Summary

Fixes #24800 — auto-compaction is not triggered during tool-use loops.

During long tool loops (LLM call → tool exec → tool result → repeat), context grows continuously but compaction only triggers on agent_end events or the next user prompt(). If context fills the window mid-loop, the API rejects the request — wasting the call and surfacing an error to the user.

This PR adds a preemptive overflow check to the transformContext hook in tool-result-context-guard.ts, which already runs before every LLM call (even mid-tool-loop). After the existing tool-result compaction (targets 75% of the context window), we re-estimate context size. If it still exceeds 90% of the window, we throw a context overflow error that triggers the existing overflow recovery cascade in run.ts.

How it works (zero changes to run.ts or attempt.ts)

1. transformContext compacts tool results as before (75% budget)
2. New: Post-enforcement check — if context still exceeds 90%, throws Error with message "Preemptive context overflow: ..."
3. Error propagates through prompt() → caught as promptError in attempt.ts
4. run.ts detects it via isLikelyContextOverflowError() (matches "context overflow:" at errors.ts:79)
5. isCompactionFailureError() returns false (message avoids the word "compaction")
6. Takes the explicit compaction branch → compactEmbeddedPiSessionDirect() runs full LLM-based session compaction
7. On success, retries the prompt with compacted context
8. MAX_OVERFLOW_COMPACTION_ATTEMPTS = 3 prevents infinite loops

Why 90%

• The guard already aggressively compacts tool results at 75%
• If context is still at 90%+ after that, non-tool-result content (assistant messages, thinking blocks) dominates — only full LLM compaction can reduce those
• Better to compact proactively at 90% than waste a failing API call at ~100%

Changes

• src/agents/pi-embedded-runner/tool-result-context-guard.ts — add PREEMPTIVE_OVERFLOW_RATIO (0.9), PREEMPTIVE_CONTEXT_OVERFLOW_MESSAGE, and post-enforcement overflow check (+19 lines)
• src/agents/pi-embedded-runner/tool-result-context-guard.test.ts — 3 new tests: throws at 90%+, doesn't throw under 90%, verifies tool results are compacted before overflow check (+60 lines)

Test plan

• pnpm test src/agents/pi-embedded-runner/tool-result-context-guard.test.ts — 11/11 pass (8 existing + 3 new)
• pnpm test src/agents/pi-embedded-runner/run.overflow-compaction.loop.test.ts — 11/11 pass (unchanged)
• pnpm test src/agents/pi-embedded-helpers.isbillingerrormessage.test.ts — 41/41 pass (error classifier tests)
• pnpm check — format, typecheck, lint all clean
• Hand-verified all 8 existing tests stay under 90% threshold after enforcement (no false triggers)
• Verified error message matches isLikelyContextOverflowError but NOT isCompactionFailureError