0,0 @@
-# 🚀 Release Highlights — v0.1.71 (2026-04-01)
-
-### 🔍 Trace Analyzer Subagents
-- **Background execution trace analysis**: Trace analyzer subagents now launch in the background after each round to analyze the previous round's execution trace and write insights into memory for next-round continuity
-- **Fix trace memory**: Corrected memory handling in execution traces
-- **Fix execution trace analyzer launch**: Trace analyzer now starts correctly
-
-### 📋 Better Evaluation Criteria
-- **Improved criteria generation**: Higher-quality, more opinionated criteria output
-
-### 🧠 System Prompt Tuning
-- **Better agent performance**: Adjusted system prompts for improved coordination across rounds
-
-### 🔧 Fixes
-- **Fix final injection**: Corrected injection behavior at the final stage
-- **Fix eval criteria GPT pre-collab**: Resolved evaluation criteria issues with GPT models during pre-collaboration phase
-- **Auto round fix for memory**: Fixed automatic round handling for memory
-
----
-
-### 📖 Getting Started
-- [**Quick Start Guide**](https://github.com/massgen/MassGen?tab=readme-ov-file#1--installation)
-- **Try It**:
-  ```bash
-  pip install massgen==0.1.71
-  uv run massgen --config @examples/features/trace_analyzer_background.yaml "Create an svg of an AI agent coding."
-  ```
diff --git a/docs/announcements/github-release-v0.1.72.md b/docs/announcements/github-release-v0.1.72.md
new file mode 100644
index 000000000..e10b2eb4f
--- /dev/null
+++ b/docs/announcements/github-release-v0.1.72.md
@@ -0,0 +1,18 @@
+# 🚀 Release Highlights — v0.1.72 (2026-04-03)
+
+### 🦎 [Grok Backend Update](https://docs.massgen.ai/en/latest/user_guide/backends.html)
+- **Backend improvements** ([#1044](https://github.com/massgen/MassGen/pull/1044)): Updated Grok backend with latest improvements
+
+### ⚡ [Circuit Breaker Phase 2](https://docs.massgen.ai/en/latest/user_guide/backends.html)
+- **Extended to all major backends** ([#1038](https://github.com/massgen/MassGen/pull/1038)): LLM API circuit breaker now covers ChatCompletions, Response API, and Gemini (was Claude-only in v0.1.68)
+- **Gemini 503 handling**: Gemini backend circuit breaker also triggers on 503 errors
+
+---
+
+### 📖 Getting Started
+- [**Quick Start Guide**](https://github.com/massgen/MassGen?tab=readme-ov-file#1--installation)
+- **Try It**:
+  ```bash
+  pip install massgen==0.1.72
+  uv run massgen --config @examples/providers/others/grok_x_search.yaml "Research the latest posts and news about AI agents in the last week, and summarize the key trends and insights."
+  ```
diff --git a/docs/source/index.rst b/docs/source/index.rst
index 1611612d5..ebffae420 100644
--- a/docs/source/index.rst
+++ b/docs/source/index.rst
@@ -200,18 +200,18 @@ Key Features
 Recent Releases
 ---------------
 
+**v0.1.72 (April 3, 2026)** - Grok Backend Update & Circuit Breaker Phase 2
+
+Grok backend update with latest improvements. LLM API circuit breaker extended to ChatCompletions, Response API, and Gemini backends (was Claude-only). Config plumbing smoke tests for all backends.
+
 **v0.1.71 (April 1, 2026)** - Trace Memory & Evaluation Polish
 
-Trace analyzer subagents now launch in the background after each round to write insights from execution traces into memory. Improved evaluation criteria generation and system prompt tuning. Stability fixes for injection, trace analyzer launch, and memory handling.
+Trace analyzer subagents now launch in the background after each round to write insights from execution traces into memory. Improved evaluation criteria generation and system prompt tuning.
 
 **v0.1.70 (March 30, 2026)** - Evaluation Criteria Redesign
 
 Redesigned three-tier evaluation criteria with anti-pattern definitions and aspiration statements. Improved checklist-gated evaluation with tighter iterative submission cycles. Fast iteration mode, WebUI review modal, and background trace analysis.
 
-**v0.1.69 (March 27, 2026)** - WebUI Automation & Improved Skill
-
-WebUI automation now auto-starts without browser interaction — open the URL at any point mid-run to monitor progress. MassGen skill redesign for increased usability and WebUI integration. Quickstart Wizard rework and Workspace Browser expansion.
-
 :doc:`Full changelog → <changelog>`
 
 Supported Models
diff --git a/massgen/__init__.py b/massgen/__init__.py
index 4d0bfff10..e3d5f529b 100644
--- a/massgen/__init__.py
+++ b/massgen/__init__.py
@@ -86,7 +86,7 @@
 from .message_templates import MessageTemplates, get_templates
 from .orchestrator import Orchestrator, create_orchestrator
 
-__version__ = "0.1.71"
+__version__ = "0.1.72"
 __author__ = "MassGen Contributors"
 
 
diff --git a/massgen/agent_config.py b/massgen/agent_config.py
index 94005e5fe..338e4d902 100644
--- a/massgen/agent_config.py
+++ b/massgen/agent_config.py
@@ -674,12 +674,21 @@ def create_claude_config(
         return cls(backend_params=backend_params)
 
     @classmethod
-    def create_grok_config(cls, model: str = "grok-2-1212", enable_web_search: bool = False, **kwargs) -> "AgentConfig":
+    def create_grok_config(
+        cls,
+        model: str = "grok-2-1212",
+        enable_web_search: bool = False,
+        enable_x_search: bool = False,
+        enable_code_execution: bool = False,
+        **kwargs,
+    ) -> "AgentConfig":
         """Create xAI Grok configuration.
 
         Args:
             model: Grok model name
-            enable_web_search: Enable Live Search feature
+            enable_web_search: Enable xAI web search
+            enable_x_search: Enable xAI X search
+            enable_code_execution: Enable xAI code execution
             **kwargs: Additional backend parameters
         """
         backend_params = {"model": model, **kwargs}
@@ -687,6 +696,10 @@ def create_grok_config(cls, model: str = "grok-2-1212", enable_web_search: bool
         # Add tool enablement to backend_params
         if enable_web_search:
             backend_params["enable_web_search"] = True
+        if enable_x_search:
+            backend_params["enable_x_search"] = True
+        if enable_code_execution:
+            backend_params["enable_code_execution"] = True
 
         return cls(backend_params=backend_params)
 
@@ -945,6 +958,8 @@ def for_computational_task(cls, model: str = "gpt-4o", backend: str = "openai")
         """
         if backend == "openai":
             return cls.create_openai_config(model, enable_code_interpreter=True)
+        elif backend == "grok":
+            return cls.create_grok_config(model, enable_code_execution=True)
         elif backend == "claude":
             return cls.create_claude_config(model, enable_code_execution=True)
         elif backend == "gemini":
diff --git a/massgen/api_params_handler/_response_api_params_handler.py b/massgen/api_params_handler/_response_api_params_handler.py
index 8b739aa08..22be03a94 100644
--- a/massgen/api_params_handler/_response_api_params_handler.py
+++ b/massgen/api_params_handler/_response_api_params_handler.py
@@ -19,6 +19,8 @@ def get_excluded_params(self) -> set[str]:
         return self.get_base_excluded_params().union(
             {
                 "enable_web_search",
+                "enable_x_search",
+                "enable_code_execution",
                 "enable_code_interpreter",
                 "allowed_tools",
                 "exclude_tools",
@@ -38,11 +40,15 @@ def get_excluded_params(self) -> set[str]:
     def get_provider_tools(self, all_params: dict[str, Any]) -> list[dict[str, Any]]:
         """Get provider tools for Response API format."""
         provider_tools = []
+        provider_name = self.backend.get_provider_name()
 
         if all_params.get("enable_web_search", False):
             provider_tools.append({"type": "web_search"})
 
-        if all_params.get("enable_code_interpreter", False):
+        if provider_name == "Grok" and all_params.get("enable_x_search", False):
+            provider_tools.append({"type": "x_search"})
+
+        if all_params.get("enable_code_interpreter", False) or all_params.get("enable_code_execution", False):
             provider_tools.append(
                 {
                     "type": "code_interpreter",
@@ -116,12 +122,18 @@ async def build_api_params(
             logger.debug(f"Using previous_response_id for reasoning continuity: {previous_response_id}")
 
         # Handle parallel_tool_calls with built-in tools constraint
-        builtin_flags = ("enable_web_search", "enable_code_interpreter", "_has_file_search_files")
+        builtin_flags = (
+            "enable_web_search",
+            "enable_x_search",
+            "enable_code_execution",
+            "enable_code_interpreter",
+            "_has_file_search_files",
+        )
         if any(all_params.get(f, False) for f in builtin_flags):
             # Built-in tools present - MUST disable parallel calling
             if all_params.get("parallel_tool_calls") is True:
                 logger.warning(
-                    "parallel_tool_calls=true is not supported with built-in tools " "(web_search, code_interpreter, file_search). " "Setting parallel_tool_calls=false.",
+                    "parallel_tool_calls=true is not supported with built-in tools " "(web_search, x_search, code_interpreter, file_search). " "Setting parallel_tool_calls=false.",
                 )
             api_params["parallel_tool_calls"] = False
         elif "parallel_tool_calls" in all_params:
diff --git a/massgen/backend/capabilities.py b/massgen/backend/capabilities.py
index b3fe9bd1b..8f38d6900 100644
--- a/massgen/backend/capabilities.py
+++ b/massgen/backend/capabilities.py
@@ -60,6 +60,7 @@ class Capability(Enum):
     """Enumeration of all possible backend capabilities."""
 
