fix: repair quality gates and land OpenWiki artifacts

.github/workflows/openwiki-update.yml

@@ -0,0 +1,62 @@
+name: OpenWiki Update
+
+on:
+  workflow_dispatch:
+  schedule:
+    - cron: "0 8 * * *"
+
+permissions:
+  contents: write
+  pull-requests: write
+
+jobs:
+  update:
+    runs-on: ubuntu-latest
+    steps:
+      - name: Check out repository
+        uses: actions/checkout@34e114876b0b11c390a56381ad16ebd13914f8d5 # v4
+        with:
+          # Full history so `openwiki code --update` can diff HEAD against the
+          # commit it last documented; a shallow clone hides that commit and the
+          # update runs against an empty change summary.
+          fetch-depth: 0
+
+      - name: Set up Node.js
+        uses: actions/setup-node@49933ea5288caeca8642d1e84afbd3f7d6820020 # v4
+        with:
+          node-version: "22"
+
+      - name: Install OpenWiki
+        # mermaid + jsdom are optional; they add high-fidelity validation of Mermaid diagrams. Remove if your wiki has none.
+        run: npm install --global openwiki@0.3.3 mermaid@11.16.0 jsdom@29.1.1
+
+      - name: Run OpenWiki
+        run: openwiki code --update --print
+        env:
+          OPENWIKI_PROVIDER: baseten
+          BASETEN_API_KEY: ${{ secrets.BASETEN_API_KEY }}
+          OPENWIKI_MODEL_ID: "moonshotai/Kimi-K3"
+          # Required for the LangSmith connector's code-mode pull to authenticate.
+          # For extra workspaces, add OPENWIKI_LANGSMITH_API_KEY_2, _3, ... as repo
+          # secrets and env entries here.
+          OPENWIKI_LANGSMITH_API_KEY: ${{ secrets.OPENWIKI_LANGSMITH_API_KEY }}
+          # Optional: also trace this workflow's own OpenWiki run to LangSmith.
+          LANGSMITH_API_KEY: ${{ secrets.LANGSMITH_API_KEY }}
+          LANGCHAIN_PROJECT: openwiki
+          LANGCHAIN_TRACING_V2: "true"
+
+      - name: Create OpenWiki update pull request
+        uses: peter-evans/create-pull-request@22a9089034f40e5a961c8808d113e2c98fb63676 # v7
+        with:
+          add-paths: |
+            openwiki
+            AGENTS.md
+            CLAUDE.md
+            .github/workflows/openwiki-update.yml
+          branch: openwiki/update
+          commit-message: "docs: update OpenWiki"
+          title: "docs: update OpenWiki"
+          body: |
+            Automated OpenWiki documentation update.
+
+            This PR was generated by the scheduled OpenWiki workflow.

AGENTS.md

@@ -103,3 +103,16 @@ Before ending a work session:
 5. Hand off remaining context clearly.
 
 Do not commit or push unrelated uncommitted user changes.
+
+<!-- OPENWIKI:START -->
+
+## OpenWiki
+
+This repository has a generated `openwiki/` evidence index. It is optional just-in-time context, not required startup reading.
+
+- Treat source code and tests as authoritative. A brief's unknowns and review items are verification gaps, not automatic requirements.
+- Prefer the narrowest quiet validation that proves the changed behavior. Preserve complete failure output.
+
+The scheduled OpenWiki GitHub Actions workflow refreshes the repository wiki. Do not hand-edit generated OpenWiki pages unless explicitly asked; prefer updating source code/docs and letting OpenWiki regenerate.
+
+<!-- OPENWIKI:END -->

CLAUDE.md

@@ -0,0 +1,7 @@
+<!-- OPENWIKI:START -->
+
+## OpenWiki
+
+See [AGENTS.md](AGENTS.md) for OpenWiki agent instructions.
+
+<!-- OPENWIKI:END -->

agentic_internet/tools/code_execution.py

@@ -1,6 +1,7 @@
 """Code execution tool for agents."""
 
 import ast
+import builtins
 import io
 import json
 import logging
@@ -20,7 +21,30 @@
 
 logger = logging.getLogger(__name__)
 
+_BUILTIN_IMPORT = builtins.__import__
+
+
+def _guarded_import(
+    name: str,
+    globals_: dict[str, Any] | None = None,
+    locals_: dict[str, Any] | None = None,
+    fromlist: tuple[str, ...] = (),
+    level: int = 0,
+) -> Any:
+    """Restricted ``__import__`` for executed code.
+
+    C extensions (e.g. numpy reductions) import submodules at runtime through
+    the frame's ``__builtins__``; without this entry they crash with
+    ``KeyError: '__import__'``. The guard enforces the same module blocklist
+    as the AST validator, so aliased indirect imports stay blocked.
+    """
+    if name.split(".")[0] in _ASTSafetyValidator.BLOCKED_MODULES:
+        raise ImportError(f"import of blocked module '{name}' is not allowed")
+    return _BUILTIN_IMPORT(name, globals_, locals_, fromlist, level)
+
+
 ALLOWED_BUILTINS = {
+    "__import__": _guarded_import,
     "print": print,
     "len": len,
     "range": range,

docs/exec-plans/active/.gitkeep

@@ -0,0 +1 @@
+In-flight exec-plans live here (see docs/PLANS.md). Completed plans move to docs/exec-plans/completed/.

