clearfront 2.5.0__tar.gz → 2.7.0__tar.gz

{clearfront-2.5.0 → clearfront-2.7.0}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: clearfront
-Version: 2.5.0
+Version: 2.7.0
 Summary: AI-powered OSINT agent, MCP server, and CLI. Interactive REPL + 30 tools. Anthropic Claude or local Ollama. For authorized security research use only.
 Author: Scott Martin Anderson
 License-Expression: MIT
@@ -23,50 +23,50 @@ Requires-Python: >=3.10
 Description-Content-Type: text/markdown
 License-File: LICENSE
 Requires-Dist: mcp>=1.28.1
-Requires-Dist: anthropic>=0.40.0
-Requires-Dist: requests>=2.31.0
+Requires-Dist: anthropic>=0.116.0
+Requires-Dist: requests>=2.34.2
 Requires-Dist: python-whois>=0.9.6
-Requires-Dist: rich>=13.0.0
-Requires-Dist: prompt_toolkit>=3.0.0
-Requires-Dist: python-dotenv>=1.0.0
+Requires-Dist: rich>=15.0.0
+Requires-Dist: prompt_toolkit>=3.0.52
+Requires-Dist: python-dotenv>=1.2.2
 Requires-Dist: aiohttp>=3.14.1
 Requires-Dist: dnspython>=2.8.0
-Requires-Dist: ddgs>=9.0.0
+Requires-Dist: ddgs>=9.14.4
 Requires-Dist: fastapi>=0.139.0
-Requires-Dist: uvicorn[standard]>=0.29.0
-Requires-Dist: httpx>=0.27.0
+Requires-Dist: uvicorn[standard]>=0.51.0
+Requires-Dist: httpx>=0.28.1
 Provides-Extra: dev
-Requires-Dist: pytest>=7.0; extra == "dev"
-Requires-Dist: pytest-asyncio>=0.21; extra == "dev"
+Requires-Dist: pytest>=9.1.1; extra == "dev"
+Requires-Dist: pytest-asyncio>=1.4.0; extra == "dev"
 Requires-Dist: pytest-cov>=7.1.0; extra == "dev"
 Requires-Dist: pytest-freezegun>=0.4.2; extra == "dev"
-Requires-Dist: ruff>=0.1.0; extra == "dev"
+Requires-Dist: ruff>=0.15.20; extra == "dev"
 Provides-Extra: shodan
 Requires-Dist: shodan>=1.31.0; extra == "shodan"
 Provides-Extra: ollama
-Requires-Dist: ollama>=0.3.0; extra == "ollama"
+Requires-Dist: ollama>=0.6.2; extra == "ollama"
 Provides-Extra: openai
-Requires-Dist: openai>=1.40.0; extra == "openai"
+Requires-Dist: openai>=2.44.0; extra == "openai"
 Provides-Extra: pdf
-Requires-Dist: reportlab>=4.0.0; extra == "pdf"
+Requires-Dist: reportlab>=5.0.0; extra == "pdf"
 Provides-Extra: censys
-Requires-Dist: censys-platform>=0.14.0; extra == "censys"
+Requires-Dist: censys-platform>=0.15.1; extra == "censys"
 Provides-Extra: web
 Requires-Dist: fastapi>=0.139.0; extra == "web"
-Requires-Dist: uvicorn[standard]>=0.29.0; extra == "web"
-Requires-Dist: sse-starlette>=1.8.0; extra == "web"
-Requires-Dist: python-dotenv>=1.0.0; extra == "web"
-Requires-Dist: anthropic>=0.25.0; extra == "web"
-Requires-Dist: playwright>=1.40.0; extra == "web"
+Requires-Dist: uvicorn[standard]>=0.51.0; extra == "web"
+Requires-Dist: sse-starlette>=3.4.5; extra == "web"
+Requires-Dist: python-dotenv>=1.2.2; extra == "web"
+Requires-Dist: anthropic>=0.116.0; extra == "web"
+Requires-Dist: playwright>=1.61.0; extra == "web"
 Provides-Extra: all
 Requires-Dist: shodan>=1.31.0; extra == "all"
-Requires-Dist: ollama>=0.3.0; extra == "all"
-Requires-Dist: openai>=1.40.0; extra == "all"
-Requires-Dist: reportlab>=4.0.0; extra == "all"
-Requires-Dist: censys-platform>=0.14.0; extra == "all"
+Requires-Dist: ollama>=0.6.2; extra == "all"
+Requires-Dist: openai>=2.44.0; extra == "all"
+Requires-Dist: reportlab>=5.0.0; extra == "all"
+Requires-Dist: censys-platform>=0.15.1; extra == "all"
 Requires-Dist: fastapi>=0.139.0; extra == "all"
-Requires-Dist: uvicorn[standard]>=0.29.0; extra == "all"
-Requires-Dist: sse-starlette>=1.8.0; extra == "all"
+Requires-Dist: uvicorn[standard]>=0.51.0; extra == "all"
+Requires-Dist: sse-starlette>=3.4.5; extra == "all"
 Dynamic: license-file
 
