fix(context): enforce canonical rendered character budget

packages/infinity_context_contracts/infinity_context_contracts/features/context_building.py

@@ -64,13 +64,19 @@ class ContextBudgetDto:
     reserved_response_tokens: int = 0
     max_items: int | None = None
     strategy: str = "balanced"
+    max_rendered_chars: int | None = None
 
     def to_dict(self) -> JsonObject:
         return {
             "max_context_tokens": self.max_context_tokens,
             "reserved_response_tokens": self.reserved_response_tokens,
             "max_items": self.max_items,
             "strategy": self.strategy,
+            **(
+                {"max_rendered_chars": self.max_rendered_chars}
+                if self.max_rendered_chars is not None
+                else {}
+            ),
         }
 
 

packages/infinity_context_core/infinity_context_core/features/context_building/application/handlers.py

@@ -19,6 +19,7 @@
     ContextBundle,
     ContextEvidenceRenderer,
     ContextQueryExpansionPolicy,
+    PromptSectionPlan,
     PromptSectionPlanner,
 )
 from infinity_context_core.features.context_building.ports import (
@@ -40,9 +41,24 @@ class PackContextHandler:
     )
 
     async def execute(self, query: PackContextQuery) -> PackContextResult:
-        plan = self.budget_policy.plan(query.candidates, query.budget)
+        plan = self.budget_policy.plan(
+            query.candidates,
+            query.budget,
+            render=lambda items: self.evidence_renderer.render_plan(
+                self.prompt_section_planner.plan(items)
+            ),
+        )
         prompt_section_plan = self.prompt_section_planner.plan(plan.selected_items)
         rendered_evidence = self.evidence_renderer.render_plan(prompt_section_plan)
+        max_rendered_chars = query.budget.max_rendered_chars
+        if max_rendered_chars is not None and len(rendered_evidence) > max_rendered_chars:
+            # Custom rendering may change between admission and final rendering.
+            # Fail explicitly for selected evidence; never truncate or discard it.
+            if plan.selected_items:
+                raise ValueError("Final rendered evidence exceeds max_rendered_chars")
+            # Empty output has a canonical representation, bypassing custom overhead.
+            prompt_section_plan = PromptSectionPlan(sections=())
+            rendered_evidence = ""
         bundle = ContextBundle(
             query=query.query,
             items=plan.selected_items,

packages/infinity_context_core/infinity_context_core/features/context_building/domain/budget.py

@@ -2,6 +2,7 @@
 
 from __future__ import annotations
 
+from collections.abc import Callable
 from dataclasses import dataclass
 
 from infinity_context_core.features.context_building.domain.context import (
@@ -12,13 +13,21 @@
 
 @dataclass(frozen=True, slots=True)
 class ContextBudget:
-    """Prompt token budget with explicit reserves outside memory evidence."""
+    """Independent token and optional rendered Unicode character limits.
+
+    Token estimates and reserves retain their existing meaning. A character cap
+    measures the complete evidence rendering with Python's ``len(str)``; zero
+    allows an empty bundle and ``None`` preserves uncapped feature callers.
+    """
 
     max_prompt_tokens: int
     reserved_response_tokens: int = 0
     reserved_system_tokens: int = 0
+    max_rendered_chars: int | None = None
 
     def __post_init__(self) -> None:
+        if self.max_rendered_chars is not None and self.max_rendered_chars < 0:
+            raise ValueError("Rendered character budget cannot be negative")
         if self.max_prompt_tokens < 1:
             raise ValueError("Context budget must allow at least one prompt token")
         if self.reserved_response_tokens < 0:
@@ -57,7 +66,13 @@ def plan(
         self,
         items: tuple[ContextItem, ...],
         budget: ContextBudget,
+        *,
+        render: Callable[[tuple[ContextItem, ...]], str] | None = None,
     ) -> ContextPackingPlan:
+        if budget.max_rendered_chars is not None and render is None:
+            from .rendering import ContextEvidenceRenderer
+
+            render = ContextEvidenceRenderer().render
         remaining = budget.available_evidence_tokens
         selected: list[ContextItem] = []
         dropped: list[ContextDroppedItem] = []
@@ -70,6 +85,21 @@ def plan(
         for _, item in ranked_items:
             token_cost = item.token_cost
             if token_cost <= remaining:
+                # Measure the actual proposed rendering, including section headings,
+                # escaped quotes, citations and numbering. Keep evidence records whole.
+                if (
+                    budget.max_rendered_chars is not None
+                    and render is not None
+                    and len(render((*selected, item))) > budget.max_rendered_chars
+                ):
+                    dropped.append(
+                        ContextDroppedItem(
+                            item_id=item.item_id,
+                            reason="rendered_character_budget_exceeded",
+                            estimated_tokens=token_cost,
+                        )
+                    )
+                    continue
                 selected.append(item)
                 remaining -= token_cost
                 used += token_cost

packages/infinity_context_server/infinity_context_server/api/v1/context.py

@@ -428,6 +428,11 @@ async def _build_context_bundle(
             as_of=request.as_of,
             budget=context_building_server.ContextBudgetHttpRequest(
                 max_context_tokens=request.token_budget,
+                max_rendered_chars=(
+                    max_rendered_chars
+                    if max_rendered_chars is not None
+                    else container.settings.max_context_chars
+                ),
                 max_items=canonical_item_limit,
             ),
             tags=(),

packages/infinity_context_server/infinity_context_server/features/context_building/contracts.py

@@ -18,6 +18,7 @@ class ContextBudgetHttpRequest(BaseModel):
 
     max_context_tokens: int = Field(default=1800, ge=1, le=64000)
     reserved_response_tokens: int = Field(default=0, ge=0, le=64000)
+    max_rendered_chars: int | None = Field(default=None, ge=0)
     max_items: int | None = Field(default=None, ge=1, le=1000)
     strategy: str = Field(default="balanced", min_length=1, max_length=80)
 
@@ -27,6 +28,7 @@ def to_contract(self) -> ContextBudgetDto:
             reserved_response_tokens=self.reserved_response_tokens,
             max_items=self.max_items,
             strategy=self.strategy,
+            max_rendered_chars=self.max_rendered_chars,
         )
 
 

packages/infinity_context_server/infinity_context_server/features/context_building/mappers.py

@@ -43,6 +43,7 @@ def build_context_query_from_contract(
         budget=context_building.ContextBudget(
             max_prompt_tokens=budget.max_context_tokens,
             reserved_response_tokens=budget.reserved_response_tokens,
+            max_rendered_chars=budget.max_rendered_chars,
         ),
         candidate_limit=budget.max_items or DEFAULT_CANDIDATE_LIMIT,
     )
@@ -97,6 +98,11 @@ def _budget_from_contract(
         ),
         max_items=_optional_positive_int(budget.get("max_items"), "budget.max_items"),
         strategy=str(budget.get("strategy") or "balanced"),
+        max_rendered_chars=(
+            _non_negative_int(budget["max_rendered_chars"], "budget.max_rendered_chars")
+            if budget.get("max_rendered_chars") is not None
+            else None
+        ),
     )