diff --git a/plugins/pr-review/README.md b/plugins/pr-review/README.md
index 47f60f3f..437beb17 100644
--- a/plugins/pr-review/README.md
+++ b/plugins/pr-review/README.md
@@ -214,6 +214,8 @@ PR reviews are automatically triggered when:
 | `acp-command` | Yes for `acp` mode | `''` | Command used to start the ACP server. The command must already be available in the runner environment or be runnable through a package manager. Examples: `npx -y @zed-industries/codex-acp@0.12.0`, `codex-acp`, `claude-agent-acp`, `npx -y @agentclientprotocol/claude-agent-acp`. |
 | `acp-prompt-timeout` | No | `'1800'` | Timeout in seconds for one ACP prompt turn |
 | `llm-base-url` | No | `''` | Custom LLM endpoint URL |
+| `llm-extra-body` | No | `''` | JSON object of extra request-body fields forwarded to the LLM provider, e.g. `{"enable_thinking": false}` to turn off default thinking on Qwen models. Ignored in ACP mode. |
+| `max-iterations` | No | `''` | Cap on agent iterations (model call + tool action) per review run. Bounds runaway exploration; empty uses the SDK default. Ignored in ACP mode. |
 | `review-style` | No | `roasted` | **[DEPRECATED]** Previously chose between `standard` and `roasted` review styles. Now ignored — the styles have been merged into a single unified skill. |
 | `require-evidence` | No | `'false'` | Require the reviewer to enforce an `Evidence` section in the PR description with end-to-end proof: screenshots/videos for frontend work, commands and runtime output for backend or scripts, and an agent conversation link when applicable. Test output alone does not qualify. |
 | `use-sub-agents` | No | `'false'` | Enable sub-agent delegation for file-level reviews in `openhands` mode. The main agent acts as a coordinator that delegates per-file review work to `file_reviewer` sub-agents via the SDK TaskToolSet, then consolidates findings into a single PR review. Useful for large PRs with many changed files. **Disabled by default** due to high token costs and potential timeouts (see [#208](https://github.com/OpenHands/extensions/issues/208)). Set to `'true'` to opt in. Ignored in ACP mode. |
diff --git a/plugins/pr-review/action.yml b/plugins/pr-review/action.yml
index e0c769e9..57a1cca5 100644
--- a/plugins/pr-review/action.yml
+++ b/plugins/pr-review/action.yml
@@ -41,6 +41,22 @@ inputs:
         description: LLM base URL (optional, for custom LLM endpoints)
         required: false
         default: ''
+    llm-extra-body:
+        description: >
+            JSON object of extra request-body fields forwarded to the LLM
+            provider (optional). Use it for provider-specific switches such as
+            '{"enable_thinking": false}' or '{"thinking_budget": 2048}' on
+            models that think by default. Ignored in ACP mode.
+        required: false
+        default: ''
+    max-iterations:
+        description: >
+            Maximum agent iterations (model call + tool action) for the review
+            run (optional). Bounds runaway exploration so the job ends with a
+            clear error instead of hitting the workflow timeout. Empty uses the
+            SDK default. Ignored in ACP mode.
+        required: false
+        default: ''
     review-style:
         description: "[DEPRECATED] Previously chose between 'standard' and 'roasted' review styles. These have been merged into a single code-review skill. The input is kept for backward compatibility but no longer changes behavior. Will be removed in a future version."
         required: false
@@ -130,6 +146,14 @@ runs:
               path: pr-repo
               submodules: recursive
 
