[chengyongru]

Summary

• Reuse the existing in-flight context governor to catch current-turn tool output that cannot fit the next model request.
• Replace the model-facing tool result with a bounded, actionable instruction before calling the provider.
• Let the normal agent loop ask the model to retry with narrower arguments, use another tool, or explain the context limitation to the user.

Fixes #2343.

Behavior

After a tool call completes, the next normal agent iteration already runs ContextGovernor.prepare_for_model() before making a provider request. If the estimated prompt exceeds the configured input budget, compact_inflight_overflow() replaces oversized in-flight tool results in the model-facing copy.

The replacement now tells the model that:

1. the previous result was too large to fit the available context;
2. the same call should not be repeated unchanged;
3. it should retry with a narrower path, query, range, or result limit, or use another tool;
4. if no smaller approach is possible, it should tell the user that the task cannot fit in the available context.

Tools remain available on the next ordinary model call. The context governor continues to leave persisted session history untouched and keeps the compacted model-facing result stable for the rest of the turn.

Deliberate boundaries

• No provider-error classification or provider-specific parsing.
• No nested model retry or second recovery state machine.
• No retry/fallback/circuit-breaker changes.
• No new stop_reason, cross-turn state, configuration, or checkpoint behavior.

Tests

• python -m pytest tests/agent/test_runner_governance.py -q — 28 passed.
• python -m pytest tests/agent/test_runner_core.py tests/agent/test_runner_governance.py tests/agent/test_loop_runner_integration.py -q — 59 passed.
• Focused runner smoke: a 50,000-character read_file result was replaced before the next provider request; the model then issued a narrower offset/limit call and completed normally.
• Targeted Ruff checks and git diff --check pass.