 <div align="center">
@@ -81,7 +81,7 @@ Dynamic: license-file
 [![Python](https://img.shields.io/badge/python-3.10%2B-blue?style=flat-square)](https://www.python.org/)
 [![License MIT](https://img.shields.io/badge/license-MIT-green?style=flat-square)](LICENSE)
 [![MCP](https://img.shields.io/badge/protocol-MCP-blueviolet?style=flat-square)](https://modelcontextprotocol.io/)
-[![Version](https://img.shields.io/badge/version-2.5.0-informational?style=flat-square)](CHANGELOG.md)
+[![Version](https://img.shields.io/badge/version-2.7.0-informational?style=flat-square)](CHANGELOG.md)
 
 **[clearfront.sh](https://clearfront.sh)** · [Disclaimer](./DISCLAIMER.md)
 

{clearfront-2.5.0 → clearfront-2.7.0}/README.md

@@ -10,7 +10,7 @@
 [![Python](https://img.shields.io/badge/python-3.10%2B-blue?style=flat-square)](https://www.python.org/)
 [![License MIT](https://img.shields.io/badge/license-MIT-green?style=flat-square)](LICENSE)
 [![MCP](https://img.shields.io/badge/protocol-MCP-blueviolet?style=flat-square)](https://modelcontextprotocol.io/)
-[![Version](https://img.shields.io/badge/version-2.5.0-informational?style=flat-square)](CHANGELOG.md)
+[![Version](https://img.shields.io/badge/version-2.7.0-informational?style=flat-square)](CHANGELOG.md)
 
 **[clearfront.sh](https://clearfront.sh)** · [Disclaimer](./DISCLAIMER.md)
 

{clearfront-2.5.0 → clearfront-2.7.0}/clearfront/agent.py

@@ -57,7 +57,8 @@ from clearfront.tools.search_wayback import run_wayback_osint
 from clearfront.tools.search_greynoise import run_greynoise_osint
 from clearfront.tools.search_hudsonrock import run_hudsonrock_osint
 from clearfront.pivot import investigate_graph_for_agent
-from clearfront.prompts import SYSTEM_PROMPT
+from clearfront.prompts import system_prompt_for_effort
+from clearfront import effort as _effort
 
 logger = logging.getLogger(__name__)
 
@@ -787,17 +788,23 @@ class OISAgent:
         self,
         api_key: str | None = None,
         model: str = "claude-sonnet-4-20250514",
+        effort: str = _effort.DEFAULT,
     ) -> None:
         self.client = anthropic.AsyncAnthropic(
             api_key=api_key or os.environ.get("ANTHROPIC_API_KEY", "")
         )
         self.model = model
+        self.effort = _effort.normalize(effort)
         self.history: list[dict[str, Any]] = []
 
     def clear_history(self) -> None:
         """Reset conversation memory."""
         self.history = []
 