+        # fetch-depth 0 brings the base branch in as origin/<base> only. The reviewer's
+        # natural `git diff <base>` otherwise fails with "unknown revision" and burns steps.
+        - name: Create local base branch
+          shell: bash
+          env:
+              BASE_REF: ${{ github.event.pull_request.base.ref }}
+          run: git -C pr-repo branch "$BASE_REF" "origin/$BASE_REF" || true
+
         - name: Set up Python
           uses: actions/setup-python@v5
           with:
@@ -258,6 +282,8 @@ runs:
               ACP_PROMPT_TIMEOUT: ${{ inputs.acp-prompt-timeout }}
               LLM_MODEL: ${{ steps.select-model.outputs.selected_model }}
               LLM_BASE_URL: ${{ inputs.llm-base-url }}
+              LLM_EXTRA_BODY: ${{ inputs.llm-extra-body }}
+              MAX_ITERATIONS: ${{ inputs.max-iterations }}
               REVIEW_STYLE: ${{ inputs.review-style }}
               REQUIRE_EVIDENCE: ${{ inputs.require-evidence }}
               COLLECT_FEEDBACK: ${{ inputs.collect-feedback }}
diff --git a/plugins/pr-review/scripts/agent_script.py b/plugins/pr-review/scripts/agent_script.py
index c8104773..2abdfb70 100644
--- a/plugins/pr-review/scripts/agent_script.py
+++ b/plugins/pr-review/scripts/agent_script.py
@@ -27,6 +27,8 @@
     LLM_API_KEY: API key for the LLM (required for OpenHands agent kind)
     LLM_MODEL: Language model to use (default: anthropic/claude-sonnet-4-5-20250929)
     LLM_BASE_URL: Optional base URL for LLM API
+    LLM_EXTRA_BODY: Optional JSON object of extra request-body fields for the LLM
+    MAX_ITERATIONS: Optional cap on agent iterations per review run
     GITHUB_TOKEN: GitHub token for API access (required)
     PR_NUMBER: Pull request number (required)
     PR_TITLE: Pull request title (required)
@@ -893,6 +895,30 @@ def validate_environment() -> dict[str, Any]:
         sys.exit(1)
 
     api_key = os.getenv("LLM_API_KEY")
+
+    extra_body: dict[str, Any] = {}
+    raw_extra_body = os.getenv("LLM_EXTRA_BODY", "")
+    if raw_extra_body:
+        try:
+            extra_body = json.loads(raw_extra_body)
+        except json.JSONDecodeError as exc:
+            logger.error(f"LLM_EXTRA_BODY is not valid JSON: {exc}")
+            sys.exit(1)
+        if not isinstance(extra_body, dict):
+            logger.error("LLM_EXTRA_BODY must be a JSON object")
+            sys.exit(1)
+
+    max_iterations: int | None = None
+    raw_max_iterations = os.getenv("MAX_ITERATIONS", "")
+    if raw_max_iterations:
+        try:
+            max_iterations = int(raw_max_iterations)
+        except ValueError:
+            logger.error("MAX_ITERATIONS must be an integer")
+            sys.exit(1)
+        if max_iterations < 1:
+            logger.error("MAX_ITERATIONS must be at least 1")
+            sys.exit(1)
     if agent_kind == "openhands" and not api_key:
         logger.error(
             "LLM_API_KEY is required when AGENT_KIND is 'openhands'"
@@ -925,6 +951,8 @@ def validate_environment() -> dict[str, Any]:
         "github_token": os.getenv("GITHUB_TOKEN"),
         "model": os.getenv("LLM_MODEL", "anthropic/claude-sonnet-4-5-20250929"),
         "base_url": os.getenv("LLM_BASE_URL"),
+        "extra_body": extra_body,
+        "max_iterations": max_iterations,
         "require_evidence": _get_bool_env("REQUIRE_EVIDENCE"),
         "collect_feedback": _get_bool_env("COLLECT_FEEDBACK"),
         "review_run_url": os.getenv("REVIEW_RUN_URL", ""),
@@ -1076,6 +1104,8 @@ def create_conversation(
     }
     if config["base_url"]:
         llm_config["base_url"] = config["base_url"]
+    if config["extra_body"]:
+        llm_config["litellm_extra_body"] = config["extra_body"]
 
     llm = LLM(**llm_config)
 
@@ -1114,6 +1144,8 @@ def create_conversation(
         conversation_kwargs["visualizer"] = DelegationVisualizer(
             name="PR Review Coordinator"
         )
+    if config["max_iterations"]:
+        conversation_kwargs["max_iteration_per_run"] = config["max_iterations"]
 
     return Conversation(**conversation_kwargs)