     WEB_SEARCH = "web_search"
+    X_SEARCH = "x_search"
     CODE_EXECUTION = "code_execution"
     BASH = "bash"
     MULTIMODAL = "multimodal"  # Legacy - being phased out
@@ -386,10 +387,12 @@ class BackendCapabilities:
         provider_name="Grok",
         supported_capabilities={
             "web_search",
+            "x_search",
+            "code_execution",
             "mcp",
             "image_understanding",
         },
-        builtin_tools=["web_search"],
+        builtin_tools=["web_search", "x_search", "code_execution"],
         filesystem_support="mcp",
         models=[
             "grok-4.20-0309-reasoning",
@@ -403,7 +406,7 @@ class BackendCapabilities:
         ],
         default_model="grok-4.20-0309-reasoning",
         env_var="XAI_API_KEY",
-        notes="Web search includes real-time data access. Image understanding capabilities.",
+        notes=("Uses xAI's Responses API tooling surface. " "Supports web_search, x_search, and code execution. " "Legacy Chat Completions search_parameters are not supported."),
         model_release_dates={
             "grok-4.20-0309-reasoning": "2026-03",
             "grok-4-1-fast-reasoning": "2025-11",
@@ -414,6 +417,7 @@ class BackendCapabilities:
             "grok-3": "2025-02",
             "grok-3-mini": "2025-05",
         },
+        base_url="https://api.x.ai/v1",
     ),
     "azure_openai": BackendCapabilities(
         backend_type="azure_openai",
@@ -854,6 +858,14 @@ def validate_backend_config(backend_type: str, config: dict) -> list[str]:
         if "web_search" not in caps.supported_capabilities:
             errors.append(f"{backend_type} does not support web_search")
 
+    if config.get("enable_x_search"):
+        if backend_type != "grok":
+            errors.append(
+                f"enable_x_search is only supported by Grok backend, not {backend_type}",
+            )
+        elif "x_search" not in caps.supported_capabilities:
+            errors.append(f"{backend_type} does not support x_search")
+
     if "enable_code_execution" in config and config["enable_code_execution"]:
         if "code_execution" not in caps.supported_capabilities:
             errors.append(f"{backend_type} does not support code_execution")
@@ -880,6 +892,13 @@ def validate_backend_config(backend_type: str, config: dict) -> list[str]:
         if "mcp" not in caps.supported_capabilities:
             errors.append(f"{backend_type} does not support MCP")
 
+    if backend_type == "grok":
+        extra_body = config.get("extra_body")
+        if isinstance(extra_body, dict) and "search_parameters" in extra_body:
+            errors.append(
+                "Grok no longer supports extra_body.search_parameters. " "Use enable_web_search and/or enable_x_search instead.",
+            )
+
     # Check for deprecated system prompt parameters (standardized across all backends)
     if "append_system_prompt" in config:
         errors.append(
diff --git a/massgen/backend/chat_completions.py b/massgen/backend/chat_completions.py
index 5ae627060..40ea6f402 100644
--- a/massgen/backend/chat_completions.py
+++ b/massgen/backend/chat_completions.py
@@ -43,6 +43,11 @@
     CustomToolChunk,
     ToolExecutionConfig,
 )
+from .llm_circuit_breaker import (
+    CircuitBreakerOpenError,
+    LLMCircuitBreaker,
+    LLMCircuitBreakerConfig,
+)
 
 
 class ChatCompletionsBackend(StreamingBufferMixin, CustomToolAndMCPBackend):
@@ -58,6 +63,8 @@ class ChatCompletionsBackend(StreamingBufferMixin, CustomToolAndMCPBackend):
     """
 
     def __init__(self, api_key: str | None = None, **kwargs):
+        # Extract circuit breaker config before passing to super
+        cb_config = self._build_circuit_breaker_config(kwargs)
         super().__init__(api_key, **kwargs)
         # Backend name is already set in MCPBackend, but we may need to override it
         self.backend_name = self.get_provider_name()
@@ -72,6 +79,27 @@ def __init__(self, api_key: str | None = None, **kwargs):
         self._stream_usage_received: bool = True  # True = no pending estimation needed
         # Track reasoning state for streaming (needed for reasoning_done transition)
         self._reasoning_active: bool = False
+        self.circuit_breaker = LLMCircuitBreaker(
+            config=cb_config,
+            backend_name=self.get_provider_name(),
+        )
+
+    @staticmethod
+    def _build_circuit_breaker_config(
+        kwargs: dict[str, Any],
+    ) -> LLMCircuitBreakerConfig:
+        """Extract circuit breaker settings from kwargs and build config."""
+        cb_kwargs: dict[str, Any] = {}
+        prefix = "llm_circuit_breaker_"
+        keys_to_pop: list[str] = []
+        for key in kwargs:
+            if key.startswith(prefix):
+                param = key[len(prefix) :]
+                cb_kwargs[param] = kwargs[key]
+                keys_to_pop.append(key)
+        for key in keys_to_pop:
+            kwargs.pop(key)
+        return LLMCircuitBreakerConfig(**cb_kwargs)
 
     def finalize_token_tracking(self) -> None:
         """Finalize token tracking by estimating tokens for any interrupted streams.
@@ -276,9 +304,19 @@ async def _stream_with_custom_and_mcp_tools(
             model=model,
             operation="stream",
         ) as llm_span:
-            # Start streaming - wrap in try/except for context length errors
+            # Start streaming - wrap with circuit breaker + context length handling
             try:
-                stream = await client.chat.completions.create(**api_params)
+
+                async def _make_api_call():
+                    return await client.chat.completions.create(**api_params)
+
+                stream = await self.circuit_breaker.call_with_retry(
+                    _make_api_call,
+                    agent_id=agent_id,
+                )
+            except CircuitBreakerOpenError:
+                self.end_api_call_timing(success=False, error="circuit_breaker_open")
+                raise
             except Exception as e:
                 if is_context_length_error(e) and not _compression_retry:
                     # Context length exceeded on initial request - compress and retry
diff --git a/massgen/backend/gemini.py b/massgen/backend/gemini.py
index 5279e633a..10976b9f6 100644
--- a/massgen/backend/gemini.py
+++ b/massgen/backend/gemini.py
@@ -51,6 +51,11 @@
     PostEvaluationResponse,
     VoteOnlyCoordinationResponse,
 )
+from .llm_circuit_breaker import (
+    CircuitBreakerOpenError,
+    LLMCircuitBreaker,
+    LLMCircuitBreakerConfig,
+)
 from .rate_limiter import GlobalRateLimiter
 
 
@@ -247,6 +252,9 @@ def __init__(self, api_key: str | None = None, **kwargs):
         # Store Gemini-specific API key before calling parent init
         gemini_api_key = api_key or os.getenv("GOOGLE_API_KEY") or os.getenv("GEMINI_API_KEY")
 
+        # Extract circuit breaker config before other kwargs processing
+        cb_config = self._build_circuit_breaker_config(kwargs)
+
         # Extract and remove enable_rate_limit and backoff config
         enable_rate_limit = kwargs.pop("enable_rate_limit", False)
         model_name = kwargs.get("model", "")
@@ -293,6 +301,12 @@ def __init__(self, api_key: str | None = None, **kwargs):
         self.backoff_retry_count = 0
         self.backoff_total_delay = 0.0
 
+        # LLM circuit breaker (opt-in, default disabled)
+        self.circuit_breaker = LLMCircuitBreaker(
+            config=cb_config,
+            backend_name="gemini",
+        )
+
         # Initialize multi-dimensional rate limiter for Gemini API
         # Supports RPM (Requests Per Minute), TPM (Tokens Per Minute), RPD (Requests Per Day)
         # Configuration loaded from massgen/config/rate_limits.yaml
@@ -335,6 +349,23 @@ def __init__(self, api_key: str | None = None, **kwargs):
             self.rate_limiter = None
             logger.info(f"[Gemini] Rate limiting disabled for '{model_name}'")
 
+    @staticmethod
+    def _build_circuit_breaker_config(
+        kwargs: dict[str, Any],
+    ) -> LLMCircuitBreakerConfig:
+        """Extract circuit breaker settings from kwargs and build config."""
+        cb_kwargs: dict[str, Any] = {}
+        prefix = "llm_circuit_breaker_"
+        keys_to_pop: list[str] = []
+        for key in kwargs:
+            if key.startswith(prefix):
+                param = key[len(prefix) :]
+                cb_kwargs[param] = kwargs[key]
+                keys_to_pop.append(key)
+        for key in keys_to_pop:
+            kwargs.pop(key)
+        return LLMCircuitBreakerConfig(**cb_kwargs)
+
     def _normalize_and_resolve_tool_name(self, tool_name: str) -> str:
         """Normalize Gemini tool names and resolve MCP aliases.
 