+    def set_effort(self, effort: str) -> None:
+        """Set the sweep effort for subsequent investigations."""
+        self.effort = _effort.normalize(effort)
+
     async def run(
         self,
         prompt: str,
@@ -825,14 +832,38 @@ class OISAgent:
             tool_calls=[],
             on_tool_call=on_tool_call,
         )
+        # Sweep effort shapes the system prompt (enrichment instruction) and caps
+        # the number of tool rounds. When the cap is reached we do not abort; we
+        # take one more turn with tools disabled so the model writes its report
+        # from the evidence gathered so far.
+        system_text = system_prompt_for_effort(self.effort)
+        max_rounds = _effort.rounds(self.effort)
+        tool_rounds = 0
         try:
             while True:
+                tool_rounds += 1
+                force_final = tool_rounds > max_rounds
+                # Prompt caching: a breakpoint on the (single) system block caches the
+                # system prompt + all tool definitions together (tools render before
+                # system), so that large stable prefix is re-billed at ~0.1x on every
+                # round instead of full price. Top-level cache_control auto-places a
+                # second breakpoint on the last message block, so the growing tool-output
+                # history is cached incrementally across rounds. Caching changes billing
+                # only; the model sees identical content and produces identical output.
                 response = await self.client.messages.create(
                     model=self.model,
                     max_tokens=_MAX_TOKENS,
-                    system=SYSTEM_PROMPT,
+                    system=[
+                        {
+                            "type": "text",
+                            "text": system_text,
+                            "cache_control": {"type": "ephemeral"},
+                        }
+                    ],
                     tools=TOOL_DEFINITIONS,  # type: ignore[arg-type]
+                    tool_choice={"type": "none"} if force_final else {"type": "auto"},
                     messages=ctx.messages,  # type: ignore[arg-type]
+                    cache_control={"type": "ephemeral"},
                 )
                 if response.stop_reason == "end_turn":
                     text = _extract_first_text(response.content)
@@ -897,15 +928,21 @@ class OllamaAgent:
         self,
         model: str = "llama3.2",
         host: str = "http://localhost:11434",
+        effort: str = _effort.DEFAULT,
     ) -> None:
         self.model = model
         self.host = host
+        self.effort = _effort.normalize(effort)
         self.history: list[dict[str, Any]] = []
 
     def clear_history(self) -> None:
         """Reset conversation memory."""
         self.history = []
 
+    def set_effort(self, effort: str) -> None:
+        """Set the sweep effort for subsequent investigations."""
+        self.effort = _effort.normalize(effort)
+
     async def run(
         self,
         prompt: str,
@@ -948,21 +985,27 @@ class OllamaAgent:
 
         self.history.append({"role": "user", "content": prompt})
         messages: list[Any] = [
-            {"role": "system", "content": SYSTEM_PROMPT},
+            {"role": "system", "content": system_prompt_for_effort(self.effort)},
             *self.history,
         ]
         ctx = _AgentRunContext(messages=messages, tool_calls=[], on_tool_call=on_tool_call)
 
+        # Sweep effort caps the tool rounds. Ollama has no reliable tool_choice
+        # switch, so on the final turn we call without tools, which forces a
+        # text answer from the evidence gathered so far.
+        max_rounds = _effort.rounds(self.effort)
+        tool_rounds = 0
         try:
             client = ollama.AsyncClient(host=self.host)
             while True:
-                response = await client.chat(
-                    model=self.model,
-                    messages=ctx.messages,
-                    tools=_OLLAMA_TOOLS,
-                )
+                tool_rounds += 1
+                force_final = tool_rounds > max_rounds
+                chat_kwargs: dict[str, Any] = {"model": self.model, "messages": ctx.messages}
+                if not force_final:
+                    chat_kwargs["tools"] = _OLLAMA_TOOLS
+                response = await client.chat(**chat_kwargs)
                 msg = response.message
-                if not msg.tool_calls:
+                if force_final or not msg.tool_calls:
                     text = msg.content or ""
                     self.history.append({"role": "assistant", "content": text})
                     return AgentResponse(content=text, tool_calls=ctx.tool_calls)
@@ -1065,17 +1108,23 @@ class OpenAICompatibleAgent:
         model: str = "gpt-4o-mini",
         base_url: str = "http://localhost:8080/v1",
         api_key: str | None = None,
+        effort: str = _effort.DEFAULT,
     ) -> None:
         self.model = model
         self.base_url = base_url
         # Many local servers ignore the key, but the SDK requires a non-empty string.
         self.api_key = api_key or os.environ.get("OPENAI_API_KEY", "") or "sk-no-key-required"
+        self.effort = _effort.normalize(effort)
         self.history: list[dict[str, Any]] = []
 
