diff --git a/.gitignore b/.gitignore
index d36f2b71..ecd28ba0 100644
--- a/.gitignore
+++ b/.gitignore
@@ -47,6 +47,7 @@ tests/*
 !tests/test_branding.py
 !tests/test_artifact_analyzer.py
 !tests/test_scala_analyzer.py
+!tests/test_agent_limits.py
 
 # Jupyter
 *.ipynb
diff --git a/codewiki/cli/adapters/doc_generator.py b/codewiki/cli/adapters/doc_generator.py
index 92c066dd..07f425fc 100644
--- a/codewiki/cli/adapters/doc_generator.py
+++ b/codewiki/cli/adapters/doc_generator.py
@@ -145,6 +145,8 @@ def generate(self) -> DocumentationJob:
                 max_token_per_leaf_module=self.config.get("max_token_per_leaf_module", 16000),
                 max_leaf_nodes_per_cluster=self.config.get("max_leaf_nodes_per_cluster", 600),
                 max_depth=self.config.get("max_depth", 2),
+                request_limit=self.config.get("request_limit", 100),
+                agent_retries=self.config.get("agent_retries", 3),
                 agent_instructions=self.config.get("agent_instructions"),
                 use_gitignore=self.config.get("use_gitignore", True),
                 prompt_caching=self.config.get("prompt_caching", True),
diff --git a/codewiki/cli/commands/config.py b/codewiki/cli/commands/config.py
index 17bed6cd..f0e6d18b 100644
--- a/codewiki/cli/commands/config.py
+++ b/codewiki/cli/commands/config.py
@@ -51,6 +51,12 @@ def config_group():
 @click.option(
     "--max-depth", type=int, help="Maximum depth for hierarchical decomposition (default: 2)"
 )
+@click.option(
+    "--request-limit", type=int, help="Maximum model requests per agent run (default: 100)"
+)
+@click.option(
+    "--agent-retries", type=int, help="Retries for a failing agent tool call (default: 3)"
+)
 @click.option(
     "--provider",
     type=click.Choice(
@@ -97,6 +103,8 @@ def config_set(
     max_token_per_module: Optional[int],
     max_token_per_leaf_module: Optional[int],
     max_depth: Optional[int],
+    request_limit: Optional[int] = None,
+    agent_retries: Optional[int] = None,
     provider: Optional[str] = None,
     aws_region: Optional[str] = None,
     api_version: Optional[str] = None,
@@ -166,6 +174,8 @@ def config_set(
                 max_token_per_module,
                 max_token_per_leaf_module,
                 max_depth,
+                request_limit is not None,
+                agent_retries is not None,
                 provider,
                 aws_region,
                 api_version,
@@ -226,6 +236,16 @@ def config_set(
                 raise ConfigurationError("max_depth must be a positive integer")
             validated_data["max_depth"] = max_depth
 
+        if request_limit is not None:
+            if request_limit < 1:
+                raise ConfigurationError("request_limit must be a positive integer")
+            validated_data["request_limit"] = request_limit
+
+        if agent_retries is not None:
+            if agent_retries < 0:
+                raise ConfigurationError("agent_retries must be zero or a positive integer")
+            validated_data["agent_retries"] = agent_retries
+
         if provider is not None:
             validated_data["provider"] = provider
 
@@ -258,6 +278,8 @@ def config_set(
             max_token_per_module=validated_data.get("max_token_per_module"),
             max_token_per_leaf_module=validated_data.get("max_token_per_leaf_module"),
             max_depth=validated_data.get("max_depth"),
+            request_limit=validated_data.get("request_limit"),
+            agent_retries=validated_data.get("agent_retries"),
             provider=validated_data.get("provider"),
             aws_region=validated_data.get("aws_region"),
             api_version=validated_data.get("api_version"),
@@ -311,6 +333,12 @@ def config_set(
         if max_depth:
             click.secho(f"✓ Max depth: {max_depth}", fg="green")
 
+        if request_limit:
+            click.secho(f"✓ Request limit: {request_limit}", fg="green")
+
+        if agent_retries is not None:
+            click.secho(f"✓ Agent retries: {agent_retries}", fg="green")
+
         if provider:
             click.secho(f"✓ Provider: {provider}", fg="green")
 