openwiki/.last-update.json

@@ -0,0 +1,8 @@
+{
+  "updatedAt": "2026-08-15T16:23:09.917Z",
+  "command": "init",
+  "gitHead": "bea3d8316ab249a8ae63910e4a1186ec147b51f4",
+  "model": "gpt-5.6-sol",
+  "status": "complete",
+  "language": "en"
+}

openwiki/INSTRUCTIONS.md

@@ -0,0 +1 @@
+A code wiki for this repository.

openwiki/agents/index.md

@@ -0,0 +1,4 @@
+# Files
+
+- [Internet and Research Agents](internet-and-research.md) - Construction, tool registration, execution, chat, and research-history behavior for the core single-agent runtime.
+- [Specialized Agents](specialized-agents.md) - Public domain-specific facades for browser automation, data analysis, content creation, market research, and technical support.

openwiki/agents/internet-and-research.md

@@ -0,0 +1,76 @@
+---
+type: component guide
+title: Internet and Research Agents
+description: Construction, tool registration, execution, chat, and research-history behavior for the core single-agent runtime.
+tags: [agents, runtime, research]
+---
+
+# Internet and Research Agents
+
+`agentic_internet/agents/internet_agent.py` owns the default single-agent path. `InternetAgent` is the base public facade; `ResearchAgent` adds depth-specific prompts and in-memory history. CLI `chat`, `run`, `research`, and default `tools` all enter here.
+
+## Construction
+
+```mermaid
+sequenceDiagram
+    participant Caller
+    participant IA as InternetAgent
+    participant MU as model_utils
+    participant CFG as settings
+    participant SA as smolagents
+    Caller->>IA: construct model tools and agent_type
+    IA->>MU: initialize_model
+    MU->>CFG: resolve alias provider and key
+    MU-->>IA: model
+    IA->>IA: assemble default tools when tools is falsey
+    IA->>SA: create ToolCallingAgent or CodeAgent
+    SA-->>IA: executable agent
+```
+
+*Construction resolves the model before default tools and the underlying smolagents agent.*
+
+`__init__` accepts `model_id`, `tools`, `verbose`, `max_iterations`, `planning_enabled`, `agent_type`, and extra authorized imports. Important invariants:
+
+- `tools or self._get_default_tools()` means `tools=[]` cannot express a tool-free agent; supply a nonempty custom list or change this behavior deliberately.
+- `_initialize_model` delegates to `initialize_model`; a falsey result raises `ModelInitializationError` during construction.
+- `agent_type == "code"` creates `CodeAgent(max_steps=max_iterations)` with fixed authorized imports (`pandas`, `numpy`, JSON/CSV/regex/date/time, `requests`, `urllib`, math/statistics/collections) plus caller additions.
+- Every other value creates `ToolCallingAgent`; unknown values are not rejected. The tool-calling branch does not pass `max_iterations`.
+- `planning_enabled` is stored but does not alter agent creation or execution.
+
+The broader isolation implications are compared in [System Architecture](../architecture/overview.md).
+
+## Default tools
+
+`_get_default_tools` uses the import-time `settings` singleton:
+
+| Gate | Registered tools |
+|---|---|
+| `settings.tools.web_search_enabled` | `WebSearchTool`, `WebScraperTool`, `NewsSearchTool` |
+| Above plus `settings.exa_api_key` | `ExaSearchTool`, `ExaFindSimilarTool` |
+| `settings.tools.browser_enabled` plus `settings.browser_use_api_key` | sync, async, and structured Browser Use tools |
+| `settings.tools.code_execution_enabled` | `PythonExecutorTool`, `DataAnalysisTool` |
+| Best-effort always | `smolagents.load_tool("calculator")`; failure is debug-logged and ignored |
+
+MCP tools are never defaults; the [MCP CLI path](../tools/mcp-integration.md) injects them explicitly. Tool contracts live under [Web Search](../tools/web-search-and-scraping.md), [Exa](../tools/exa-search.md), [Browser Automation](../tools/browser-automation.md), and [Code Execution](../tools/code-execution-and-data.md).
+
+## Run and chat lifecycle
+
+`run(task, show_result=True, **kwargs)` optionally renders the task, calls `self.agent.run(task, **kwargs)`, optionally renders returned Markdown, and returns `str(result)`. It catches every exception, logs the traceback, and returns `"Error executing task: ..."`. Callers must inspect outcomes; exception-to-string conversion often leaves CLI exit status zero.
+
+`chat()` is a terminal loop. `exit`, `quit`, and `bye` stop; `help` and `tools` display local help; all other input calls `run`. A local `{user, agent}` history is appended but never returned or persisted. `KeyboardInterrupt` exits, while other loop errors are printed and the loop continues. `_show_tools` expects each tool to expose a string `description`.
+
+## ResearchAgent
+
+`ResearchAgent` delegates construction and initializes `research_history`. `research(topic, depth)` selects fixed `quick`, `moderate`, or `deep` prompts, calls `run(show_result=False)`, then appends and returns:
+
+```text
+{topic, depth, findings, timestamp}
+```
+
+The timestamp is `pandas.Timestamp.now().isoformat()`. An unknown depth uses the moderate prompt but preserves the unknown value in returned metadata; the CLI prevents this through its own validation. `get_research_history()` returns the mutable internal list, not a defensive copy.
+
+## Extension and validation
+
+Override `_initialize_model`, `_get_default_tools`, or `_create_agent` for controlled variants; prefer injecting a nonempty tool list for consumers. When adding a default tool, implement and export it under `agentic_internet.tools`, add the correct setting/key gate here, update any K-LLM inventory separately, and test both enabled and disabled registration.
+
+Focused coverage is currently narrow: `tests/test_exa_search.py` proves conditional Exa registration. There is no isolated test for model failure, agent construction, `run`, chat, `ResearchAgent`, planning, or tool-calling iteration limits. Use `uv run pytest tests/test_exa_search.py` plus the focused suite for the tool/model code changed; run `uv run pytest` for constructor-wide changes.

