Merge pull request #114 from FSoft-AI4Code/fix/whole-repo-update-loop

updater: stop whole-repository --update from looping (#113)

Author: anhnh2002

codewiki/src/be/agent_tools/generate_sub_module_documentations.py

@@ -3,7 +3,7 @@
 from pydantic_ai import RunContext, Tool, Agent
 
 from codewiki.src.be.agent_tools.deps import CodeWikiDeps
-from codewiki.src.be.module_naming import normalize_sub_module_specs
+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
@@ -25,9 +25,10 @@ async def generate_sub_module_documentation(
         sub_module_specs: A dictionary mapping sub-module names to their core component IDs.
             Example: {"authentication": ["auth_handler.py::AuthHandler", "auth_middleware.py::verify_token"], "database": ["db_client.py::DBClient", "models.py::UserModel"]}
             Each key is a descriptive sub-module name, and the value is a list of component IDs from the current module's core components that belong to that sub-module.
-            Sub-module names must be unique across the whole wiki; a name that is already used by another
-            module is automatically prefixed with the current module name, and the tool result reports the
-            final file names actually saved.
+            Sub-module names must be unique across the whole wiki. A name already used by another module is
+            prefixed with the current module name; a name that is already documented (plain or prefixed)
+            is SKIPPED, not regenerated. The tool result reports the final file names actually saved and
+            every skipped request.
     """
 
     deps = ctx.deps
@@ -38,15 +39,21 @@ async def generate_sub_module_documentation(
 
     # Resolve name collisions against the module tree and files already on disk
     # before touching the tree (issue #76): docs live in one flat directory.
-    name_map = normalize_sub_module_specs(
+    # A request that is already documented (plain or parent-prefixed name) is
+    # skipped rather than renamed x_2, x_3, ... (issue #113).
+    plan = plan_sub_module_specs(
         sub_module_specs,
         previous_module_name,
         deps.module_tree,
         deps.absolute_docs_path,
     )
+    name_map = plan.name_map
+    if not name_map:
+        return _skipped_report(plan.skipped, deps.current_module_name)
     final_specs = {
         name_map[requested_name]: core_component_ids
         for requested_name, core_component_ids in sub_module_specs.items()
+        if requested_name in name_map
     }
 
     # add the sub-module to the module tree
@@ -135,9 +142,22 @@ async def generate_sub_module_documentation(
     if missing:
         report += f" MISSING (generation did not produce these files): {', '.join(missing)}."
         logger.warning("Sub-module documentation missing after generation: %s", ", ".join(missing))
+    if plan.skipped:
+        report += " " + _skipped_report(plan.skipped, deps.current_module_name)
     return report
 
 
+def _skipped_report(skipped: dict[str, str], current_module_name: str) -> str:
+    """Tell the parent agent, unambiguously, not to retry skipped sub-modules."""
+    if not skipped:
+        return "No sub-modules were generated."
+    items = ", ".join(f"'{name}' ({reason})" for name, reason in skipped.items())
+    return (
+        f"Skipped sub-modules: {items}. Do NOT call generate_sub_module_documentation again "
+        f"for these; link the existing pages from `{current_module_name}.md` instead."
+    )
+
+
 generate_sub_module_documentation_tool = Tool(
     function=generate_sub_module_documentation,
     name="generate_sub_module_documentation",

codewiki/src/be/caw_toolkit.py

@@ -28,7 +28,7 @@
 from mcp.server.fastmcp import Context
 
 from codewiki.src.be.agent_tools.deps import CodeWikiDeps
-from codewiki.src.be.module_naming import normalize_sub_module_specs
+from codewiki.src.be.module_naming import plan_sub_module_specs
 
 if TYPE_CHECKING:
     from codewiki.src.be.caw_backend import CawBackend
@@ -242,9 +242,10 @@ async def str_replace_editor(
             "sub_module_specs: a dictionary mapping sub-module names to their core component IDs. "
             "Example: {'authentication': ['auth_handler.py::AuthHandler'], "
             "'database': ['db_client.py::DBClient']}\n"
-            "Sub-module names must be unique across the whole wiki; a name already used by another "
-            "module is automatically prefixed with the current module name, and the tool result "
-            "reports the final file names actually saved."
+            "Sub-module names must be unique across the whole wiki. A name already used by another "
+            "module is prefixed with the current module name; a name that is already documented "
+            "(plain or prefixed) is SKIPPED, not regenerated. The result reports the files actually "
+            "saved and every skipped request."
         )
     )
     async def generate_sub_module_documentation(
@@ -284,15 +285,21 @@ def _run_sub_modules(self, sub_module_specs: dict[str, list[str]]) -> str:
 