@@ -384,6 +412,8 @@ def config_show(output_json: bool):
                 "max_token_per_module": config.max_token_per_module if config else 36369,
                 "max_token_per_leaf_module": config.max_token_per_leaf_module if config else 16000,
                 "max_depth": config.max_depth if config else 2,
+                "request_limit": config.request_limit if config else 100,
+                "agent_retries": config.agent_retries if config else 3,
                 "use_gitignore": config.use_gitignore if config else True,
                 "prompt_caching": config.prompt_caching if config else True,
                 "agent_instructions": config.agent_instructions.to_dict()
@@ -450,6 +480,8 @@ def config_show(output_json: bool):
             click.secho("Decomposition Settings", fg="cyan", bold=True)
             if config:
                 click.echo(f"  Max Depth:               {config.max_depth}")
+                click.echo(f"  Request Limit:           {config.request_limit}")
+                click.echo(f"  Agent Retries:           {config.agent_retries}")
                 click.echo(f"  Use Gitignore:           {config.use_gitignore}")
 
             click.echo()
diff --git a/codewiki/cli/commands/generate.py b/codewiki/cli/commands/generate.py
index 2e476c86..b0df3a84 100644
--- a/codewiki/cli/commands/generate.py
+++ b/codewiki/cli/commands/generate.py
@@ -296,6 +296,18 @@ def _find_affected(tree, parent_names=None):
     default=None,
     help="Maximum depth for hierarchical decomposition (overrides config)",
 )
+@click.option(
+    "--request-limit",
+    type=click.IntRange(min=1),
+    default=None,
+    help="Maximum model requests per agent run (overrides config)",
+)
+@click.option(
+    "--agent-retries",
+    type=click.IntRange(min=0),
+    default=None,
+    help="Retries for a failing agent tool call (overrides config)",
+)
 @click.option(
     "--prompt-caching/--no-prompt-caching",
     default=None,
@@ -398,6 +410,8 @@ def generate_command(
     max_token_per_module: int | None,
     max_token_per_leaf_module: int | None,
     max_depth: int | None,
+    request_limit: int | None,
+    agent_retries: int | None,
     prompt_caching: bool | None,
     artifacts: bool = True,
     artifact_token_budget: int = 200_000,
@@ -717,6 +731,13 @@ def generate_command(
                 else config.max_token_per_leaf_module,
                 # Max depth setting (runtime override takes precedence)
                 "max_depth": max_depth if max_depth is not None else config.max_depth,
+                # Agent run limits (runtime override takes precedence)
+                "request_limit": request_limit
+                if request_limit is not None
+                else config.request_limit,
+                "agent_retries": agent_retries
+                if agent_retries is not None
+                else config.agent_retries,
                 # Gitignore setting (runtime override takes precedence)
                 "use_gitignore": use_gitignore
                 if use_gitignore is not None
diff --git a/codewiki/cli/config_manager.py b/codewiki/cli/config_manager.py
index 3cbadcff..15f9f820 100644
--- a/codewiki/cli/config_manager.py
+++ b/codewiki/cli/config_manager.py
@@ -136,6 +136,8 @@ def save(
         max_token_per_module: Optional[int] = None,
         max_token_per_leaf_module: Optional[int] = None,
         max_depth: Optional[int] = None,
+        request_limit: Optional[int] = None,
+        agent_retries: Optional[int] = None,
         provider: Optional[str] = None,
         aws_region: Optional[str] = None,
         api_version: Optional[str] = None,
@@ -157,6 +159,8 @@ def save(
             max_token_per_module: Maximum tokens per module for clustering
             max_token_per_leaf_module: Maximum tokens per leaf module
             max_depth: Maximum depth for hierarchical decomposition
+            request_limit: Maximum model requests per agent run
+            agent_retries: Retries for a failing tool call
             provider: LLM provider type (openai-compatible, anthropic, bedrock, azure-openai)
             aws_region: AWS region for Bedrock provider
             api_version: Azure OpenAI API version
@@ -205,6 +209,10 @@ def save(
             self._config.max_token_per_leaf_module = max_token_per_leaf_module
         if max_depth is not None:
             self._config.max_depth = max_depth
+        if request_limit is not None:
+            self._config.request_limit = request_limit
+        if agent_retries is not None:
+            self._config.agent_retries = agent_retries
         if provider is not None:
             self._config.provider = provider
         if aws_region is not None:
diff --git a/codewiki/cli/models/config.py b/codewiki/cli/models/config.py
index c1cabe07..f0f99cff 100644
--- a/codewiki/cli/models/config.py
+++ b/codewiki/cli/models/config.py
@@ -130,6 +130,8 @@ class Configuration:
         max_token_per_module: Maximum tokens per module for clustering (default: 36369)
         max_token_per_leaf_module: Maximum tokens per leaf module (default: 16000)
         max_depth: Maximum depth for hierarchical decomposition (default: 2)
+        request_limit: Maximum model requests per agent run (default: 100)
+        agent_retries: Retries for a failing tool call (default: 3)
         use_gitignore: Apply Git ignore rules during repository analysis
         prompt_caching: Add prompt-cache breakpoints to agentic LLM calls (default: True)
         agent_instructions: Custom agent instructions for documentation generation
@@ -148,6 +150,8 @@ class Configuration:
     max_token_per_module: int = 36369
     max_token_per_leaf_module: int = 16000
     max_depth: int = 2
+    request_limit: int = 100
+    agent_retries: int = 3
     use_gitignore: bool = True
     prompt_caching: bool = True
     agent_instructions: AgentInstructions = field(default_factory=AgentInstructions)
@@ -187,6 +191,8 @@ def to_dict(self) -> dict:
             "max_token_per_module": self.max_token_per_module,
             "max_token_per_leaf_module": self.max_token_per_leaf_module,
             "max_depth": self.max_depth,
+            "request_limit": self.request_limit,
+            "agent_retries": self.agent_retries,
             "use_gitignore": self.use_gitignore,
             "prompt_caching": self.prompt_caching,
             "fallback_model": self.fallback_model,
@@ -224,6 +230,8 @@ def from_dict(cls, data: dict) -> "Configuration":
             max_token_per_module=data.get("max_token_per_module", 36369),
             max_token_per_leaf_module=data.get("max_token_per_leaf_module", 16000),
             max_depth=data.get("max_depth", 2),
+            request_limit=data.get("request_limit", 100),
+            agent_retries=data.get("agent_retries", 3),
             use_gitignore=data.get("use_gitignore", True),
             prompt_caching=data.get("prompt_caching", True),
             agent_instructions=agent_instructions,
@@ -300,6 +308,8 @@ def to_backend_config(
             max_token_per_module=self.max_token_per_module,
             max_token_per_leaf_module=self.max_token_per_leaf_module,
             max_depth=self.max_depth,
+            request_limit=self.request_limit,
+            agent_retries=self.agent_retries,
             agent_instructions=final_instructions.to_dict() if final_instructions else None,
             use_gitignore=self.use_gitignore,
             prompt_caching=self.prompt_caching,
diff --git a/codewiki/src/be/agent_tools/generate_sub_module_documentations.py b/codewiki/src/be/agent_tools/generate_sub_module_documentations.py
index 938a34cd..9583bd61 100644
--- a/codewiki/src/be/agent_tools/generate_sub_module_documentations.py
+++ b/codewiki/src/be/agent_tools/generate_sub_module_documentations.py
@@ -1,13 +1,18 @@
 import os
 