@@ -777,6 +808,11 @@ async def stream_with_tools(self, messages: list[dict[str, Any]], tools: list[di
             last_response_with_candidates = None
 
             cfg = self.backoff_config
+
+            # Circuit breaker gate
+            if self.circuit_breaker.should_block():
+                raise CircuitBreakerOpenError("Circuit breaker is open for gemini")
+
             first_token_recorded = False
             for stream_attempt in range(1, cfg.max_attempts + 1):
                 try:
@@ -863,6 +899,7 @@ async def stream_with_tools(self, messages: list[dict[str, Any]], tools: list[di
 
                     # End API call timing on successful completion
                     self.end_api_call_timing(success=True)
+                    self.circuit_breaker.record_success()
                     break
 
                 except Exception as stream_exc:
@@ -873,6 +910,10 @@ async def stream_with_tools(self, messages: list[dict[str, Any]], tools: list[di
 
                     if not is_retryable or stream_attempt >= cfg.max_attempts:
                         if is_retryable:
+                            self.circuit_breaker.record_failure(
+                                error_type=f"exhausted_{status_code or 'unknown'}",
+                                error_message=f"Max retries exhausted: {error_msg[:200]}",
+                            )
                             yield StreamChunk(
                                 type="error",
                                 error=f"⚠️ Rate limit exceeded after {cfg.max_attempts} retries. Please try again later.",
@@ -1443,6 +1484,10 @@ def tool_config_for_call(call: dict[str, Any]) -> ToolExecutionConfig:
                     cont_first_token_recorded = False
 
                     # Retry for continuation with backoff
+                    # Circuit breaker gate
+                    if self.circuit_breaker.should_block():
+                        raise CircuitBreakerOpenError("Circuit breaker is open for gemini")
+
                     for cont_attempt in range(1, cfg.max_attempts + 1):
                         try:
                             # Start API call timing for continuation
@@ -1519,16 +1564,21 @@ def tool_config_for_call(call: dict[str, Any]) -> ToolExecutionConfig:
 
                             # End API call timing on successful completion
                             self.end_api_call_timing(success=True)
+                            self.circuit_breaker.record_success()
                             break
 
                         except Exception as cont_exc:
                             # End API call timing with failure
                             self.end_api_call_timing(success=False, error=str(cont_exc))
-                            is_retryable, status_code, _ = _is_retryable_gemini_error(cont_exc, cfg.retry_statuses)
+                            is_retryable, status_code, error_msg = _is_retryable_gemini_error(cont_exc, cfg.retry_statuses)
 
                             if not is_retryable or cont_attempt >= cfg.max_attempts:
                                 # Yield user-friendly error before raising
                                 if is_retryable:
+                                    self.circuit_breaker.record_failure(
+                                        error_type=f"exhausted_{status_code or 'unknown'}",
+                                        error_message=f"Max retries exhausted: {error_msg[:200]}",
+                                    )
                                     yield StreamChunk(
                                         type="error",
                                         error=f"⚠️ Rate limit exceeded after {cfg.max_attempts} retries. Please try again later.",
diff --git a/massgen/backend/grok.py b/massgen/backend/grok.py
index c9f6f7002..f6a90d952 100644
--- a/massgen/backend/grok.py
+++ b/massgen/backend/grok.py
@@ -1,81 +1,36 @@
-"""
-Grok/xAI backend is using the chat_completions backend for streaming.
-It overrides methods for Grok-specific features (Grok Live Search).
+"""xAI Grok backend implemented on the shared Responses API stack."""
 
-✅ TESTED: Backend works correctly with architecture
-- ✅ Grok API integration working (through chat_completions)
-- ✅ Streaming functionality working correctly
-- ✅ SingleAgent integration working
-- ✅ Error handling and pricing calculations implemented
-- ✅ Web search is working through Grok Live Search
-- ✅ MCP is working
-
-TODO for future releases:
-- Test multi-agent orchestrator integration
-- Validate advanced Grok-specific features
-"""
-
-# -*- coding: utf-8 -*-
 from __future__ import annotations
 
-import logging
 import os
 from typing import Any
 
-from openai import AsyncOpenAI
-
-from ..logger_config import log_stream_chunk
-from .chat_completions import ChatCompletionsBackend
+from .response import ResponseBackend
 
-logger = logging.getLogger(__name__)
 
+class GrokBackend(ResponseBackend):
+    """Grok backend using xAI's Responses-compatible API surface."""
 
-class GrokBackend(ChatCompletionsBackend):
-    """Grok backend using xAI's OpenAI-compatible API."""
+    XAI_BASE_URL = "https://api.x.ai/v1"
 
     def __init__(self, api_key: str | None = None, **kwargs):
-        super().__init__(api_key, **kwargs)
+        self._reject_legacy_search_parameters(kwargs)
+        kwargs.setdefault("base_url", self.XAI_BASE_URL)
+        super().__init__(api_key=api_key, **kwargs)
         self.api_key = api_key or os.getenv("XAI_API_KEY")
-        self.base_url = "https://api.x.ai/v1"
-
-    def _create_client(self, **kwargs) -> AsyncOpenAI:
-        """Create OpenAI client configured for xAI's Grok API."""
-        import openai
-
-        return openai.AsyncOpenAI(api_key=self.api_key, base_url=self.base_url)
-
-    def _customize_api_params(
-        self,
-        api_params: dict[str, Any],
-        all_params: dict[str, Any],
-    ) -> dict[str, Any]:
-        """Add Grok Live Search parameters to API params if web search is enabled.
-
-        This hook is called by the parent class before making the API call.
-        """
-        enable_web_search = False  # all_params.get("enable_web_search", False)
-        logger.debug(f"GrokBackend: Live search is currently deprecated, enable_web_search={enable_web_search} for now")
-
-        if enable_web_search:
-            # Check for conflict with manually specified search_parameters
-            existing_extra = api_params.get("extra_body", {})
-            if isinstance(existing_extra, dict) and "search_parameters" in existing_extra:
-                error_message = "Conflict: Cannot use both 'enable_web_search: true' and manual 'extra_body.search_parameters'. Use one or the other."
-                log_stream_chunk("backend.grok", "error", error_message, self.agent_id)
-                raise ValueError(error_message)
-
-            # Merge search_parameters into existing extra_body
-            search_params = {"mode": "auto", "return_citations": True}
-            merged_extra = existing_extra.copy() if existing_extra else {}
-            merged_extra["search_parameters"] = search_params
-            api_params["extra_body"] = merged_extra
 
-        return api_params
+    @staticmethod
+    def _reject_legacy_search_parameters(kwargs: dict[str, Any]) -> None:
+        extra_body = kwargs.get("extra_body")
+        if isinstance(extra_body, dict) and "search_parameters" in extra_body:
+            raise ValueError(
+                "Grok no longer supports legacy extra_body.search_parameters. " "Use enable_web_search: true and/or enable_x_search: true instead.",
+            )
 
     def get_provider_name(self) -> str:
-        """Get the name of this provider."""
+        """Get the provider name."""
         return "Grok"
 
     def get_supported_builtin_tools(self) -> list[str]:
-        """Get list of builtin tools supported by Grok."""
-        return ["web_search"]
+        """Get the builtin tools exposed through the Grok backend."""
+        return ["web_search", "x_search", "code_execution"]
diff --git a/massgen/backend/response.py b/massgen/backend/response.py
index 0d8723164..f2dbf39ac 100644
--- a/massgen/backend/response.py
+++ b/massgen/backend/response.py
@@ -35,6 +35,11 @@
     ToolExecutionConfig,
     UploadFileError,
 )
+from .llm_circuit_breaker import (
+    CircuitBreakerOpenError,
+    LLMCircuitBreaker,
+    LLMCircuitBreakerConfig,
+)
 
 
 class _WSEvent:
@@ -84,7 +89,12 @@ def __repr__(self) -> str:
 class ResponseBackend(StreamingBufferMixin, CustomToolAndMCPBackend):
     """Backend using the standard Response API format with multimodal support."""
 
+    _XAI_X_SEARCH_CUSTOM_TOOL_CALL_ID_PREFIX = "xs_"
+    _XAI_X_SEARCH_CUSTOM_TOOL_NAME_PREFIX = "x_"
+
     def __init__(self, api_key: str | None = None, **kwargs):
+        # Extract circuit breaker config before passing to super
+        cb_config = self._build_circuit_breaker_config(kwargs)
         super().__init__(api_key, **kwargs)
         self.api_key = api_key or os.getenv("OPENAI_API_KEY")
         self.formatter = ResponseFormatter()
@@ -106,7 +116,40 @@ def __init__(self, api_key: str | None = None, **kwargs):
         self._vector_store_ids: list[str] = []
         self._uploaded_file_ids: list[str] = []
 
+        # xAI currently surfaces some builtin tools as custom_tool_call items.
+        # Keep minimal per-stream state so we can label midstream input chunks.
+        self._xai_custom_tool_names_by_call_id: dict[str, str] = {}
+        self._xai_custom_tool_inputs_by_call_id: dict[str, str] = {}
+        self._xai_custom_tool_calls_reported: set[str] = set()
+
         # Note: _streaming_buffer is provided by StreamingBufferMixin
+        self.circuit_breaker = LLMCircuitBreaker(
+            config=cb_config,
+            backend_name="response_api",
+        )
+
+    @staticmethod
+    def _build_circuit_breaker_config(
+        kwargs: dict[str, Any],
+    ) -> LLMCircuitBreakerConfig:
+        """Extract circuit breaker settings from kwargs and build config."""
+        cb_kwargs: dict[str, Any] = {}
+        prefix = "llm_circuit_breaker_"
+        keys_to_pop: list[str] = []
+        for key in kwargs:
+            if key.startswith(prefix):
+                param = key[len(prefix) :]
+                cb_kwargs[param] = kwargs[key]
+                keys_to_pop.append(key)
+        for key in keys_to_pop:
+            kwargs.pop(key)
+        return LLMCircuitBreakerConfig(**cb_kwargs)
+
+    def _reset_provider_tool_state(self) -> None:
+        """Clear per-stream provider tool state."""
+        self._xai_custom_tool_names_by_call_id.clear()
+        self._xai_custom_tool_inputs_by_call_id.clear()
+        self._xai_custom_tool_calls_reported.clear()
 
     def supports_upload_files(self) -> bool:
         return True
@@ -123,6 +166,7 @@ async def stream_with_tools(
         """
         # Clear streaming buffer at start of new stream (mixin respects _compression_retry)
         self._clear_streaming_buffer(**kwargs)
+        self._reset_provider_tool_state()
         agent_id = kwargs.get("agent_id", self.agent_id)
 
         ws_transport = None
@@ -244,12 +288,20 @@ async def _stream_without_custom_and_mcp_tools(
         _compression_retry = kwargs.get("_compression_retry", False)
         ws_transport = kwargs.get("_ws_transport")
 
