Add ToolOutputManagement capability

PLAN.md

@@ -0,0 +1,57 @@
+# Plan: ToolOutputManagement capability
+
+Closes #82
+
+## Context
+
+When tools return large outputs (file contents, search results, command output), they can consume most of the context window. A single `grep -r` or verbose test output can crowd out all useful conversation history.
+
+Issue #82 originally proposed this as harness infrastructure baked into the agent loop, but per maintainer feedback, this is better implemented as a configurable capability that uses the existing `after_tool_execute` hook on `AbstractCapability`.
+
+## Design
+
+A `ToolOutputManagement` capability (dataclass extending `AbstractCapability`) that:
+
+1. **Intercepts tool results** via `after_tool_execute` -- the standard hook that fires after every tool execution, before the result enters the model's context
+2. **Measures output size** by converting the result to its string representation
+3. **Truncates when over limit** using one of three strategies:
+   - `head` -- keep first N chars
+   - `tail` -- keep last N chars (good for build/test output)
+   - `head_tail` (default) -- keep first 60% + last 40% with middle elided
+4. **Supports per-tool overrides** via `per_tool_limits` and `per_tool_strategies` dicts
+5. **Supports custom summarization** via an optional `summarize_fn(tool_name, output) -> str` (sync or async), with truncation as a safety net if the summary still exceeds the limit
+
+### What it does NOT do (deliberately)
+
+- **Spill-to-file**: The issue proposed writing full output to files and returning paths. This requires filesystem access and assumptions about the execution environment. Better to leave this to a more specialized capability or to the `summarize_fn` hook.
+- **Token-based limits**: Character limits are a simple, model-independent proxy. Token counting requires model-specific tokenizers and adds complexity. Can be added later.
+- **Model-aware scaling**: Adjusting limits based on `ModelProfile.context_window` is a good idea but depends on #35 (ContextWindowTracker) which doesn't exist yet.
+
+### Key property: original preserved upstream
+
+The `after_tool_execute` hook modifies only what the model sees. The original full result is already captured in telemetry/trajectory before this hook fires, so no data is lost.
+
+## Files
+
+- `src/pydantic_harness/tool_output_management.py` -- the capability
+- `src/pydantic_harness/__init__.py` -- re-exports `ToolOutputManagement` and `TruncationStrategy`
+- `tests/test_tool_output_management.py` -- unit + integration tests (27 tests)
+
+## Usage
+
+```python
+from pydantic_ai import Agent
+from pydantic_harness import ToolOutputManagement, TruncationStrategy
+
+agent = Agent(
+    'openai:gpt-4o',
+    capabilities=[
+        ToolOutputManagement(
+            max_output_chars=8000,
+            strategy=TruncationStrategy.head_tail,
+            per_tool_limits={'bash': 4000},
+            per_tool_strategies={'bash': TruncationStrategy.tail},
+        ),
+    ],
+)
+```

src/pydantic_harness/__init__.py

@@ -7,4 +7,9 @@
 # Each capability module is imported and re-exported here.
 # Capabilities are listed alphabetically.
 
-__all__: list[str] = []
+from pydantic_harness.tool_output_management import ToolOutputManagement, TruncationStrategy
+
+__all__: list[str] = [
+    'ToolOutputManagement',
+    'TruncationStrategy',
+]