 from pydantic_ai import RunContext, Tool, Agent
+from pydantic_ai.usage import UsageLimits
 
 from codewiki.src.be.agent_tools.deps import CodeWikiDeps
 from codewiki.src.be.module_naming import plan_sub_module_specs
 from codewiki.src.be.agent_tools.read_code_components import read_code_components_tool
 from codewiki.src.be.agent_tools.str_replace_editor import str_replace_editor_tool
 from codewiki.src.be.llm_services import create_fallback_models
-from codewiki.src.be.prompt_template import SYSTEM_PROMPT, LEAF_SYSTEM_PROMPT, format_user_prompt
+from codewiki.src.be.prompt_template import (
+    format_leaf_system_prompt,
+    format_system_prompt,
+    format_user_prompt,
+)
 from codewiki.src.be.utils import is_complex_module, count_tokens
 from codewiki.src.be.cluster_modules import format_potential_core_components
 
@@ -83,9 +88,8 @@ async def generate_sub_module_documentation(
                 model=fallback_models,
                 name=sub_module_name,
                 deps_type=CodeWikiDeps,
-                system_prompt=SYSTEM_PROMPT.format(
-                    module_name=sub_module_name, custom_instructions=ctx.deps.custom_instructions
-                ),
+                system_prompt=format_system_prompt(sub_module_name, ctx.deps.custom_instructions),
+                retries=ctx.deps.config.agent_retries,
                 tools=[
                     read_code_components_tool,
                     str_replace_editor_tool,
@@ -97,9 +101,10 @@ async def generate_sub_module_documentation(
                 model=fallback_models,
                 name=sub_module_name,
                 deps_type=CodeWikiDeps,
-                system_prompt=LEAF_SYSTEM_PROMPT.format(
-                    module_name=sub_module_name, custom_instructions=ctx.deps.custom_instructions
+                system_prompt=format_leaf_system_prompt(
+                    sub_module_name, ctx.deps.custom_instructions
                 ),
+                retries=ctx.deps.config.agent_retries,
                 tools=[read_code_components_tool, str_replace_editor_tool],
             )
 
@@ -117,6 +122,7 @@ async def generate_sub_module_documentation(
                 module_tree=ctx.deps.module_tree,
             ),
             deps=ctx.deps,
+            usage_limits=UsageLimits(request_limit=ctx.deps.config.request_limit),
         )
 
         # remove the sub-module name from the path to current module and the module tree
diff --git a/codewiki/src/be/backend.py b/codewiki/src/be/backend.py
index 294a3fc8..d31733f9 100644
--- a/codewiki/src/be/backend.py
+++ b/codewiki/src/be/backend.py
@@ -89,8 +89,9 @@ def complete(
         prompt: str,
         *,
         model: str | None = None,
+        system_prompt: str | None = None,
     ) -> str:
-        """Single-shot text completion."""
+        """Single-shot text completion, with an optional system message."""
 