+        # Start API call timing for non-MCP path
+        model = api_params.get("model", "unknown")
+        self.start_api_call_timing(model)
+
         try:
             stream = await self._create_response_stream(
                 api_params,
                 client,
                 ws_transport,
+                agent_id=agent_id,
             )
+        except CircuitBreakerOpenError:
+            self.end_api_call_timing(success=False, error="circuit_breaker_open")
+            raise
         except Exception as e:
             # Debug: Catch input[N].content format errors and print the problematic message
             error_str = str(e)
@@ -271,6 +323,7 @@ async def _stream_without_custom_and_mcp_tools(
             from ._context_errors import is_context_length_error
 
             if is_context_length_error(e) and not _compression_retry:
+                self.end_api_call_timing(success=False, error=str(e))
                 logger.warning(
                     f"[{self.get_provider_name()}] Context length exceeded, " f"attempting compression recovery...",
                 )
@@ -307,6 +360,7 @@ async def _stream_without_custom_and_mcp_tools(
                     api_params,
                     client,
                     ws_transport,
+                    agent_id=agent_id,
                 )
 
                 # Notify user that compression succeeded
@@ -323,11 +377,129 @@ async def _stream_without_custom_and_mcp_tools(
                     f"[{self.get_provider_name()}] Compression recovery successful via summarization " f"({input_count} items)",
                 )
             else:
+                self.end_api_call_timing(success=False, error=str(e))
                 raise
 
         async for chunk in self._process_stream(stream, all_params, agent_id):
             yield chunk
 
+    @staticmethod
+    def _get_response_item_field(item: Any, field: str, default: Any = None) -> Any:
+        """Read a field from either an SDK object or a plain dict."""
+        if isinstance(item, dict):
+            return item.get(field, default)
+        return getattr(item, field, default)
+
+    @staticmethod
+    def _parse_response_item_json(value: Any) -> dict[str, Any]:
+        """Parse a JSON-ish input payload from a provider output item."""
+        if isinstance(value, dict):
+            return value
+        if isinstance(value, str) and value:
+            try:
+                parsed = json.loads(value)
+            except json.JSONDecodeError:
+                return {}
+            return parsed if isinstance(parsed, dict) else {}
+        return {}
+
+    def _is_xai_x_search_custom_tool_item(self, item: Any) -> bool:
+        """Return True when xAI surfaces X search as provider custom_tool_call items.
+
+        xAI's Responses API currently emits X search work as ``custom_tool_call``
+        items (for example ``x_keyword_search`` and ``x_thread_fetch``) rather
+        than the ``response.x_search_call.*`` events we initially expected.
+        """
+        if self.get_provider_name() != "Grok":
+            return False
+        if self._get_response_item_field(item, "type") != "custom_tool_call":
+            return False
+        name = self._get_response_item_field(item, "name", "") or ""
+        call_id = self._get_response_item_field(item, "call_id", "") or ""
+        return call_id.startswith(self._XAI_X_SEARCH_CUSTOM_TOOL_CALL_ID_PREFIX) or name.startswith(
+            self._XAI_X_SEARCH_CUSTOM_TOOL_NAME_PREFIX,
+        )
+
+    def _record_xai_custom_tool_start(self, item: Any) -> None:
+        """Track xAI custom tool call metadata for subsequent input delta events."""
+        if not self._is_xai_x_search_custom_tool_item(item):
+            return
+        call_id = self._get_response_item_field(item, "call_id", "") or ""
+        name = self._get_response_item_field(item, "name", "") or ""
+        if call_id and name:
+            self._xai_custom_tool_names_by_call_id[call_id] = name
+
+    def _format_xai_x_search_custom_tool_input_chunk(
+        self,
+        call_id: str,
+        payload: dict[str, Any],
+        agent_id: str | None,
+    ) -> TextStreamChunk | None:
+        """Format a midstream xAI custom tool input payload when enough is known."""
+        tool_name = self._xai_custom_tool_names_by_call_id.get(call_id, "")
+        if not tool_name:
+            if payload.get("post_id"):
+                tool_name = "x_thread_fetch"
+            elif payload.get("query"):
+                tool_name = "x_search"
+            else:
+                return None
+
+        if call_id in self._xai_custom_tool_calls_reported:
+            return None
+
+        if not (payload.get("query") or payload.get("post_id")):
+            return None
+
+        self._xai_custom_tool_calls_reported.add(call_id)
+        return self._format_xai_x_search_custom_tool_chunk(
+            {
+                "name": tool_name,
+                "input": payload,
+            },
+            agent_id,
+        )
+
+    def _format_xai_x_search_custom_tool_chunk(
+        self,
+        item: Any,
+        agent_id: str | None,
+        *,
+        strip_newlines: bool = False,
+    ) -> TextStreamChunk:
+        """Format xAI provider custom tool items into user-visible stream chunks."""
+        tool_name = self._get_response_item_field(item, "name", "x_search")
+        payload = self._parse_response_item_json(self._get_response_item_field(item, "input"))
+        query = payload.get("query", "")
+        post_id = payload.get("post_id", "")
+
+        if tool_name in {"x_keyword_search", "x_semantic_search"}:
+            if query:
+                log_stream_chunk("backend.response", "x_search_query", query, agent_id)
+                content = f"\n🔎 [X Search Query] '{query}' via {tool_name}\n"
+            else:
+                log_stream_chunk("backend.response", "x_search", f"Running {tool_name}", agent_id)
+                content = f"\n🔎 [Provider Tool: X Search] Running {tool_name}...\n"
+        elif tool_name in {"x_thread_fetch", "x_post_fetch"}:
+            if post_id:
+                log_stream_chunk("backend.response", "x_thread_fetch", post_id, agent_id)
+                content = f"\n🧵 [X Thread Fetch] {post_id}\n"
+            else:
+                log_stream_chunk("backend.response", "x_thread_fetch", f"Running {tool_name}", agent_id)
+                content = f"\n🧵 [X Thread Fetch] Running {tool_name}...\n"
+        else:
+            log_stream_chunk("backend.response", "x_search", f"Running {tool_name}", agent_id)
+            content = f"\n🔎 [Provider Tool: X Search] Running {tool_name}...\n"
+
+        if strip_newlines:
+            content = content.strip("\n")
+
+        return TextStreamChunk(
+            type=ChunkType.CONTENT,
+            content=content,
+            source="response_api",
+        )
+
     def _append_tool_result_message(
         self,
         updated_messages: list[dict[str, Any]],
@@ -471,7 +643,11 @@ async def _stream_with_custom_and_mcp_tools(
                 api_params,
                 client,
                 ws_transport,
+                agent_id=agent_id,
             )
+        except CircuitBreakerOpenError:
+            self.end_api_call_timing(success=False, error="circuit_breaker_open")
+            raise
         except Exception as e:
             # Debug: Catch input[N].content format errors and print the problematic message
             error_str = str(e)
@@ -533,6 +709,7 @@ async def _stream_with_custom_and_mcp_tools(
                     api_params,
                     client,
                     ws_transport,
+                    agent_id=agent_id,
                 )
 
                 # Notify user that compression succeeded
@@ -1431,6 +1608,69 @@ def _process_stream_chunk(self, chunk, agent_id) -> TextStreamChunk | StreamChun
                 source="response_api",
             )
 