     def clear_history(self) -> None:
         """Reset conversation memory."""
         self.history = []
 
+    def set_effort(self, effort: str) -> None:
+        """Set the sweep effort for subsequent investigations."""
+        self.effort = _effort.normalize(effort)
+
     async def run(
         self,
         prompt: str,
@@ -1110,19 +1159,26 @@ class OpenAICompatibleAgent:
 
         self.history.append({"role": "user", "content": prompt})
         messages: list[Any] = [
-            {"role": "system", "content": SYSTEM_PROMPT},
+            {"role": "system", "content": system_prompt_for_effort(self.effort)},
             *self.history,
         ]
         ctx = _AgentRunContext(messages=messages, tool_calls=[], on_tool_call=on_tool_call)
 
+        # Sweep effort caps the tool rounds. On the final turn we set
+        # tool_choice="none" so the model writes its report instead of calling
+        # another tool.
+        max_rounds = _effort.rounds(self.effort)
+        tool_rounds = 0
         try:
             client = openai.AsyncOpenAI(base_url=self.base_url, api_key=self.api_key)
             while True:
+                tool_rounds += 1
+                force_final = tool_rounds > max_rounds
                 response = await client.chat.completions.create(
                     model=self.model,
                     messages=ctx.messages,
                     tools=_OPENAI_TOOLS,
-                    tool_choice="auto",
+                    tool_choice="none" if force_final else "auto",
                     max_tokens=_MAX_TOKENS,
                 )
                 if not response.choices:
@@ -1134,7 +1190,7 @@ class OpenAICompatibleAgent:
                         ),
                     )
                 msg = response.choices[0].message
-                if not msg.tool_calls:
+                if force_final or not msg.tool_calls:
                     text = msg.content or ""
                     self.history.append({"role": "assistant", "content": text})
                     return AgentResponse(content=text, tool_calls=ctx.tool_calls)

{clearfront-2.5.0 → clearfront-2.7.0}/clearfront/cli.py

@@ -213,6 +213,17 @@ def _build_parser() -> argparse.ArgumentParser:
         dest="is_pdf_disabled",
         help="Disable automatic PDF generation alongside Markdown reports.",
     )
+    parser.add_argument(
+        "--effort",
+        type=str,
+        default="deeper",
+        choices=["faster", "balanced", "deeper"],
+        help=(
+            "Sweep effort for the interactive analyst (default: deeper).  "
+            "faster: fewer sources, quick pass.  balanced: main sources.  "
+            "deeper: follows every lead.  Change it live with the 'effort' command."
+        ),
+    )
 
     subparsers = parser.add_subparsers(dest="command", metavar="command")
 
@@ -1389,6 +1400,7 @@ async def _async_main() -> None:
             openai_model=getattr(args, "openai_model", "gpt-4o-mini"),
             openai_api_key=getattr(args, "openai_api_key", None),
             is_pdf_disabled=getattr(args, "is_pdf_disabled", False),
+            effort=getattr(args, "effort", "deeper"),
         )
         await repl.run()
         return