     @abc.abstractmethod
     async def run_module_agent(
diff --git a/codewiki/src/be/caw_backend.py b/codewiki/src/be/caw_backend.py
index 6a2c6970..f10c46a7 100644
--- a/codewiki/src/be/caw_backend.py
+++ b/codewiki/src/be/caw_backend.py
@@ -246,6 +246,7 @@ def complete(
         prompt: str,
         *,
         model: str | None = None,
+        system_prompt: str | None = None,
     ) -> str:
         # Blocks the calling thread for the lifetime of the claude/codex
         # subprocess.  Callers running this from an async context (e.g. the
@@ -256,6 +257,7 @@ def complete(
             provider=self._caw_provider,
             model=effective_model,
             tools=ToolGroup.READER,
+            system_prompt=system_prompt,
         )
         traj = agent.completion(prompt)
         self.last_usage = usage_to_dict(getattr(traj, "total_usage", None))
diff --git a/codewiki/src/be/documentation_generator.py b/codewiki/src/be/documentation_generator.py
index 3f1a5b9d..fad342a2 100644
--- a/codewiki/src/be/documentation_generator.py
+++ b/codewiki/src/be/documentation_generator.py
@@ -24,6 +24,7 @@
     MODULE_OVERVIEW_PROMPT,
     REPO_OVERVIEW_ARTIFACT_ADDENDUM,
     REPO_OVERVIEW_PROMPT,
+    format_overview_system_prompt,
 )
 from codewiki.src.config import (
     FIRST_MODULE_TREE_FILENAME,
@@ -354,7 +355,10 @@ async def generate_parent_module_docs(
         logger.debug(f"Overview prompt for {module_name}: {len(prompt)} chars")
 
         try:
-            parent_docs = self.backend.complete(prompt)
+            parent_docs = self.backend.complete(
+                prompt,
+                system_prompt=format_overview_system_prompt(self.config.get_prompt_addition()),
+            )
             if not parent_docs:
                 raise RuntimeError(
                     f"LLM returned empty content for {module_name} overview "
diff --git a/codewiki/src/be/llm_services.py b/codewiki/src/be/llm_services.py
index 38fca92b..f3b4acb2 100644
--- a/codewiki/src/be/llm_services.py
+++ b/codewiki/src/be/llm_services.py
@@ -13,7 +13,12 @@
 
 from openai.types import chat
 
-from pydantic_ai.exceptions import ModelHTTPError
+from pydantic_ai.exceptions import ModelHTTPError, UnexpectedModelBehavior
+
+try:  # pydantic-ai >= 1.x adds ModelAPIError as the parent of ModelHTTPError
+    from pydantic_ai.exceptions import ModelAPIError
+except ImportError:  # older releases (e.g. the pinned 1.0.6) only have ModelHTTPError
+    ModelAPIError = ModelHTTPError
 from pydantic_ai.models.openai import OpenAIChatModel, OpenAIChatModelSettings
 from pydantic_ai.models.fallback import FallbackModel
 from pydantic_ai.providers.openai import OpenAIProvider
@@ -254,7 +259,11 @@ def create_fallback_models(config: Config) -> FallbackModel:
     """Create fallback models chain from configuration."""
     main = create_main_model(config)
     fallback = create_fallback_model(config)
-    return FallbackModel(main, fallback)
+    # The default fallback_on (ModelAPIError, or ModelHTTPError on older pydantic-ai)
+    # misses UnexpectedModelBehavior, which pydantic-ai raises for a 200 response whose
+    # body does not parse (e.g. choices=None from an OpenAI-compatible gateway); without
+    # it such a response skips the fallback model.
+    return FallbackModel(main, fallback, fallback_on=(ModelAPIError, UnexpectedModelBehavior))
 
 
 def create_openai_client(config: Config) -> OpenAI:
@@ -308,7 +317,15 @@ def _extract_content(response, model: str) -> Optional[str]:
     return content
 
 
-def call_llm(prompt: str, config: Config, model: str = None) -> Optional[str]:
+def _messages(prompt: str, system_prompt: str | None) -> list[dict[str, str]]:
+    messages = [{"role": "system", "content": system_prompt}] if system_prompt else []
+    messages.append({"role": "user", "content": prompt})
+    return messages
+
+
+def call_llm(
+    prompt: str, config: Config, model: str = None, system_prompt: str | None = None
+) -> Optional[str]:
     """
     Call LLM with the given prompt.
 
