Merge pull request 'depth flag now drives constraint defaults' (#33) from feat/depth-presets into main

Reviewed-on: #33

cli/main.py

@@ -40,27 +40,28 @@ log = get_logger("marchwarden.cli")
 async def call_research_tool(
     question: str,
     depth: str,
-    max_iterations: int,
+    max_iterations: Optional[int],
-    token_budget: int,
+    token_budget: Optional[int],
 ) -> ResearchResult:
-    """Spawn the web researcher MCP server and call its `research` tool."""
+    """Spawn the web researcher MCP server and call its `research` tool.
+
+    ``max_iterations`` and ``token_budget`` are optional — when None,
+    the MCP server uses the depth preset (Issue #30).
+    """
     params = StdioServerParameters(
         command=sys.executable,
         args=["-m", "researchers.web.server"],
         env=os.environ.copy(),
     )
+    arguments: dict = {"question": question, "depth": depth}
+    if max_iterations is not None:
+        arguments["max_iterations"] = max_iterations
+    if token_budget is not None:
+        arguments["token_budget"] = token_budget
     async with stdio_client(params) as (read, write):
         async with ClientSession(read, write) as session:
             await session.initialize()
-            result = await session.call_tool(
+            result = await session.call_tool("research", arguments=arguments)
-                "research",
-                arguments={
-                    "question": question,
-                    "depth": depth,
-                    "max_iterations": max_iterations,
-                    "token_budget": token_budget,
-                },
-            )
             # FastMCP returns the tool's string return as a TextContent block.
             payload = result.content[0].text
             return ResearchResult.model_validate_json(payload)
@@ -178,21 +179,20 @@ def cli() -> None:
     "--budget",
     "token_budget",
     type=int,
-    default=20_000,
+    default=None,
-    show_default=True,
+    help="Token budget for the research loop. Overrides the depth preset.",
-    help="Token budget for the research loop.",
 )
 @click.option(
     "--max-iterations",
     type=int,
-    default=5,
+    default=None,
-    show_default=True,
+    help="Max research loop iterations. Overrides the depth preset.",
 )
 def ask(
     question: str,
     depth: str,
-    token_budget: int,
+    token_budget: Optional[int],
-    max_iterations: int,
+    max_iterations: Optional[int],
 ) -> None:
     """Ask the web researcher a QUESTION."""
     console = Console()

researchers/web/agent.py

@@ -25,6 +25,7 @@ from researchers.web.models import (
     OpenQuestion,
     ResearchConstraints,
     ResearchResult,
+    constraints_for_depth,
 )
 from researchers.web.tools import SearchResult, fetch_url, tavily_search
 from researchers.web.trace import TraceLogger
@@ -203,7 +204,10 @@ class WebResearcher:
         Returns:
             A ResearchResult conforming to the v1 contract.
         """
-        constraints = constraints or ResearchConstraints()
+        # If the caller didn't supply explicit constraints, build them
+        # from the depth preset (Issue #30). Callers that DO pass a
+        # ResearchConstraints are taken at their word — explicit wins.
+        constraints = constraints or constraints_for_depth(depth)
         trace = TraceLogger(trace_dir=self.trace_dir)
         start_time = time.time()
         total_tokens = 0

researchers/web/models.py

@@ -41,6 +41,43 @@ class ResearchConstraints(BaseModel):
     )
 
 
+# Depth presets — choosing a depth picks sensible defaults for the
+# constraint fields. Explicit overrides (--max-iterations, --budget,
+# explicit ResearchConstraints) always win over the preset.
+#
+# `balanced` matches the historical defaults so existing callers see
+# no behavior change. `shallow` and `deep` are tuned for "quick lookup"
+# and "thorough investigation" respectively. These are starting points;
+# Phase 3 stress testing will inform calibration.
+DEPTH_PRESETS: dict[str, dict[str, int]] = {
+    "shallow": {"max_iterations": 2, "token_budget": 5_000, "max_sources": 5},
+    "balanced": {"max_iterations": 5, "token_budget": 20_000, "max_sources": 10},
+    "deep": {"max_iterations": 8, "token_budget": 60_000, "max_sources": 20},
+}
+
+
+def constraints_for_depth(
+    depth: str,
+    *,
+    max_iterations: Optional[int] = None,
+    token_budget: Optional[int] = None,
+    max_sources: Optional[int] = None,
+) -> ResearchConstraints:
+    """Build a ResearchConstraints from a depth preset, with optional overrides.
+
+    Any non-None override wins over the preset value. Unknown depths
+    fall back to ``balanced``.
+    """
+    preset = DEPTH_PRESETS.get(depth, DEPTH_PRESETS["balanced"]).copy()
+    if max_iterations is not None:
+        preset["max_iterations"] = max_iterations
+    if token_budget is not None:
+        preset["token_budget"] = token_budget
+    if max_sources is not None:
+        preset["max_sources"] = max_sources
+    return ResearchConstraints(**preset)
+
+
 # ---------------------------------------------------------------------------
 # Output types — Citation
 # ---------------------------------------------------------------------------

