#19593: feat(compaction): proactive handover before context overflow

## Summary Proactive handover monitors token usage after each agent turn and triggers compaction **before** context overflow occurs, rather than waiting for the API to reject a prompt that's too large. ## Problem Currently, compaction is only triggered reactively when the API returns a context overflow error. By that point the conversation may already be degraded — the agent loses continuity and the user experiences an unexpected interruption. ## Solution After each successful agent turn, estimate current token usage and compare against a configurable threshold. If remaining tokens fall below the threshold, proactively trigger compaction to preserve conversation quality. ### How it works 1. After each turn completes (no error, not aborted), estimate total message tokens via \estimateMessagesTokens()\ 2. Compare remaining tokens (\contextWindow - currentTokens\) against \compaction.proactiveThresholdTokens\ 3. If below threshold **and** at least 5 user messages have occurred (minimum guard), trigger \compactEmbeddedPiSessionDirect()\ 4. On success, inject a system note into the session so the agent is aware of the handover on the next turn 5. Set a cooldown flag to prevent re-triggering within the same run ### Safety guards - **Minimum turn guard**: Requires at least 5 user messages before triggering (avoids compacting near-empty sessions) - **Cooldown flag**: \proactiveHandoverTriggered\ prevents multiple compactions in the same run - **Graceful failure**: If compaction fails, logs a warning and continues normally — no user-facing error - **Opt-in**: Disabled by default; requires explicit configuration ## Configuration \\\yaml agents: defaults: compaction: proactiveThresholdTokens: 20000 # trigger when <= 20k tokens remain \\\ Set to \ <!-- greptile_comment --> <h3>Greptile Summary</h3> This PR implements proactive context compaction that triggers before overflow occurs. After each successful agent turn, it estimates token usage and triggers compaction when remaining tokens fall below a configurable threshold. **Key changes:** - Added `proactiveThresholdTokens` config option to control when proactive handover triggers - Implemented token estimation and threshold check after successful turns - Added safety guards: minimum 5 user messages, cooldown flag, graceful failure handling - Injects notification message into session after successful compaction - Added `proactiveHandover` field to run metadata for observability The implementation is opt-in (disabled by default) and follows existing patterns for reactive compaction. Guards prevent premature triggering on short sessions and multiple compactions within the same run. <h3>Confidence Score: 5/5</h3> - This PR is safe to merge with minimal risk - The implementation is well-designed with multiple safety guards, follows existing code patterns, handles errors gracefully, and is opt-in by default. The logic is straightforward and the changes are isolated to the agent runner. - No files require special attention _{Last reviewed commit: efdf074} <!-- greptile_other_comments_section --> _{(5/5) You can turn off certain types of comments like style [here](https://app.greptile.com/review/github)!} <!-- /greptile_comment -->