@@ -322,6 +339,7 @@ def call_llm(prompt: str, config: Config, model: str = None) -> Optional[str]:
         prompt: The prompt to send
         config: Configuration containing LLM settings
         model: Model name (defaults to config.main_model)
+        system_prompt: Optional system message sent before the prompt
 
     Returns:
         LLM response text, or None when the provider returned no content
@@ -333,10 +351,10 @@ def call_llm(prompt: str, config: Config, model: str = None) -> Optional[str]:
     provider = getattr(config, "provider", "openai-compatible")
 
     if provider in ("bedrock", "anthropic"):
-        return _call_llm_via_litellm(prompt, config, model)
+        return _call_llm_via_litellm(prompt, config, model, system_prompt)
 
     if provider == "azure-openai":
-        return _call_llm_via_azure(prompt, config, model)
+        return _call_llm_via_azure(prompt, config, model, system_prompt)
 
     # Default: OpenAI-compatible
     client = create_openai_client(config)
@@ -349,7 +367,7 @@ def call_llm(prompt: str, config: Config, model: str = None) -> Optional[str]:
 
     base_kwargs = {
         "model": model,
-        "messages": [{"role": "user", "content": prompt}],
+        "messages": _messages(prompt, system_prompt),
     }
 
     try:
@@ -387,7 +405,9 @@ def _is_unsupported_token_param_error(err: BadRequestError, param: str) -> bool:
     return "unsupported parameter" in msg and param in msg
 
 
-def _call_llm_via_litellm(prompt: str, config: Config, model: str) -> Optional[str]:
+def _call_llm_via_litellm(
+    prompt: str, config: Config, model: str, system_prompt: str | None = None
+) -> Optional[str]:
     """
     Call LLM via litellm for Bedrock/Anthropic providers.
 
@@ -407,14 +427,16 @@ def _call_llm_via_litellm(prompt: str, config: Config, model: str) -> Optional[s
 
     response = litellm.completion(
         model=litellm_model,
-        messages=[{"role": "user", "content": prompt}],
+        messages=_messages(prompt, system_prompt),
         max_tokens=config.max_tokens,
         api_key=config.llm_api_key if config.provider != "bedrock" else None,
     )
     return _extract_content(response, litellm_model)
 
 
-def _call_llm_via_azure(prompt: str, config: Config, model: str) -> Optional[str]:
+def _call_llm_via_azure(
+    prompt: str, config: Config, model: str, system_prompt: str | None = None
+) -> Optional[str]:
     """
     Call LLM via Azure OpenAI.
 
@@ -436,7 +458,7 @@ def _call_llm_via_azure(prompt: str, config: Config, model: str) -> Optional[str
 
     response = client.chat.completions.create(
         model=deployment,
-        messages=[{"role": "user", "content": prompt}],
+        messages=_messages(prompt, system_prompt),
         max_tokens=config.max_tokens,
     )
     return _extract_content(response, deployment)
diff --git a/codewiki/src/be/prompt_template.py b/codewiki/src/be/prompt_template.py
index ed9f736b..20429318 100644
--- a/codewiki/src/be/prompt_template.py
+++ b/codewiki/src/be/prompt_template.py
@@ -645,3 +645,20 @@ def format_leaf_system_prompt(module_name: str, custom_instructions: str | None
     return LEAF_SYSTEM_PROMPT.format(
         module_name=module_name, custom_instructions=custom_section
     ).strip()
+
+
+def format_overview_system_prompt(custom_instructions: str | None = None) -> str | None:
+    """
+    System message for the module / repository overview completions.
+
+    MODULE_OVERVIEW_PROMPT and REPO_OVERVIEW_PROMPT have no custom-instructions
+    slot, so without this the user's instructions (language, audience, style)
+    never reach overview pages. Returns None when there is nothing to add.
+    """
+    if not custom_instructions:
+        return None
+    return (
+        f"<CUSTOM_INSTRUCTIONS>\n{custom_instructions}\n</CUSTOM_INSTRUCTIONS>\n\n"
+        "Follow these instructions when writing the overview; they take precedence over "
+        "conflicting defaults in the request (language, audience, structure, style)."
+    )
diff --git a/codewiki/src/be/pydantic_ai_backend.py b/codewiki/src/be/pydantic_ai_backend.py
index 23ad08e9..3be7dada 100644
--- a/codewiki/src/be/pydantic_ai_backend.py
+++ b/codewiki/src/be/pydantic_ai_backend.py
@@ -15,6 +15,7 @@
 from typing import Any
 
 from pydantic_ai import Agent
+from pydantic_ai.usage import UsageLimits
 
 from codewiki.src.be.agent_tools.deps import CodeWikiDeps
 from codewiki.src.be.agent_tools.generate_sub_module_documentations import (
@@ -63,9 +64,10 @@ def complete(
         prompt: str,
         *,
         model: str | None = None,
+        system_prompt: str | None = None,
     ) -> str:
         pop_last_usage()
-        result = call_llm(prompt, self._config, model=model)
+        result = call_llm(prompt, self._config, model=model, system_prompt=system_prompt)
         self.last_usage = pop_last_usage()
         return result
 