researchers/web/server.py

@@ -13,7 +13,7 @@ from mcp.server.fastmcp import FastMCP
 
 from obs import configure_logging, get_logger
 from researchers.web.agent import WebResearcher
-from researchers.web.models import ResearchConstraints
+from researchers.web.models import constraints_for_depth
 
 log = get_logger("marchwarden.mcp")
 
@@ -52,17 +52,18 @@ async def research(
     question: str,
     context: Optional[str] = None,
     depth: str = "balanced",
-    max_iterations: int = 5,
+    max_iterations: Optional[int] = None,
-    token_budget: int = 20000,
+    token_budget: Optional[int] = None,
 ) -> str:
     """Research a question using web search and return a structured answer.
 
     Args:
         question: The question to investigate.
         context: What the caller already knows (optional).
-        depth: Research depth — "shallow", "balanced", or "deep".
+        depth: Research depth — "shallow", "balanced", or "deep". Each
-        max_iterations: Maximum number of search/fetch iterations (1-20).
+            depth picks default max_iterations / token_budget / max_sources.
-        token_budget: Maximum tokens to spend (1000-100000).
+        max_iterations: Override the depth preset for iterations (1-20).
+        token_budget: Override the depth preset for token budget.
 
     Returns:
         JSON string containing the full ResearchResult with answer,
@@ -70,7 +71,8 @@ async def research(
         and cost_metadata.
     """
     researcher = _get_researcher()
-    constraints = ResearchConstraints(
+    constraints = constraints_for_depth(
+        depth,
         max_iterations=max_iterations,
         token_budget=token_budget,
     )

tests/test_models.py

@@ -443,3 +443,56 @@ class TestResearchResult:
             "recency",
         }
         assert set(data["confidence_factors"].keys()) == cf_keys
+
+
+# ---------------------------------------------------------------------------
+# Depth presets (Issue #30)
+# ---------------------------------------------------------------------------
+
+
+from researchers.web.models import DEPTH_PRESETS, constraints_for_depth
+
+
+class TestDepthPresets:
+    def test_shallow_preset(self):
+        c = constraints_for_depth("shallow")
+        assert c.max_iterations == 2
+        assert c.token_budget == 5_000
+        assert c.max_sources == 5
+
+    def test_balanced_preset_matches_historical_defaults(self):
+        # Backward compat: balanced must equal the original ResearchConstraints defaults
+        c = constraints_for_depth("balanced")
+        default = ResearchConstraints()
+        assert c.max_iterations == default.max_iterations == 5
+        assert c.token_budget == default.token_budget == 20_000
+        assert c.max_sources == default.max_sources == 10
+
+    def test_deep_preset(self):
+        c = constraints_for_depth("deep")
+        assert c.max_iterations == 8
+        assert c.token_budget == 60_000
+        assert c.max_sources == 20
+
+    def test_unknown_depth_falls_back_to_balanced(self):
+        c = constraints_for_depth("nonsense")
+        assert c.max_iterations == DEPTH_PRESETS["balanced"]["max_iterations"]
+        assert c.token_budget == DEPTH_PRESETS["balanced"]["token_budget"]
+
+    def test_explicit_overrides_win(self):
+        c = constraints_for_depth(
+            "shallow",
+            max_iterations=10,
+            token_budget=42_000,
+            max_sources=15,
+        )
+        assert c.max_iterations == 10
+        assert c.token_budget == 42_000
+        assert c.max_sources == 15
+
+    def test_partial_override(self):
+        # Only one field overridden — others stay at the preset
+        c = constraints_for_depth("deep", token_budget=100_000)
+        assert c.token_budget == 100_000
+        assert c.max_iterations == 8  # deep preset
+        assert c.max_sources == 20  # deep preset