clearfront-2.7.0/clearfront/effort.py

@@ -0,0 +1,65 @@
+# clearfront/effort.py
+"""Sweep effort: shared level metadata and the tool-round ceiling per level.
+
+Faster, Balanced, and Deeper trade collection breadth (the number of tool
+rounds the analyst runs) and the matching enrichment instruction, never the
+analyst's reasoning or output quality. Deeper is the full default fan-out; its
+ceiling stays env-tunable via ``OIS_MAX_TOOL_ROUNDS``. The lighter levels are
+capped below it and never exceed it.
+
+This module holds the pieces that are identical across surfaces: the level list
+(used by the web console chip and the REPL ``effort`` command) and the per-level
+round ceiling (used by both agent loops). The prompt wording itself is
+surface-specific and lives with each surface: the web console keeps its own
+graph-flavoured enrichment text in ``web_server``; the terminal agent's version
+lives in ``prompts`` (see ``system_prompt_for_effort``).
+"""
+
+from __future__ import annotations
+
+import os
+
+DEFAULT = "deeper"
+
+# User-facing labels and descriptions. The descriptions are kept identical to
+# the web console's sweep-effort menu (clearfront/web/index.html effortLevels) so
+# every surface describes the feature the same way.
+LEVELS: list[dict[str, str]] = [
+    {
+        "v": "faster",
+        "name": "Faster",
+        "desc": "Checks fewer sources for a quick sweep. Finishes fast and uses less of your API budget.",
+    },
+    {
+        "v": "balanced",
+        "name": "Balanced",
+        "desc": "Covers the main sources. A middle ground on time and cost.",
+    },
+    {
+        "v": "deeper",
+        "name": "Deeper",
+        "desc": "Follows every lead for the most complete map, but takes longer and uses more of your API budget.",
+    },
+]
+
+_VALUES = {level["v"] for level in LEVELS}
+
+
+def normalize(effort: str | None) -> str:
+    """Return a valid effort value, falling back to the default for anything unknown."""
+    return effort if effort in _VALUES else DEFAULT
+
+
+def rounds(effort: str) -> int:
+    """Tool-round ceiling for a effort level. Deeper stays env-tunable (default 12);
+    the lighter levels are capped below it and never exceed it."""
+    ceiling = int(os.environ.get("OIS_MAX_TOOL_ROUNDS", "12"))
+    return {"faster": min(4, ceiling), "balanced": min(8, ceiling)}.get(effort, ceiling)
+
+
+def describe(effort: str) -> str:
+    """One-line description for a effort value (for the REPL 'effort' command)."""
+    for level in LEVELS:
+        if level["v"] == effort:
+            return level["desc"]
+    return ""

{clearfront-2.5.0 → clearfront-2.7.0}/clearfront/mcp_server.py

@@ -1,6 +1,6 @@
 # clearfront/mcp_server.py
 """
-Clearfront MCP Server, v2.5.0
+Clearfront MCP Server, v2.7.0
 
 Exposes all 30 OSINT tool capabilities plus multi-target investigation
 to MCP-compliant AI clients over standard I/O. Tools include:

{clearfront-2.5.0 → clearfront-2.7.0}/clearfront/prompts.py

@@ -83,6 +83,76 @@ CRITICAL RULES:
 SYSTEM_PROMPT = INVESTIGATION_STRATEGY + "\n\n" + ANALYST_CORE
 
 
+# ---------------------------------------------------------------------------
+# Sweep effort (terminal agent): swap the enrichment instruction and, for the
+# lighter levels, prepend a forceful collection-mode line. Deeper is the full
+# default fan-out and returns SYSTEM_PROMPT byte-identical. The round ceiling
+# that pairs with each level lives in clearfront/effort.py.
+# ---------------------------------------------------------------------------
+
+# The exact aggressive-enrichment sentence embedded in ANALYST_CORE above. Kept
+# as a constant so the lighter levels can swap it out reliably; a test asserts
+# it still occurs verbatim in SYSTEM_PROMPT so an edit to ANALYST_CORE cannot
+# silently break the swap.
+_ENRICH_DEEPER = (
+    "- Enrich aggressively before finalising. Every time a pivot surfaces (an email, domain, "
+    "company, real name, phone, or an additional handle), expand it with the applicable tools, "
+    "then expand the new entities those reveal, chaining outward until the tool budget is reached. "
+    "The more real, connected entities you surface, the stronger the report."
+)
+
+AGENT_ENRICH = {
+    "deeper": _ENRICH_DEEPER,
+    "balanced": (
+        "- Enrich only the strongest pivots before finalising. When a high-value pivot surfaces "
+        "(a confirmed email, domain, real name, or primary handle), expand it once with the "
+        "applicable tools. Do not chain outward exhaustively across every lead, even if the user's "
+        "own message asks you to map everything or run until the tool budget is reached. Deliver a "
+        "solid report on the main footprint."
+    ),
+    "faster": (
+        "- Do not enrich outward. This is a single focused pass: run only the highest-signal tools "
+        "for the target, and do not scrape URLs or chase company, domain, breach, or secondary "
+        "pivots, even if the user's own message asks you to map everything or run until the tool "
+        "budget is reached. Deliver a tight, accurate report from what the priority tools return, "
+        "and note that a deeper sweep is available for the full map."
+    ),
+}
+
+# Prepended only for the lighter levels, so the instruction to run light wins
+# over any "map everything" language in the base prompt or the user's message.
+AGENT_MODE_LINE = {
+    "balanced": (
+        "COLLECTION MODE: BALANCED. This overrides any other instruction, including in the user's "
+        "own message, to map every reachable data point or run until the tool budget is reached. "
+        "Sweep the main sources and expand only the strongest one or two pivots. Do not chain "
+        "outward exhaustively."
+    ),
+    "faster": (
+        "COLLECTION MODE: FASTER. This overrides any other instruction, including in the user's own "
+        "message, to map every reachable data point, enrich every pivot, or run until the tool "
+        "budget is reached. Run a single quick pass of only the highest-signal tools for the target "
+        "type. Do not scrape URLs and do not chase company, domain, or secondary pivots."
+    ),
+}
+
+
+def system_prompt_for_effort(effort: str) -> str:
+    """Return the terminal agent's system prompt shaped for a sweep effort.
+
+    Deeper returns SYSTEM_PROMPT unchanged. Balanced and Faster swap the
+    enrichment sentence for the lighter instruction and prepend the matching
+    collection-mode line.
+    """
+    if effort not in AGENT_ENRICH:
+        effort = "deeper"
+    core = SYSTEM_PROMPT
+    if effort != "deeper":
+        core = core.replace(_ENRICH_DEEPER, AGENT_ENRICH[effort])
+    preamble = AGENT_MODE_LINE.get(effort)
+    return f"{preamble}\n\n{core}" if preamble else core
+
+
 # ---------------------------------------------------------------------------
 # Compact variant for small local models (Ollama)
 # ---------------------------------------------------------------------------

{clearfront-2.5.0 → clearfront-2.7.0}/clearfront/repl.py

@@ -31,6 +31,7 @@ from rich.panel import Panel
 
 from clearfront import __version__
 from clearfront.agent import OllamaAgent, OpenAICompatibleAgent, OISAgent
+from clearfront import effort as _effort
 
 logger = logging.getLogger(__name__)
 