@@ -81,9 +83,14 @@ async def run_update_agent(
             deps_type=CodeWikiDeps,
             tools=[read_code_components_tool, str_replace_editor_tool],
             system_prompt=system_prompt,
+            retries=self._config.agent_retries,
         )
         started = time.time()
-        result = await agent.run(user_prompt, deps=deps)
+        result = await agent.run(
+            user_prompt,
+            deps=deps,
+            usage_limits=UsageLimits(request_limit=self._config.request_limit),
+        )
         seconds = time.time() - started
         usage = _run_usage(result)
         self.last_usage = usage
@@ -127,6 +134,7 @@ async def run_module_agent(
                     generate_sub_module_documentation_tool,
                 ],
                 system_prompt=format_system_prompt(module_name, self._custom_instructions),
+                retries=config.agent_retries,
             )
         else:
             agent = Agent(
@@ -135,6 +143,7 @@ async def run_module_agent(
                 deps_type=CodeWikiDeps,
                 tools=[read_code_components_tool, str_replace_editor_tool],
                 system_prompt=format_leaf_system_prompt(module_name, self._custom_instructions),
+                retries=config.agent_retries,
             )
 
         deps = CodeWikiDeps(
@@ -160,6 +169,7 @@ async def run_module_agent(
                     module_tree=deps.module_tree,
                 ),
                 deps=deps,
+                usage_limits=UsageLimits(request_limit=config.request_limit),
             )
             self.last_usage = _run_usage(result)
             file_manager.save_json(deps.module_tree, module_tree_path)
diff --git a/codewiki/src/config.py b/codewiki/src/config.py
index b98208ed..63b98f98 100644
--- a/codewiki/src/config.py
+++ b/codewiki/src/config.py
@@ -19,6 +19,11 @@
 DEFAULT_MAX_TOKENS = 32_768
 DEFAULT_MAX_TOKEN_PER_MODULE = 36_369
 DEFAULT_MAX_TOKEN_PER_LEAF_MODULE = 4_000
+# pydantic-ai caps each agent run at 50 model requests by default; complex
+# modules that read many components and delegate sub-modules can need more.
+DEFAULT_REQUEST_LIMIT = 100
+# pydantic-ai retries a failing tool call (and output validation) once by default.
+DEFAULT_AGENT_RETRIES = 3
 # Super-group the flat top level into architectural subsystems only when it
 # has more than this many modules; 0 or negative disables the pass.
 DEFAULT_MIN_MODULES_FOR_SUPER_GROUPING = 3
@@ -91,6 +96,9 @@ class Config:
     max_token_per_leaf_module: int = DEFAULT_MAX_TOKEN_PER_LEAF_MODULE
     min_modules_for_super_grouping: int = DEFAULT_MIN_MODULES_FOR_SUPER_GROUPING
     max_leaf_nodes_per_cluster: int = DEFAULT_MAX_LEAF_NODES_PER_CLUSTER
+    # Agent run limits (model requests per run, retries per failing tool call)
+    request_limit: int = DEFAULT_REQUEST_LIMIT
+    agent_retries: int = DEFAULT_AGENT_RETRIES
     # Prompt caching for agentic/multi-turn calls (auto-disables per model if
     # the provider rejects cache_control markers)
     prompt_caching: bool = True
@@ -217,6 +225,8 @@ def from_cli(
         min_modules_for_super_grouping: int = DEFAULT_MIN_MODULES_FOR_SUPER_GROUPING,
         max_leaf_nodes_per_cluster: int = DEFAULT_MAX_LEAF_NODES_PER_CLUSTER,
         max_depth: int = MAX_DEPTH,
+        request_limit: int = DEFAULT_REQUEST_LIMIT,
+        agent_retries: int = DEFAULT_AGENT_RETRIES,
         agent_instructions: dict[str, Any] | None = None,
         use_gitignore: bool = True,
         prompt_caching: bool = True,
@@ -248,6 +258,8 @@ def from_cli(
             max_leaf_nodes_per_cluster: Partition clustering inputs into
                 structure-based batches of at most this many leaf nodes
             max_depth: Maximum depth for hierarchical decomposition
+            request_limit: Maximum model requests per agent run
+            agent_retries: Retries for a failing tool call or output validation
             agent_instructions: Custom agent instructions dict
             use_gitignore: Whether to apply Git ignore rules
             prompt_caching: Whether to add prompt-cache breakpoints to agentic calls
@@ -281,6 +293,8 @@ def from_cli(
             max_token_per_leaf_module=max_token_per_leaf_module,
             min_modules_for_super_grouping=min_modules_for_super_grouping,
             max_leaf_nodes_per_cluster=max_leaf_nodes_per_cluster,
+            request_limit=request_limit,
+            agent_retries=agent_retries,
             agent_instructions=agent_instructions,
             use_gitignore=use_gitignore,
             prompt_caching=prompt_caching,
diff --git a/pyproject.toml b/pyproject.toml
index 78d72c69..5a3c0716 100644
--- a/pyproject.toml
+++ b/pyproject.toml
@@ -108,6 +108,7 @@ packages = [
     "codewiki.src.be.dependency_analyzer.analyzers",
     "codewiki.src.be.dependency_analyzer.models",
     "codewiki.src.be.dependency_analyzer.utils",
+    "codewiki.src.be.updater",
     "codewiki.src.fe",
     "codewiki.mcp",
     "codewiki.mcp.tools"
diff --git a/tests/test_agent_limits.py b/tests/test_agent_limits.py
new file mode 100644