         # Resolve name collisions against the module tree and files already on
         # disk before touching the tree (issue #76): docs live in one flat directory.
-        name_map = normalize_sub_module_specs(
+        # A request that is already documented (plain or parent-prefixed name) is
+        # skipped rather than renamed x_2, x_3, ... (issue #113).
+        plan = plan_sub_module_specs(
             sub_module_specs,
             previous_module_name,
             deps.module_tree,
             deps.absolute_docs_path,
         )
+        name_map = plan.name_map
+        if not name_map:
+            return _skipped_report(plan.skipped, deps.current_module_name)
         final_specs = {
             name_map[requested_name]: core_ids
             for requested_name, core_ids in sub_module_specs.items()
+            if requested_name in name_map
         }
 
         # Add sub-modules to the in-memory module tree.
@@ -352,4 +359,17 @@ def _run_sub_modules(self, sub_module_specs: dict[str, list[str]]) -> str:
             logger.warning(
                 "Sub-module documentation missing after generation: %s", ", ".join(missing)
             )
+        if plan.skipped:
+            report += " " + _skipped_report(plan.skipped, deps.current_module_name)
         return report
+
+
+def _skipped_report(skipped: dict[str, str], current_module_name: str) -> str:
+    """Tell the parent agent, unambiguously, not to retry skipped sub-modules."""
+    if not skipped:
+        return "No sub-modules were generated."
+    items = ", ".join(f"'{name}' ({reason})" for name, reason in skipped.items())
+    return (
+        f"Skipped sub-modules: {items}. Do NOT call generate_sub_module_documentation again "
+        f"for these; link the existing pages from `{current_module_name}.md` instead."
+    )

codewiki/src/be/module_naming.py

@@ -8,6 +8,7 @@
 """
 
 import os
+from dataclasses import dataclass, field
 from typing import Any, Dict, List, Optional, Set
 
 import logging
@@ -97,6 +98,66 @@ def normalize_sub_module_specs(
     return name_map
 
 
+@dataclass
+class SubModulePlan:
+    """What ``generate_sub_module_documentation`` will actually create.
+
+    ``name_map``: requested name -> final, unique file stem to generate.
+    ``skipped``: requested name -> reason it is not generated (already
+    documented under that name or its parent-prefixed variant).
+    """
+
+    name_map: Dict[str, str] = field(default_factory=dict)
+    skipped: Dict[str, str] = field(default_factory=dict)
+
+
+def plan_sub_module_specs(
+    sub_module_specs: Dict[str, Any],
+    parent_name: Optional[str],
+    module_tree: Dict[str, Any],
+    working_dir: str,
+) -> SubModulePlan:
+    """Decide which requested sub-modules to generate (issue #113).
+
+    Like :func:`normalize_sub_module_specs`, a name that collides with the
+    tree, a ``.md`` on disk or a reserved stem gets the parent prefix. Unlike
+    it, a request whose plain *and* prefixed names are both taken is treated
+    as a repeat of something already documented and skipped, never renamed
+    with a numeric suffix: that suffixing is what let one agent regenerate the
+    same modules as ``x_2``, ``x_3``, ... without ever converging.
+    """
+    taken = collect_module_tree_names(module_tree)
+    taken |= _existing_doc_stems(working_dir)
+    taken |= RESERVED_STEMS
+
+    plan = SubModulePlan()
+    for requested_name in sub_module_specs:
+        name = sanitize_module_name(requested_name)
+        if name not in taken:
+            final_name = name
+        else:
+            prefixed = f"{sanitize_module_name(parent_name)}_{name}" if parent_name else name
+            if prefixed in taken:
+                existing = name if name in taken else prefixed
+                reason = f"already documented as {existing}.md; do not request it again"
+                plan.skipped[requested_name] = reason
+                logger.info(
+                    "Sub-module '%s' already documented as '%s'; skipping duplicate request.",
+                    requested_name,
+                    existing,
+                )
+                continue
+            final_name = prefixed
+            logger.info(
+                "Sub-module name '%s' collides with an existing module or file; renamed to '%s'.",
+                requested_name,
+                final_name,
+            )
+        taken.add(final_name)
+        plan.name_map[requested_name] = final_name
+    return plan
+
+
 def dedupe_module_tree_names(module_tree: Dict[str, Any]) -> Dict[str, Any]:
     """Sanitize and uniquify all module names in a freshly clustered tree.
 