openwiki/agents/specialized-agents.md

@@ -0,0 +1,38 @@
+---
+type: component guide
+title: Specialized Agents
+description: Public domain-specific facades for browser automation, data analysis, content creation, market research, and technical support.
+tags: [agents, public-api]
+---
+
+# Specialized Agents
+
+`agentic_internet/agents/specialized_agents.py` provides five root-exported subclasses of [InternetAgent](internet-and-research.md). They do not introduce new execution engines: each chooses constructor defaults, creates a detailed prompt, calls inherited `run`, and sometimes wraps the string in metadata. Prompt requests are intentions, not postcondition enforcement.
+
+## Agent contracts
+
+| Class and defaults | Methods and returned shape |
+|---|---|
+| `BrowserAutomationAgent`, `agent_type="code"`, `max_iterations=20` | `scrape_structured_data(url, data_schema)` requests JSON and returns a decoded dict only when the final text is a JSON object; arrays/scalars/invalid JSON become `{"raw_result": result}`. `fill_form` returns text. `monitor_website` performs one check and returns URL, target, findings, and a suggested interval—it does not schedule monitoring. |
+| `DataAnalysisAgent`, code mode | `analyze_dataset(data, analysis_type)` returns `{analysis_type, insights}`. `compare_datasets` turns optional criteria into prompt bullets and returns text. Inputs are string-interpolated without size or schema controls. |
+| `ContentCreationAgent`, tool-calling mode | `write_article(topic, style, word_count, sources_required)` returns topic/style/content/target count; it does not verify count or citations. `summarize_content(content, summary_type)` returns text. |
+| `MarketResearchAgent`, tool-calling mode, `max_iterations=15` | `analyze_competitor` defaults to products, pricing, position, strengths, and weaknesses and returns `{company, aspects, analysis}`. `market_trends` returns `{industry, timeframe, trends}`. |
+| `TechnicalSupportAgent`, code mode | `troubleshoot(problem_description, system_info)` JSON-renders optional system information and returns `{problem, system_info, solution}`. `code_review(code, language, focus_areas)` embeds a language fence and returns language, normalized focus list, and review. |
+
+All inherited failures usually appear inside a result string because `InternetAgent.run` catches exceptions. A metadata wrapper can therefore look structurally successful while its content starts with `Error executing task:`. Code-mode classes inherit the local execution/network implications documented in [System Architecture](../architecture/overview.md).
+
+## Ownership and dependencies
+
+The module depends only on JSON handling and `InternetAgent`; actual models and tools come from the base class. `agentic_internet/__init__.py` and `agents/__init__.py` expose all five classes. `examples/advanced_usage.py` demonstrates selected methods but requires real model/provider configuration and is not a hermetic test.
+
+## Extension recipe
+
+For another domain facade:
+
+1. Subclass `InternetAgent` and set defaults with `kwargs.setdefault` so callers can override them.
+2. Keep side effects in registered tools rather than the prompt wrapper.
+3. Define whether the method returns raw text or a stable dict. If parsing model output, preserve malformed output explicitly rather than silently discarding it.
+4. Export from `agents/__init__.py` and root `agentic_internet/__init__.py` if it is public.
+5. Add a focused test with construction bypassed or model/tool mocks, checking prompt-critical behavior and malformed/failure output—not only metadata.
+
+`tests/test_specialized_agents.py` verifies object-versus-nonobject JSON handling for `scrape_structured_data` and article metadata. Constructor defaults and the other helper methods have no focused coverage. Run that file for facade changes and the base/tool suites when changing inherited behavior.

openwiki/architecture/index.md

@@ -0,0 +1,3 @@
+# Files
+
+- [System Architecture](overview.md) - Runtime layers, dependency flow, orchestration paths, state ownership, and cross-cutting trust boundaries for Agentic Internet.