index 00000000..fffa1fd8
--- /dev/null
+++ b/tests/test_agent_limits.py
@@ -0,0 +1,121 @@
+"""Agent run limits, fallback triggers, overview instructions and packaging."""
+
+import asyncio
+import tomllib
+from pathlib import Path
+from types import SimpleNamespace
+
+from pydantic_ai.exceptions import UnexpectedModelBehavior
+
+import codewiki.src.be.agent_tools.generate_sub_module_documentations as sub_mod
+import codewiki.src.be.llm_services as llm_services
+from codewiki.cli.models.config import Configuration
+from codewiki.src.be.agent_tools.deps import CodeWikiDeps
+from codewiki.src.be.prompt_template import format_overview_system_prompt
+from codewiki.src.config import DEFAULT_AGENT_RETRIES, DEFAULT_REQUEST_LIMIT, Config
+
+ROOT = Path(__file__).resolve().parents[1]
+
+
+def test_configuration_round_trips_agent_limits():
+    cfg = Configuration.from_dict(
+        {
+            "base_url": "http://x",
+            "main_model": "m",
+            "cluster_model": "m",
+            "request_limit": 250,
+            "agent_retries": 5,
+        }
+    )
+    assert (cfg.request_limit, cfg.agent_retries) == (250, 5)
+    assert Configuration.from_dict(cfg.to_dict()).request_limit == 250
+    backend = cfg.to_backend_config(repo_path=".", output_dir="/tmp/out", api_key="k")
+    assert (backend.request_limit, backend.agent_retries) == (250, 5)
+
+
+def test_backend_config_defaults():
+    cfg = Config.from_cli(
+        repo_path=".",
+        output_dir="/tmp/out",
+        llm_base_url="http://x",
+        llm_api_key="k",
+        main_model="m",
+        cluster_model="m",
+    )
+    assert cfg.request_limit == DEFAULT_REQUEST_LIMIT == 100
+    assert cfg.agent_retries == DEFAULT_AGENT_RETRIES == 3
+
+
+def test_fallback_covers_malformed_responses(monkeypatch):
+    captured = {}
+    monkeypatch.setattr(llm_services, "create_main_model", lambda config: "main")
+    monkeypatch.setattr(llm_services, "create_fallback_model", lambda config: "fallback")
+    monkeypatch.setattr(
+        llm_services, "FallbackModel", lambda *models, **kwargs: captured.update(kwargs)
+    )
+    llm_services.create_fallback_models(SimpleNamespace())
+    # a 200 response with an unparseable body must also switch to the fallback model
+    assert set(captured["fallback_on"]) == {llm_services.ModelAPIError, UnexpectedModelBehavior}
+
+
+def test_overview_system_prompt():
+    assert format_overview_system_prompt("") is None
+    assert format_overview_system_prompt(None) is None
+    text = format_overview_system_prompt("Write in Portuguese.")
+    assert "<CUSTOM_INSTRUCTIONS>\nWrite in Portuguese.\n</CUSTOM_INSTRUCTIONS>" in text
+
+
+def test_messages_put_system_first():
+    assert llm_services._messages("hi", None) == [{"role": "user", "content": "hi"}]
+    assert llm_services._messages("hi", "sys") == [
+        {"role": "system", "content": "sys"},
+        {"role": "user", "content": "hi"},
+    ]
+
+
+def test_sub_module_agents_get_limits_and_no_literal_none(tmp_path, monkeypatch):
+    seen = {}
+
+    class FakeAgent:
+        def __init__(self, *args, **kwargs):
+            seen["init"] = kwargs
+
+        async def run(self, prompt, deps, **kwargs):
+            seen["run"] = kwargs
+            Path(deps.absolute_docs_path, f"{deps.current_module_name}.md").write_text("# x\n")
+            return SimpleNamespace(output="ok")
+
+    monkeypatch.setattr(sub_mod, "Agent", FakeAgent)
+    monkeypatch.setattr(sub_mod, "create_fallback_models", lambda config: None)
+    deps = CodeWikiDeps(
+        absolute_docs_path=str(tmp_path),
+        absolute_repo_path=str(tmp_path),
+        registry={},
+        components={},
+        path_to_current_module=[],
+        current_module_name="root",
+        module_tree={},
+        max_depth=2,
+        current_depth=1,
+        config=SimpleNamespace(max_token_per_leaf_module=4000, agent_retries=7, request_limit=321),
+        custom_instructions=None,
+    )
+    asyncio.run(
+        sub_mod.generate_sub_module_documentation(
+            SimpleNamespace(deps=deps), {"billing": ["b.py::fb"]}
+        )
+    )
+    assert seen["init"]["retries"] == 7
+    assert seen["run"]["usage_limits"].request_limit == 321
+    # custom_instructions=None used to be formatted into the prompt as the text "None"
+    assert not seen["init"]["system_prompt"].endswith("None")
+
+
+def test_every_package_is_listed_in_pyproject():
+    listed = set(
+        tomllib.loads((ROOT / "pyproject.toml").read_text())["tool"]["setuptools"]["packages"]
+    )
+    on_disk = {
+        ".".join(p.parent.relative_to(ROOT).parts) for p in (ROOT / "codewiki").rglob("__init__.py")
+    }
+    assert on_disk - listed == set()
diff --git a/tests/test_processing_order_update.py b/tests/test_processing_order_update.py
index 8a3a08ce..d3449bf1 100644
--- a/tests/test_processing_order_update.py
+++ b/tests/test_processing_order_update.py
@@ -78,7 +78,7 @@ async def run_module_agent(
         Path(working_dir, f"{module_name}.md").write_text(f"# {module_name}\n")
         return module_tree
 