@@ -134,6 +135,7 @@ def _print_help() -> None:
                     "[bold]Commands:[/]",
                     "",
                     "  [#00ff88]<target>[/]          Investigate any target (email, username, domain, IP, name)",
+                    "  [#00ff88]effort[/]             Set sweep effort (faster, balanced, deeper)",
                     "  [#00ff88]clear[/]             Clear conversation memory",
                     "  [#00ff88]save[/]              Save last report to reports/",
                     "  [#00ff88]tools[/]             List available OSINT tools",
@@ -216,6 +218,7 @@ def _print_config(
     ollama_host: str,
     is_pdf_disabled: bool,
     openai_base_url: str = "",
+    effort: str = _effort.DEFAULT,
 ) -> None:
     masked = ("*" * 20 + api_key[-6:]) if api_key and len(api_key) > 6 else "not set"
     rows = [
@@ -229,6 +232,7 @@ def _print_config(
     else:
         rows.append(f"[bold]Ollama:[/]   {ollama_host}")
     rows += [
+        f"[bold]Effort:[/]    {effort}",
         "[bold]Reports:[/]  ./reports/",
         f"[bold]PDF:[/]      {'disabled' if is_pdf_disabled else 'enabled'}",
     ]
@@ -276,6 +280,7 @@ class OISRepl:
         openai_model: str = "gpt-4o-mini",
         openai_api_key: str | None = None,
         is_pdf_disabled: bool = False,
+        effort: str = _effort.DEFAULT,
     ) -> None:
         self._api_key = api_key or os.environ.get("ANTHROPIC_API_KEY", "")
         self._provider = provider
@@ -285,12 +290,14 @@ class OISRepl:
         self._openai_model = openai_model
         self._openai_api_key = openai_api_key
         self._is_pdf_disabled = is_pdf_disabled
+        self._effort = _effort.normalize(effort)
 
         self._agent: OISAgent | OllamaAgent | OpenAICompatibleAgent
         if provider == "ollama":
             self._agent = OllamaAgent(
                 model=ollama_model,
                 host=ollama_host,
+                effort=self._effort,
             )
             self._display_model = ollama_model
         elif provider == "openai":
@@ -298,10 +305,11 @@ class OISRepl:
                 model=openai_model,
                 base_url=openai_base_url,
                 api_key=openai_api_key,
+                effort=self._effort,
             )
             self._display_model = openai_model
         else:
-            self._agent = OISAgent(api_key=self._api_key)
+            self._agent = OISAgent(api_key=self._api_key, effort=self._effort)
             self._display_model = "claude-sonnet-4-20250514"
 
         self._last_response: str = ""
@@ -449,7 +457,7 @@ class OISRepl:
             return False
         os.environ["ANTHROPIC_API_KEY"] = key
         self._api_key = key
-        self._agent = OISAgent(api_key=key)
+        self._agent = OISAgent(api_key=key, effort=self._effort)
         saved = self._persist_key_to_env(key)
         if saved:
             console.print(f"\n  [dim]Key saved to {saved} and active for this session.[/]\n")
@@ -512,9 +520,38 @@ class OISRepl:
                         self._ollama_host,
                         self._is_pdf_disabled,
                         self._openai_base_url,
+                        self._effort,
                     )
                     continue
 
+                # 'effort' shows the current level; 'effort <level>' sets it. Only
+                # treat 'effort <arg>' as a command when <arg> is a real level, so
+                # a genuine query that happens to start with 'effort' still runs.
+                effort_parts = user_input.split(maxsplit=1)
+                if effort_parts[0].lower() == "effort" and (
+                    len(effort_parts) == 1
+                    or effort_parts[1].strip().lower() in {lvl["v"] for lvl in _effort.LEVELS}
+                ):
+                    if len(effort_parts) == 2:
+                        chosen = effort_parts[1].strip().lower()
+                        self._effort = chosen
+                        self._agent.set_effort(chosen)
+                        console.print(
+                            f"  [dim]Sweep effort set to[/] [#00ff88]{chosen}[/][dim]. "
+                            f"{_effort.describe(chosen)}[/]\n"
+                        )
+                    else:
+                        console.print(f"\n  [bold]Sweep effort:[/] [#00ff88]{self._effort}[/]")
+                        for lvl in _effort.LEVELS:
+                            marker = "[#00ff88]›[/]" if lvl["v"] == self._effort else " "
+                            console.print(
+                                f"  {marker} [#00ff88]{lvl['v']:<9}[/][dim]{lvl['desc']}[/]"
+                            )
+                        console.print(
+                            "  [dim]Set with: effort faster | balanced | deeper[/]\n"
+                        )
+                    continue
+
                 if user_input.lower() == "save":
                     if self._last_response:
                         path = _save_report(self._last_response)