+        elif chunk_type == "response.x_search_call.in_progress":
+            log_stream_chunk("backend.response", "x_search", "Starting X search", agent_id)
+            return TextStreamChunk(
+                type=ChunkType.CONTENT,
+                content="\n🔎 [Provider Tool: X Search] Starting search...",
+                source="response_api",
+            )
+        elif chunk_type == "response.x_search_call.searching":
+            log_stream_chunk("backend.response", "x_search", "Searching X", agent_id)
+            return TextStreamChunk(
+                type=ChunkType.CONTENT,
+                content="\n🔎 [Provider Tool: X Search] Searching...",
+                source="response_api",
+            )
+        elif chunk_type == "response.x_search_call.completed":
+            log_stream_chunk("backend.response", "x_search", "X search completed", agent_id)
+            return TextStreamChunk(
+                type=ChunkType.CONTENT,
+                content="\n✅ [Provider Tool: X Search] Search completed",
+                source="response_api",
+            )
+        elif chunk_type == "response.output_item.added":
+            if hasattr(chunk, "item") and chunk.item and self._is_xai_x_search_custom_tool_item(chunk.item):
+                self._record_xai_custom_tool_start(chunk.item)
+                tool_name = self._get_response_item_field(chunk.item, "name", "x_search")
+                log_stream_chunk("backend.response", "x_search", f"Starting {tool_name}", agent_id)
+                return TextStreamChunk(
+                    type=ChunkType.CONTENT,
+                    content=f"\n🔎 [Provider Tool: X Search] Running {tool_name}...\n",
+                    source="response_api",
+                )
+        elif chunk_type == "response.custom_tool_call_input.delta":
+            call_id = getattr(chunk, "call_id", "") or ""
+            if call_id:
+                previous = self._xai_custom_tool_inputs_by_call_id.get(call_id, "")
+                delta = getattr(chunk, "delta", "") or ""
+                current = previous + delta
+                self._xai_custom_tool_inputs_by_call_id[call_id] = current
+                payload = self._parse_response_item_json(current)
+                formatted = self._format_xai_x_search_custom_tool_input_chunk(
+                    call_id,
+                    payload,
+                    agent_id,
+                )
+                if formatted is not None:
+                    return formatted
+        elif chunk_type == "response.custom_tool_call_input.done":
+            call_id = getattr(chunk, "call_id", "") or ""
+            if call_id:
+                final_input = getattr(chunk, "input", None)
+                if isinstance(final_input, str) and final_input:
+                    self._xai_custom_tool_inputs_by_call_id[call_id] = final_input
+                payload = self._parse_response_item_json(
+                    self._xai_custom_tool_inputs_by_call_id.get(call_id, ""),
+                )
+                formatted = self._format_xai_x_search_custom_tool_input_chunk(
+                    call_id,
+                    payload,
+                    agent_id,
+                )
+                if formatted is not None:
+                    return formatted
+
         elif chunk_type == "response.code_interpreter_call.in_progress":
             log_stream_chunk("backend.response", "code_interpreter", "Starting execution", agent_id)
             return TextStreamChunk(
@@ -1438,13 +1678,31 @@ def _process_stream_chunk(self, chunk, agent_id) -> TextStreamChunk | StreamChun
                 content="\n💻 [Provider Tool: Code Interpreter] Starting execution...",
                 source="response_api",
             )
-        elif chunk_type == "response.code_interpreter_call.executing":
+        elif chunk_type in {"response.code_interpreter_call.executing", "response.code_interpreter_call.interpreting"}:
             log_stream_chunk("backend.response", "code_interpreter", "Executing", agent_id)
             return TextStreamChunk(
                 type=ChunkType.CONTENT,
                 content="\n💻 [Provider Tool: Code Interpreter] Executing...",
                 source="response_api",
             )
+        elif chunk_type == "response.code_interpreter_call_code.delta":
+            code_delta = getattr(chunk, "delta", "") or ""
+            if code_delta:
+                log_stream_chunk("backend.response", "code_executed", code_delta, agent_id)
+                return TextStreamChunk(
+                    type=ChunkType.CONTENT,
+                    content=f"💻 [Code Executed]\n```\n{code_delta}\n```\n",
+                    source="response_api",
+                )
+        elif chunk_type == "response.code_interpreter_call_code.done":
+            code_text = getattr(chunk, "code", None) or getattr(chunk, "text", None) or ""
+            if code_text:
+                log_stream_chunk("backend.response", "code_executed", code_text, agent_id)
+                return TextStreamChunk(
+                    type=ChunkType.CONTENT,
+                    content=f"💻 [Code Executed]\n```\n{code_text}\n```\n",
+                    source="response_api",
+                )
         elif chunk_type == "response.code_interpreter_call.completed":
             log_stream_chunk("backend.response", "code_interpreter", "Execution completed", agent_id)
             return TextStreamChunk(
@@ -1498,6 +1756,18 @@ def _process_stream_chunk(self, chunk, agent_id) -> TextStreamChunk | StreamChun
                                 content=f"\n🔍 [Search Query] '{search_query}'\n",
                                 source="response_api",
                             )
+                elif hasattr(chunk.item, "type") and chunk.item.type == "x_search_call":
+                    if hasattr(chunk.item, "action") and ("query" in chunk.item.action):
+                        search_query = chunk.item.action["query"]
+                        if search_query:
+                            log_stream_chunk("backend.response", "x_search_query", search_query, agent_id)
+                            return TextStreamChunk(
+                                type=ChunkType.CONTENT,
+                                content=f"\n🔎 [X Search Query] '{search_query}'\n",
+                                source="response_api",
+                            )
+                elif self._is_xai_x_search_custom_tool_item(chunk.item):
+                    return self._format_xai_x_search_custom_tool_chunk(chunk.item, agent_id)
                 elif hasattr(chunk.item, "type") and chunk.item.type == "code_interpreter_call":
                     if hasattr(chunk.item, "code") and chunk.item.code:
                         # Format code as a proper code block - don't assume language
@@ -1661,6 +1931,23 @@ def _process_stream_chunk(self, chunk, agent_id) -> TextStreamChunk | StreamChun
                                     content=content,
                                     source="response_api",
                                 )
+                        elif item.get("type") == "x_search_call":
+                            status = item.get("status", "unknown")
+                            query = item.get("action", {}).get("query", "")
+                            results = item.get("results")
+
+                            if query:
+                                content = f"\n🔧 X Search [{status.title()}]: {query}"
+                                if results:
+                                    content += f" → Found {len(results)} results"
+                                log_stream_chunk("backend.response", "x_search_result", content, agent_id)
+                                return TextStreamChunk(
+                                    type=ChunkType.CONTENT,
+                                    content=content,
+                                    source="response_api",
+                                )
+                        elif self._is_xai_x_search_custom_tool_item(item):
+                            return self._format_xai_x_search_custom_tool_chunk(item, agent_id, strip_newlines=True)
                         elif item.get("type") == "image_generation_call":
                             # Image generation result in completed response
                             status = item.get("status", "unknown")
@@ -1758,12 +2045,19 @@ def extract_tool_result_content(self, tool_result_message: dict[str, Any]) -> st
         """Extract content from OpenAI Responses API tool result message."""
         return tool_result_message.get("output", "")
 
-    async def _create_response_stream(self, api_params, client, ws_transport=None):
+    async def _create_response_stream(self, api_params, client, ws_transport=None, agent_id=None):
         """Create a response stream via HTTP or websocket transport."""
         if ws_transport is not None and ws_transport.is_connected:
             logger.debug("[WebSocket] Sending response.create via WebSocket")
             return self._ws_event_stream(ws_transport, api_params)
-        return await client.responses.create(**api_params)
+
+        async def _make_api_call():
+            return await client.responses.create(**api_params)
+
+        return await self.circuit_breaker.call_with_retry(
+            _make_api_call,
+            agent_id=agent_id,
+        )
 
     async def _ws_event_stream(self, ws_transport, api_params):
         """Wrap websocket JSON events as objects matching SDK stream chunks."""
diff --git a/massgen/chat_agent.py b/massgen/chat_agent.py
index bb26bcd9e..d8eccf09c 100644
--- a/massgen/chat_agent.py
+++ b/massgen/chat_agent.py
@@ -1033,7 +1033,8 @@ def get_status(self) -> dict[str, Any]:
                 "config": self.config.to_dict(),
                 "capabilities": {
                     "web_search": self.config.backend_params.get("enable_web_search", False),
-                    "code_execution": self.config.backend_params.get("enable_code_interpreter", False),
+                    "x_search": self.config.backend_params.get("enable_x_search", False),
+                    "code_execution": (self.config.backend_params.get("enable_code_execution", False) or self.config.backend_params.get("enable_code_interpreter", False)),
                 },
             },
         )
diff --git a/massgen/config_validator.py b/massgen/config_validator.py
index 722cf7dab..ae451d203 100644
--- a/massgen/config_validator.py
+++ b/massgen/config_validator.py
@@ -614,6 +614,7 @@ def _validate_backend(self, backend_config: dict[str, Any], location: str, resul
         # Validate boolean fields
         boolean_fields = [
             "enable_web_search",
+            "enable_x_search",
             "enable_code_execution",
             "enable_code_interpreter",
             "enable_programmatic_flow",
diff --git a/massgen/configs/README.md b/massgen/configs/README.md
index 02a81b8ab..fc44347a4 100644
--- a/massgen/configs/README.md
+++ b/massgen/configs/README.md
@@ -227,21 +227,27 @@ Most configurations use environment variables for API keys:so
 
 ## Release History & Examples
 
-### v0.1.71 - Latest
-**New Features:** Trace Memory & Evaluation Polish
+### v0.1.72 - Latest
+**New Features:** Grok Backend Update & Circuit Breaker Phase 2
 
 **Key Features:**
-- **Trace Analyzer Subagents**: Background trace analysis after each round — writes insights from execution traces into memory
-- **Better Evaluation Criteria**: Improved criteria generation for higher-quality, more opinionated output
-- **System Prompt Tuning**: Adjusted system prompts for better agent performance across coordination rounds
-- **Stability Fixes**: Fixed final injection, eval criteria GPT pre-collab, trace analyzer launch, and memory handling
+- **Grok Backend Update**: Updated Grok backend with latest improvements
+- **Circuit Breaker Phase 2**: LLM API circuit breaker extended to ChatCompletions, Response API, and Gemini backends (was Claude-only)
+- **Config Plumbing Smoke Tests**: Verify circuit breaker wiring for all backends
 
 **Try It:**
 ```bash
-pip install massgen==0.1.71
-uv run massgen --config @examples/features/trace_analyzer_background.yaml "Create an svg of an AI agent coding."
+pip install massgen==0.1.72
+uv run massgen --config @examples/providers/others/grok_x_search.yaml "Research the latest posts and news about AI agents in the last week, and summarize the key trends and insights."
 ```
 
+### v0.1.71
+**New Features:** Trace Memory & Evaluation Polish
+
+**Key Features:**
+- **Trace Analyzer Subagents**: Background trace analysis after each round — writes insights from execution traces into memory
+- **Better Evaluation Criteria**: Improved criteria generation for higher-quality, more opinionated output
+
 ### v0.1.70
 **New Features:** Evaluation Criteria Redesign
 
diff --git a/massgen/configs/features/trace_analyzer_background.yaml b/massgen/configs/features/trace_analyzer_background.yaml
index 7d5ea2ce9..dcb4f8b67 100644
--- a/massgen/configs/features/trace_analyzer_background.yaml
+++ b/massgen/configs/features/trace_analyzer_background.yaml
@@ -1,10 +1,67 @@
 # uv run massgen --config @examples/features/trace_analyzer_background.yaml "Create an svg of an AI agent coding."
 