-    def complete(self, prompt, model=None):
+    def complete(self, prompt, model=None, system_prompt=None):
         self.complete_calls += 1
         return "<OVERVIEW>overview</OVERVIEW>"
 
@@ -86,7 +86,9 @@ def complete(self, prompt, model=None):
 def _generator(docs_dir: Path) -> tuple[DocumentationGenerator, FakeBackend]:
     # Bypass __init__: it wires a real LLM backend and dependency analyzer.
     gen = object.__new__(DocumentationGenerator)
-    gen.config = SimpleNamespace(docs_dir=str(docs_dir), repo_path=str(docs_dir))
+    gen.config = SimpleNamespace(
+        docs_dir=str(docs_dir), repo_path=str(docs_dir), get_prompt_addition=lambda: ""
+    )
     gen.backend = FakeBackend(docs_dir)
     return gen, gen.backend
 
diff --git a/tests/test_sub_module_dedupe.py b/tests/test_sub_module_dedupe.py
index 83a78a3c..0c0e0728 100644
--- a/tests/test_sub_module_dedupe.py
+++ b/tests/test_sub_module_dedupe.py
@@ -21,7 +21,7 @@ class FakeAgent:
     def __init__(self, *args, **kwargs):
         self.name = kwargs.get("name")
 
-    async def run(self, prompt, deps):
+    async def run(self, prompt, deps, **kwargs):
         FakeAgent.runs.append(deps.current_module_name)
         page = f"{deps.absolute_docs_path}/{deps.current_module_name}.md"
         with open(page, "w", encoding="utf-8") as f:
@@ -53,7 +53,7 @@ def _deps(tmp_path) -> CodeWikiDeps:
         module_tree={},
         max_depth=2,
         current_depth=1,
-        config=SimpleNamespace(max_token_per_leaf_module=4000),
+        config=SimpleNamespace(max_token_per_leaf_module=4000, agent_retries=3, request_limit=100),
         custom_instructions="",
     )
 
diff --git a/tests/test_updater_orchestrator.py b/tests/test_updater_orchestrator.py
index be1892d4..292e4c05 100644
--- a/tests/test_updater_orchestrator.py
+++ b/tests/test_updater_orchestrator.py
@@ -29,7 +29,7 @@ def __init__(self, own_verdict="patch"):
         self.complete_calls = []
         self.last_usage = None
 
-    def complete(self, prompt, *, model=None):
+    def complete(self, prompt, *, model=None, system_prompt=None):
         self.complete_calls.append(prompt[:80])
         self.last_usage = {"prompt_tokens": 10, "completion_tokens": 5}
         return "<OVERVIEW>regenerated overview</OVERVIEW>"