codewiki/src/be/updater/orchestrator.py

@@ -9,7 +9,7 @@
 from typing import Any
 
 from codewiki.src.be.backend import LLMBackend
-from codewiki.src.be.cluster_modules import cluster_modules
+from codewiki.src.be.cluster_modules import cluster_modules, get_clustering_input_token_count
 from codewiki.src.be.dependency_analyzer.models.core import Node
 from codewiki.src.be.updater import pages as P
 from codewiki.src.be.updater import tree as T
@@ -41,7 +41,12 @@
 )
 from codewiki.src.be.updater.routing import RoutingAgent
 from codewiki.src.be.updater.stale_scan import StaleScanner
-from codewiki.src.be.updater.tree_repair import RepairResult, repair_tree
+from codewiki.src.be.updater.tree_repair import (
+    RULE_AGENT,
+    RepairResult,
+    RoutingDecision,
+    repair_tree,
+)
 from codewiki.src.config import MODULE_TREE_FILENAME, Config
 from codewiki.src.utils import file_manager
 
@@ -85,6 +90,17 @@ def _load_state(self, old_graph_path: str) -> tuple[dict[str, Node], dict[str, A
             self.record.detector_notes.append("whole-repository mode: one virtual leaf (overview)")
         return old_graph, tree
 
+    def _route_to_overview(
+        self, orphans: list[str], context: dict[str, Any]
+    ) -> list[RoutingDecision]:
+        """Whole-repository mode: every orphan belongs to the single page."""
+        return [
+            RoutingDecision(
+                cid, RULE_AGENT, (P.OVERVIEW_STEM,), detail="whole-repository mode: single page"
+            )
+            for cid in orphans
+        ]
+
     def _module_path(self, path: tuple[str, ...]) -> list[str]:
         if self.whole_repo and path == (P.OVERVIEW_STEM,):
             return []
@@ -253,6 +269,12 @@ async def _run(
             if self.opts.use_routing_agent
             else None
         )
+        if self.whole_repo:
+            # One page documents everything: orphans go to the virtual leaf
+            # without an LLM call and without creating leaves. Each created
+            # leaf would otherwise map back onto the overview page and rewrite
+            # it once more in Step 5 (issue #113).
+            router = self._route_to_overview
         repair: RepairResult = repair_tree(
             old_tree, diff, new_graph, tracked_new, self.opts, route_orphans=router
         )
@@ -288,11 +310,43 @@ async def _run(
         ratios["fired"] = (
             ratios["r_leaf"] >= self.opts.tau_full or ratios["r_tree"] >= self.opts.tau_tree
         )
-        if self.whole_repo and ratios["fired"]:
-            # One virtual leaf: any change is 100% active by construction, and
-            # patching that single page is exactly the incremental step.
-            ratios["fired"] = False
-            ratios["note"] = "whole-repository mode: fallback rule not applied"
+        if self.whole_repo:
+            # One virtual leaf: any change is 100% active by construction, so
+            # r_leaf says nothing. What matters is whether the *current* scope
+            # still fits one page under the same threshold a fresh build uses
+            # to skip clustering. A whole-repo baseline built from a narrow
+            # --include and then updated against the full repo does not, and
+            # feeding that to the single-page agent runs unbounded (issue #113).
+            scope = [c for c in leaf_nodes if c in new_graph]
+            tokens = get_clustering_input_token_count(scope, new_graph)
+            threshold = self.config.max_token_per_module
+            ratios["clustering_tokens"] = tokens
+            ratios["tau_cluster"] = threshold
+            if tokens <= threshold:
+                ratios["fired"] = False
+                note = "whole-repository mode: scope still fits one page; fallback rule not applied"
+                ratios["note"] = note
+                logger.info(
+                    "Whole-repository mode: current scope is %d clustering tokens "
+                    "(threshold %d, %d leaf nodes); updating the single page in place",
+                    tokens,
+                    threshold,
+                    len(scope),
+                )
+            else:
+                ratios["fired"] = True
+                ratios["note"] = (
+                    "whole-repository baseline but current scope exceeds the clustering "
+                    "threshold; a fresh build would cluster, so fall back to a full build"
+                )
+                rec.detector_notes.append(ratios["note"])
+                logger.warning(
+                    "Whole-repository baseline cannot be updated in place: current scope is "
+                    "%d clustering tokens (threshold %d, %d leaf nodes); falling back to a full build",
+                    tokens,
+                    threshold,
+                    len(scope),
+                )
         rec.fallback = ratios
         if ratios["fired"]:
             rec.outcome = OUTCOME_FULL_FALLBACK
@@ -311,6 +365,13 @@ async def _run(
 
         # ---- Step 5: sequential leaf agents
         order = order_active(active, new_tree, new_graph)
+        if self.whole_repo and len(order) > 1:
+            # Every active unit is the same overview page; regenerate it once.
+            overview_path = (P.OVERVIEW_STEM,)
+            order = [overview_path if overview_path in reports else order[0]]
+            rec.detector_notes.append(
+                f"whole-repository mode: {len(active)} active units collapsed into one overview run"
+            )
         dep = T.leaf_dependents(new_tree, new_graph)
         inv = inverse(ref_index)
         rec.active = [