 agents:
-- id: agent_a
+# - id: agent_a
+#   backend:
+#     type: codex
+#     model: gpt-5.4
+#     reasoning:
+#       effort: high
+#       summary: auto
+#     cwd: workspace
+#     enable_code_based_tools: true
+#     exclude_file_operation_mcps: true
+#     enable_mcp_command_line: true
+#     command_line_execution_mode: docker
+#     command_line_docker_image: ghcr.io/massgen/mcp-runtime-sudo:latest
+#     command_line_docker_network_mode: bridge
+#     command_line_docker_enable_sudo: true
+#     command_line_docker_credentials:
+#       env_file: .env
+#       env_vars_from_file:
+#       - OPENAI_API_KEY
+#       - ANTHROPIC_API_KEY
+#       - GOOGLE_API_KEY
+#       - GEMINI_API_KEY
+#       - XAI_API_KEY
+#     shared_tools_directory: shared_tools
+#     auto_discover_custom_tools: false
+#     enable_web_search: true
+# - id: agent_b
+#   backend:
+#     # type: claude_code
+#     # model: claude-sonnet-4-6
+#     # reasoning:
+#     #   type: disabled
+#     type: gemini
+#     model: gemini-3.1-pro-preview
+#     cwd: workspace
+#     enable_code_based_tools: true
+#     exclude_file_operation_mcps: true
+#     enable_mcp_command_line: true
+#     command_line_execution_mode: docker
+#     command_line_docker_image: ghcr.io/massgen/mcp-runtime-sudo:latest
+#     command_line_docker_network_mode: bridge
+#     command_line_docker_enable_sudo: true
+#     command_line_docker_credentials:
+#       env_file: .env
+#       env_vars_from_file:
+#       - OPENAI_API_KEY
+#       - ANTHROPIC_API_KEY
+#       - GOOGLE_API_KEY
+#       - GEMINI_API_KEY
+#       - XAI_API_KEY
+#     shared_tools_directory: shared_tools
+#     auto_discover_custom_tools: false
+#     enable_web_search: true
+- id: agent_c
   backend:
-    type: codex
-    model: gpt-5.4
+    # type: claude_code
+    # model: claude-sonnet-4-6
+    # reasoning:
+    #   type: disabled
+    type: grok
+    model: grok-4.20-0309-reasoning
     cwd: workspace
     enable_code_based_tools: true
     exclude_file_operation_mcps: true
@@ -20,36 +77,11 @@ agents:
       - ANTHROPIC_API_KEY
       - GOOGLE_API_KEY
       - GEMINI_API_KEY
+      - XAI_API_KEY
     shared_tools_directory: shared_tools
     auto_discover_custom_tools: false
-    enable_web_search: false
-    reasoning:
-      effort: high
-      summary: auto
-- id: agent_b
-  backend:
-    type: claude_code
-    model: claude-sonnet-4-6
-    reasoning:
-      type: disabled
-    cwd: workspace
-    enable_code_based_tools: true
-    exclude_file_operation_mcps: true
-    enable_mcp_command_line: true
-    command_line_execution_mode: docker
-    command_line_docker_image: ghcr.io/massgen/mcp-runtime-sudo:latest
-    command_line_docker_network_mode: bridge
-    command_line_docker_enable_sudo: true
-    command_line_docker_credentials:
-      env_file: .env
-      env_vars_from_file:
-      - OPENAI_API_KEY
-      - ANTHROPIC_API_KEY
-      - GOOGLE_API_KEY
-      - GEMINI_API_KEY
-    shared_tools_directory: shared_tools
-    auto_discover_custom_tools: false
-    enable_web_search: false
+    # enable_web_search: true
+    enable_x_search: true
 orchestrator:
   snapshot_storage: snapshots
   agent_temporary_workspace: temp_workspaces
@@ -75,8 +107,8 @@ orchestrator:
     task_planning_filesystem_mode: true
     enable_memory_filesystem_mode: true
     # for pre-collab eval criteria generation
-    evaluation_criteria_generator:
-       enabled: true
+    # evaluation_criteria_generator:
+    #    enabled: true
     write_mode: auto
     auto_trace_analysis: true
   context_paths: []
diff --git a/massgen/configs/providers/others/grok_single_agent.yaml b/massgen/configs/providers/others/grok_single_agent.yaml
index 55f6cce64..c6e5dffe2 100644
--- a/massgen/configs/providers/others/grok_single_agent.yaml
+++ b/massgen/configs/providers/others/grok_single_agent.yaml
@@ -3,17 +3,13 @@ agent:
   id: "grok_agent"
   backend:
     type: "grok"
-    model: "grok-3-mini"
-    enable_web_search: false
-    extra_body:
-        search_parameters:
-           mode: "auto"                 # Search strategy (see Grok API docs for valid values)
-           return_citations: true       # Include search result citations
-           max_search_results: 3        # Max number of search results to return
+    model: "grok-4"
+    enable_web_search: true
+    # enable_x_search: true
     # enable_code_execution: true
   system_message: "You are a helpful assistant"
 
 # Display configuration
 ui:
   display_type: "textual_terminal"
-  logging_enabled: true
\ No newline at end of file
+  logging_enabled: true
diff --git a/massgen/tests/test_backend_capabilities.py b/massgen/tests/test_backend_capabilities.py
index 9e28bfcc7..1ca772c35 100644
--- a/massgen/tests/test_backend_capabilities.py
+++ b/massgen/tests/test_backend_capabilities.py
@@ -72,6 +72,7 @@ def test_capability_strings_are_valid(self):
         """Ensure capability strings follow conventions."""
         valid_capabilities = {
             "web_search",
+            "x_search",
             "code_execution",
             "bash",
             "multimodal",  # Legacy - being phased out
@@ -222,6 +223,23 @@ def test_validate_code_execution_variants(self):
         errors = validate_backend_config("claude", config_claude)
         assert len(errors) == 0
 
+        # Grok now supports enable_code_execution through xAI Responses
+        config_grok = {"type": "grok", "enable_code_execution": True}
+        errors = validate_backend_config("grok", config_grok)
+        assert len(errors) == 0
+
+    def test_validate_grok_x_search(self):
+        """Grok should accept X search as a backend-specific builtin tool."""
+        config = {"type": "grok", "enable_x_search": True}
+        errors = validate_backend_config("grok", config)
+        assert len(errors) == 0
+
+    def test_validate_x_search_rejected_for_non_grok(self):
+        """X search is currently Grok-specific."""
+        config = {"type": "openai", "enable_x_search": True}
+        errors = validate_backend_config("openai", config)
+        assert any("enable_x_search" in error for error in errors)
+
     def test_validate_mcp_servers(self):
         """Test validation of MCP server configuration."""
         # Valid MCP config for backend that supports it
@@ -319,6 +337,18 @@ def test_gemini_cli_model_release_dates(self):
         for model in caps.models:
             assert model in caps.model_release_dates, f"Model {model} missing from model_release_dates"
 
+    def test_grok_capabilities(self):
+        """Grok should advertise the xAI Responses search/code surface."""
+        caps = get_capabilities("grok")
+        assert "web_search" in caps.supported_capabilities
+        assert "x_search" in caps.supported_capabilities
+        assert "code_execution" in caps.supported_capabilities
+        assert "mcp" in caps.supported_capabilities
+        assert caps.filesystem_support == "mcp"
+        assert caps.env_var == "XAI_API_KEY"
+        assert "web_search" in caps.builtin_tools
+        assert "x_search" in caps.builtin_tools
+
     def test_gemini_cli_builtin_tools_use_actual_names(self):
         """Test Gemini CLI builtin_tools use actual Gemini CLI tool names."""
         caps = get_capabilities("gemini_cli")
diff --git a/massgen/tests/test_backend_event_loop_all.py b/massgen/tests/test_backend_event_loop_all.py
index 2a35d68b7..d2c726452 100644
--- a/massgen/tests/test_backend_event_loop_all.py
+++ b/massgen/tests/test_backend_event_loop_all.py
@@ -93,21 +93,10 @@ def _factory(*args: Any, **kwargs: Any) -> _FakeOpenAIClient:
     assert created[0]._closed is True
 
 
-# ---- GrokBackend test ----
-class _FakeChatCompletions:
-    async def create(self, **kwargs: Any):
-        # Yield a single finishing chunk similar to Chat Completions
-        def _item():
-            choice = SimpleNamespace(delta=None, finish_reason="stop")
-            return SimpleNamespace(choices=[choice], usage=None)
-
-        return _FakeStreamSingleStop(_item)
-
-
 class _FakeOpenAIClientForGrok(_FakeAsyncClientBase):
     def __init__(self, *args: Any, **kwargs: Any):
         super().__init__(*args, **kwargs)
-        self.chat = SimpleNamespace(completions=_FakeChatCompletions())
+        self.responses = _FakeResponses()
 
 
 @pytest.mark.asyncio
@@ -127,7 +116,7 @@ def _factory(*args: Any, **kwargs: Any) -> _FakeOpenAIClientForGrok:
     backend = GrokBackend()
     messages = [{"role": "user", "content": "hi"}]
 
-    async for _ in backend.stream_with_tools(messages, tools=[], model="grok-2-mini"):
+    async for _ in backend.stream_with_tools(messages, tools=[], model="grok-4"):
         pass
 
     assert len(created) == 1
diff --git a/massgen/tests/test_llm_cb_backends.py b/massgen/tests/test_llm_cb_backends.py
new file mode 100644
index 000000000..fb4553a6b
--- /dev/null
+++ b/massgen/tests/test_llm_cb_backends.py
@@ -0,0 +1,567 @@
+"""Tests for LLM circuit breaker integration in ChatCompletions, Response API, and Gemini backends."""
+
+from __future__ import annotations
+
+import time
+from unittest.mock import MagicMock, patch
+
+import pytest
+
+from massgen.backend.llm_circuit_breaker import (
+    CircuitState,
+    LLMCircuitBreaker,
+    LLMCircuitBreakerConfig,
+)
+
+# ---------------------------------------------------------------------------
+# Helpers
+# ---------------------------------------------------------------------------
+
+
+class FakeAPIError(Exception):
+    """Test double for provider API errors."""
+
+
+def _make_openai_429(retry_after=None):
+    """Create a fake OpenAI-style 429 error with optional Retry-After."""
+    exc = FakeAPIError("HTTP 429")
+    exc.status_code = 429
+    response = MagicMock()
+    response.status_code = 429
+    response.headers = {"retry-after": str(retry_after)} if retry_after is not None else {}
+    exc.response = response
+    return exc
+
+
+def _enabled_config(**overrides) -> LLMCircuitBreakerConfig:
+    """Return an enabled config with fast settings for tests."""
+    defaults = {"enabled": True, "max_failures": 3, "reset_time_seconds": 1}
+    defaults.update(overrides)
+    return LLMCircuitBreakerConfig(**defaults)
+
+
+# ---------------------------------------------------------------------------
+# Config plumbing smoke tests
+# ---------------------------------------------------------------------------
+
+
+class TestChatCompletionsConfigPlumbing:
+    """ChatCompletionsBackend accepts and strips CB kwargs correctly."""
+
+    def test_accepts_cb_kwargs_without_error(self):
+        from massgen.backend.chat_completions import ChatCompletionsBackend
+
+        backend = ChatCompletionsBackend(
+            llm_circuit_breaker_enabled=False,
+            llm_circuit_breaker_max_failures=5,
+        )
+        assert isinstance(backend.circuit_breaker, LLMCircuitBreaker)
+
+    def test_cb_disabled_by_default(self):
+        from massgen.backend.chat_completions import ChatCompletionsBackend
+
+        backend = ChatCompletionsBackend()
+        assert backend.circuit_breaker.config.enabled is False
+
+    def test_cb_kwargs_stripped_from_config(self):
+        from massgen.backend.chat_completions import ChatCompletionsBackend
+
+        backend = ChatCompletionsBackend(
+            llm_circuit_breaker_enabled=True,
+            llm_circuit_breaker_max_failures=7,
+        )
+        assert "llm_circuit_breaker_enabled" not in backend.config
+        assert "llm_circuit_breaker_max_failures" not in backend.config
+
+    def test_cb_enabled_when_requested(self):
+        from massgen.backend.chat_completions import ChatCompletionsBackend
+
+        backend = ChatCompletionsBackend(llm_circuit_breaker_enabled=True)
+        assert backend.circuit_breaker.config.enabled is True
+
+    def test_cb_max_failures_propagated(self):
+        from massgen.backend.chat_completions import ChatCompletionsBackend
+
+        backend = ChatCompletionsBackend(
+            llm_circuit_breaker_enabled=True,
+            llm_circuit_breaker_max_failures=10,
+        )
+        assert backend.circuit_breaker.config.max_failures == 10
+
+
+class TestResponseBackendConfigPlumbing:
+    """ResponseBackend accepts and strips CB kwargs correctly."""
+
+    def test_accepts_cb_kwargs_without_error(self):
+        from massgen.backend.response import ResponseBackend
+
+        backend = ResponseBackend(
+            llm_circuit_breaker_enabled=False,
+            llm_circuit_breaker_max_failures=5,
+        )
+        assert isinstance(backend.circuit_breaker, LLMCircuitBreaker)
+
+    def test_cb_disabled_by_default(self):
+        from massgen.backend.response import ResponseBackend
+
+        backend = ResponseBackend()
+        assert backend.circuit_breaker.config.enabled is False
+
+    def test_cb_kwargs_stripped_from_config(self):
+        from massgen.backend.response import ResponseBackend
+
+        backend = ResponseBackend(
+            llm_circuit_breaker_enabled=True,
+            llm_circuit_breaker_max_failures=7,
+        )
+        assert "llm_circuit_breaker_enabled" not in backend.config
+        assert "llm_circuit_breaker_max_failures" not in backend.config
+
+    def test_cb_enabled_when_requested(self):
+        from massgen.backend.response import ResponseBackend
+
+        backend = ResponseBackend(llm_circuit_breaker_enabled=True)
+        assert backend.circuit_breaker.config.enabled is True
+
+    def test_cb_max_failures_propagated(self):
+        from massgen.backend.response import ResponseBackend
+
+        backend = ResponseBackend(
+            llm_circuit_breaker_enabled=True,
+            llm_circuit_breaker_max_failures=10,
+        )
+        assert backend.circuit_breaker.config.max_failures == 10
+
+
+class TestGeminiBackendConfigPlumbing:
+    """GeminiBackend accepts and strips CB kwargs correctly."""
+
+    def test_accepts_cb_kwargs_without_error(self):
+        from massgen.backend.gemini import GeminiBackend
+
+        backend = GeminiBackend(
+            llm_circuit_breaker_enabled=False,
+            llm_circuit_breaker_max_failures=5,
+        )
+        assert isinstance(backend.circuit_breaker, LLMCircuitBreaker)
+
+    def test_cb_disabled_by_default(self):
+        from massgen.backend.gemini import GeminiBackend
+
+        backend = GeminiBackend()
+        assert backend.circuit_breaker.config.enabled is False
+
+    def test_cb_kwargs_stripped_from_config(self):
+        from massgen.backend.gemini import GeminiBackend
+
+        backend = GeminiBackend(
+            llm_circuit_breaker_enabled=True,
+            llm_circuit_breaker_max_failures=7,
+        )
+        assert "llm_circuit_breaker_enabled" not in backend.config
+        assert "llm_circuit_breaker_max_failures" not in backend.config
+
+    def test_cb_enabled_when_requested(self):
+        from massgen.backend.gemini import GeminiBackend
+
+        backend = GeminiBackend(llm_circuit_breaker_enabled=True)
+        assert backend.circuit_breaker.config.enabled is True
+
+    def test_cb_max_failures_propagated(self):
+        from massgen.backend.gemini import GeminiBackend
+
+        backend = GeminiBackend(
+            llm_circuit_breaker_enabled=True,
+            llm_circuit_breaker_max_failures=10,
+        )
+        assert backend.circuit_breaker.config.max_failures == 10
+
+
+# ---------------------------------------------------------------------------
+# State transitions via CB instance on backends
+# ---------------------------------------------------------------------------
+
+
+class TestBackendCBStateTransitions:
+    """Verify CB instance on backend transitions correctly through states."""
+
+    def _make_cb(self, **cfg_overrides) -> LLMCircuitBreaker:
+        """Create a standalone enabled CB for state-transition tests."""
+        return LLMCircuitBreaker(config=_enabled_config(**cfg_overrides))
+
+    def test_closed_to_open_after_max_failures(self):
+        cb = self._make_cb(max_failures=3)
+        assert cb.state == CircuitState.CLOSED
+        for _ in range(3):
+            cb.record_failure(error_type="test")
+        assert cb.state == CircuitState.OPEN
+
+    def test_open_to_half_open_after_reset_time(self):
+        cb = self._make_cb(max_failures=1, reset_time_seconds=1)
+        cb.record_failure()
+        assert cb.state == CircuitState.OPEN
+        time.sleep(1.1)
+        # should_block triggers the OPEN -> HALF_OPEN transition
+        blocked = cb.should_block()
+        assert blocked is False
+        assert cb.state == CircuitState.HALF_OPEN
+
+    def test_half_open_to_closed_on_success(self):
+        cb = self._make_cb(max_failures=1, reset_time_seconds=1)
+        cb.record_failure()
+        time.sleep(1.1)
+        cb.should_block()  # OPEN -> HALF_OPEN
+        assert cb.state == CircuitState.HALF_OPEN
+        cb.record_success()
+        assert cb.state == CircuitState.CLOSED
+        assert cb.failure_count == 0
+
+    def test_half_open_to_open_on_failure(self):
+        cb = self._make_cb(max_failures=1, reset_time_seconds=1)
+        cb.record_failure()
+        time.sleep(1.1)
+        cb.should_block()  # OPEN -> HALF_OPEN
+        cb.record_failure()
+        assert cb.state == CircuitState.OPEN
+
+    def test_chatcompletions_backend_cb_state_transition(self):
+        """CB on ChatCompletionsBackend transitions CLOSED -> OPEN -> HALF_OPEN -> CLOSED."""
+        from massgen.backend.chat_completions import ChatCompletionsBackend
+
+        backend = ChatCompletionsBackend(
+            llm_circuit_breaker_enabled=True,
+            llm_circuit_breaker_max_failures=2,
+            llm_circuit_breaker_reset_time_seconds=1,
+        )
+        cb = backend.circuit_breaker
+        assert cb.state == CircuitState.CLOSED
+
+        cb.record_failure()
+        cb.record_failure()
+        assert cb.state == CircuitState.OPEN
+
+        time.sleep(1.1)
+        cb.should_block()
+        assert cb.state == CircuitState.HALF_OPEN
+
+        cb.record_success()
+        assert cb.state == CircuitState.CLOSED
+
+
+# ---------------------------------------------------------------------------
+# Disabled bypass
+# ---------------------------------------------------------------------------
+
+
+class TestDisabledBypassViaBackend:
+    """When enabled=False (default), should_block() always returns False."""
+
+    def test_chatcompletions_disabled_never_blocks(self):
+        from massgen.backend.chat_completions import ChatCompletionsBackend
+
+        backend = ChatCompletionsBackend()
+        cb = backend.circuit_breaker
+        # Force many failures -- should have no effect when disabled
+        for _ in range(100):
+            cb.record_failure()
+        assert cb.should_block() is False
+        assert cb.failure_count == 0
+
+    def test_response_disabled_never_blocks(self):
+        from massgen.backend.response import ResponseBackend
+
+        backend = ResponseBackend()
+        cb = backend.circuit_breaker
+        for _ in range(100):
+            cb.record_failure()
+        assert cb.should_block() is False
+        assert cb.failure_count == 0
+
+    def test_gemini_disabled_never_blocks(self):
+        from massgen.backend.gemini import GeminiBackend
+
+        backend = GeminiBackend()
+        cb = backend.circuit_breaker
+        for _ in range(100):
+            cb.record_failure()
+        assert cb.should_block() is False
+        assert cb.failure_count == 0
+
+
+# ---------------------------------------------------------------------------
+# 429 classification via call_with_retry (LLMCircuitBreaker directly)
+# ---------------------------------------------------------------------------
+
+
+class TestCallWithRetry429Classifications:
+    """429 WAIT / STOP / CAP classification through call_with_retry."""
+
+    @pytest.mark.asyncio
+    async def test_wait_short_retry_after_retries_no_failure_increment(self):
+        """WAIT: short Retry-After -> retry, failure counter stays 0."""
+        cb = LLMCircuitBreaker(config=_enabled_config(retry_after_threshold_seconds=60))
+        call_count = 0
+
+        async def api_call():
+            nonlocal call_count
+            call_count += 1
+            if call_count == 1:
+                raise _make_openai_429(retry_after=10)
+            return "ok"
+
+        with patch("massgen.backend.llm_circuit_breaker.asyncio.sleep") as mock_sleep:
+            mock_sleep.return_value = None
+            result = await cb.call_with_retry(api_call)
+
+        assert result == "ok"
+        assert call_count == 2
+        # WAIT honors Retry-After value
+        mock_sleep.assert_awaited_once_with(10)
+        # WAIT does not increment failure counter
+        assert cb.failure_count == 0
+
+    @pytest.mark.asyncio
+    async def test_stop_long_retry_after_forces_cb_open(self):
+        """STOP: long Retry-After -> CB forced open, no retry."""
+        cb = LLMCircuitBreaker(config=_enabled_config(retry_after_threshold_seconds=60))
+        call_count = 0
+
+        async def api_call():
+            nonlocal call_count
+            call_count += 1
+            raise _make_openai_429(retry_after=600)
+
+        with pytest.raises(FakeAPIError):
+            await cb.call_with_retry(api_call)
+
+        # Only one attempt -- STOP does not retry
+        assert call_count == 1
+        assert cb.state == CircuitState.OPEN
+
+    @pytest.mark.asyncio
+    async def test_cap_no_retry_after_increments_failure_counter(self):
+        """CAP: no Retry-After -> failure counter increments."""
+        cb = LLMCircuitBreaker(config=_enabled_config(max_failures=10))
+
+        async def api_call():
+            raise _make_openai_429(retry_after=None)  # CAP
+
+        with patch("massgen.backend.llm_circuit_breaker.asyncio.sleep") as mock_sleep:
+            mock_sleep.return_value = None
+            with pytest.raises(FakeAPIError):
+                await cb.call_with_retry(api_call, max_retries=3)
+
+        # 3 CAP failures recorded
+        assert cb.failure_count == 3
+
+    @pytest.mark.asyncio
+    async def test_cap_retries_before_exhaustion(self):
+        """CAP: backoff-retry succeeds on second attempt."""
+        cb = LLMCircuitBreaker(config=_enabled_config(max_failures=10))
+        call_count = 0
+
+        async def api_call():
+            nonlocal call_count
+            call_count += 1
+            if call_count == 1:
+                raise _make_openai_429(retry_after=None)
+            return "recovered"
+
+        with patch("massgen.backend.llm_circuit_breaker.asyncio.sleep") as mock_sleep:
+            mock_sleep.return_value = None
+            result = await cb.call_with_retry(api_call, max_retries=3)
+
+        assert result == "recovered"
+        assert call_count == 2
+        # record_success resets the counter
+        assert cb.failure_count == 0
+
+
+# ---------------------------------------------------------------------------
+# Gemini 503 handling
+# ---------------------------------------------------------------------------
+
+
+class TestGemini503Handling:
+    """record_failure('exhausted_503', ...) increments counter; after max_failures, blocks."""
+
+    def test_503_increments_failure_counter(self):
+        cb = LLMCircuitBreaker(config=_enabled_config(max_failures=5))
+        cb.record_failure("exhausted_503", "Service Unavailable")
+        assert cb.failure_count == 1
+        assert cb.state == CircuitState.CLOSED
+
+    def test_503_repeated_failures_open_cb(self):
+        cb = LLMCircuitBreaker(config=_enabled_config(max_failures=3))
+        for _ in range(3):
+            cb.record_failure("exhausted_503", "Service Unavailable")
+        assert cb.state == CircuitState.OPEN
+        assert cb.should_block() is True
+
+    @pytest.mark.asyncio
+    async def test_503_via_call_with_retry_records_failure(self):
+        """HTTP 503 in retryable_status_codes increments failure counter."""
+        cb = LLMCircuitBreaker(config=_enabled_config(max_failures=10))
+
+        exc = FakeAPIError("HTTP 503")
+        exc.status_code = 503
+
+        async def api_call():
+            raise exc
+
+        with patch("massgen.backend.llm_circuit_breaker.asyncio.sleep") as mock_sleep:
+            mock_sleep.return_value = None
+            with pytest.raises(FakeAPIError):
+                await cb.call_with_retry(api_call, max_retries=2)
+
+        assert cb.failure_count == 2
+
+    def test_503_does_not_increment_when_disabled(self):
+        """When CB is disabled, record_failure is a no-op."""
+        cb = LLMCircuitBreaker(config=LLMCircuitBreakerConfig(enabled=False))
+        for _ in range(100):
+            cb.record_failure("exhausted_503", "Service Unavailable")
+        assert cb.should_block() is False
+        assert cb.state == CircuitState.CLOSED
+
+
+# ---------------------------------------------------------------------------
+# Config validation
+# ---------------------------------------------------------------------------
+
+
+class TestCBConfigValidation:
+    """Backends reject invalid CB config at construction time."""
+
+    def test_chatcompletions_rejects_negative_max_failures(self):
+        from massgen.backend.chat_completions import ChatCompletionsBackend
+
+        with pytest.raises(ValueError, match="max_failures"):
+            ChatCompletionsBackend(
+                llm_circuit_breaker_enabled=True,
+                llm_circuit_breaker_max_failures=-1,
+            )
+
+    def test_response_rejects_negative_max_failures(self):
+        from massgen.backend.response import ResponseBackend
+
+        with pytest.raises(ValueError, match="max_failures"):
+            ResponseBackend(
+                llm_circuit_breaker_enabled=True,
+                llm_circuit_breaker_max_failures=-1,
+            )
+
+    def test_gemini_rejects_negative_max_failures(self):
+        from massgen.backend.gemini import GeminiBackend
+
+        with pytest.raises(ValueError, match="max_failures"):
+            GeminiBackend(
+                llm_circuit_breaker_enabled=True,
+                llm_circuit_breaker_max_failures=-1,
+            )
+
+    def test_config_dataclass_rejects_max_failures_zero(self):
+        with pytest.raises(ValueError, match="max_failures"):
+            LLMCircuitBreakerConfig(max_failures=0)
+
+    def test_config_dataclass_rejects_max_failures_negative(self):
+        with pytest.raises(ValueError, match="max_failures"):
+            LLMCircuitBreakerConfig(max_failures=-1)
+
+    def test_config_dataclass_rejects_invalid_reset_time(self):
+        with pytest.raises(ValueError, match="reset_time_seconds"):
+            LLMCircuitBreakerConfig(reset_time_seconds=0)
+
+    def test_chatcompletions_rejects_invalid_reset_time(self):
+        from massgen.backend.chat_completions import ChatCompletionsBackend
+
+        with pytest.raises(ValueError, match="reset_time_seconds"):
+            ChatCompletionsBackend(
+                llm_circuit_breaker_enabled=True,
+                llm_circuit_breaker_reset_time_seconds=0,
+            )
+
+    def test_response_rejects_invalid_reset_time(self):
+        from massgen.backend.response import ResponseBackend
+
+        with pytest.raises(ValueError, match="reset_time_seconds"):
+            ResponseBackend(
+                llm_circuit_breaker_enabled=True,
+                llm_circuit_breaker_reset_time_seconds=0,
+            )
+
+    def test_gemini_rejects_invalid_reset_time(self):
+        from massgen.backend.gemini import GeminiBackend
+
+        with pytest.raises(ValueError, match="reset_time_seconds"):
+            GeminiBackend(
+                llm_circuit_breaker_enabled=True,
+                llm_circuit_breaker_reset_time_seconds=0,
+            )
+
+
+# ---------------------------------------------------------------------------
+# Integration: CB OPEN raises CircuitBreakerOpenError through backend
+# ---------------------------------------------------------------------------
+
+
+class TestCBOpenRaisesErrorThroughBackend:
+    """When CB is OPEN, backend API calls raise CircuitBreakerOpenError."""
+
+    @pytest.mark.asyncio
+    async def test_chatcompletions_raises_when_cb_open(self):
+        from massgen.backend.chat_completions import ChatCompletionsBackend
+        from massgen.backend.llm_circuit_breaker import CircuitBreakerOpenError
+
+        backend = ChatCompletionsBackend(
+            llm_circuit_breaker_enabled=True,
+            llm_circuit_breaker_max_failures=1,
+        )
+        # Force CB open
+        backend.circuit_breaker.record_failure("test", "force open")
+        assert backend.circuit_breaker.state == CircuitState.OPEN
+
+        # call_with_retry should raise immediately
+        async def _dummy():
+            return "should not reach"
+
+        with pytest.raises(CircuitBreakerOpenError):
+            await backend.circuit_breaker.call_with_retry(_dummy)
+
+    @pytest.mark.asyncio
+    async def test_response_raises_when_cb_open(self):
+        from massgen.backend.llm_circuit_breaker import CircuitBreakerOpenError
+        from massgen.backend.response import ResponseBackend
+
+        backend = ResponseBackend(
+            llm_circuit_breaker_enabled=True,
+            llm_circuit_breaker_max_failures=1,
+        )
+        backend.circuit_breaker.record_failure("test", "force open")
+        assert backend.circuit_breaker.state == CircuitState.OPEN
+
+        async def _dummy():
+            return "should not reach"
+
+        with pytest.raises(CircuitBreakerOpenError):
+            await backend.circuit_breaker.call_with_retry(_dummy)
+
+    @pytest.mark.asyncio
+    async def test_gemini_raises_when_cb_open(self):
+        """Gemini uses should_block() gate -- verify CB blocks via call_with_retry."""
+        from massgen.backend.gemini import GeminiBackend
+        from massgen.backend.llm_circuit_breaker import CircuitBreakerOpenError
+
+        backend = GeminiBackend(
+            llm_circuit_breaker_enabled=True,
+            llm_circuit_breaker_max_failures=1,
+        )
+        backend.circuit_breaker.record_failure("test", "force open")
+        assert backend.circuit_breaker.state == CircuitState.OPEN
+
+        # Gemini uses should_block() directly, but call_with_retry also blocks
+        async def _dummy():
+            return "should not reach"
+
+        with pytest.raises(CircuitBreakerOpenError, match="Circuit breaker is open"):
+            await backend.circuit_breaker.call_with_retry(_dummy)