#21072: feat: add compaction retry config (maxAttempts, retryDelayMs)

## Summary - **Problem:** Pre-emptive compaction hardcoded `maxAttempts=1`, causing session resets on transient provider errors (OAuth 401, timeouts) - **Why it matters:** Single transient failure kills active conversations; overflow path had 3 retries but pre-emptive didn't - **What changed:** Added `compaction.maxAttempts` (default: 3) and `compaction.retryDelayMs` (default: 1000ms) config options; both compaction paths now use config values - **What did NOT change:** Default retry behavior improved from 1 to 3 attempts; no breaking changes to existing sessions ## Change Type (select all) - [x] Feature ## Scope (select all touched areas) - [x] Gateway / orchestration ## Linked Issue/PR - Closes #20873 ## User-visible / Behavior Changes New optional config fields: ```json { "agents": { "defaults": { "compaction": { "maxAttempts": 3, "retryDelayMs": 1000 } } } } ``` - Pre-emptive compaction now retries 3 times by default (was 1) - Overflow compaction uses configured `maxAttempts` (was hardcoded 3) - Retry attempts logged with `[compaction-retry]` prefix ## Security Impact (required) - New permissions/capabilities? **No** - Secrets/tokens handling changed? **No** - New/changed network calls? **No** (just retries existing calls) - Command/tool execution surface changed? **No** - Data access scope changed? **No** ## Repro + Verification ### Environment - OS: Any - Runtime: Node 22+ - Provider: OAuth-based (e.g., google-antigravity) or any with transient failures - Config: `compaction.mode: "safeguard"` ### Steps 1. Configure compaction with new options (or rely on defaults) 2. Trigger pre-emptive compaction (reach `maxHistoryShare` threshold) 3. Simulate transient provider error (OAuth 401, timeout) ### Expected - Compaction retries up to `maxAttempts` times with `retryDelayMs` delay - Session continues if retry succeeds ### Actual - ✓ Retry logic executes with configured attempts - ✓ Warnings logged on retry - ✓ Defaults applied when fields omitted ## Evidence - [x] Config schema updated with JSDoc and Zod validation - [x] Both compaction paths (pre-emptive and overflow) use config values - [x] `retryAsync` utility handles retry logic ## Human Verification (required) - Verified scenarios: - Config schema accepts new fields - Defaults applied when omitted - TypeScript compilation clean - Edge cases checked: - Zero/negative values rejected by Zod - Existing sessions unaffected - What you did **not** verify: - Live OAuth 401 retry behavior (requires provider setup) ## Compatibility / Migration - Backward compatible? **Yes** - Config/env changes? **No** (optional fields) - Migration needed? **No** <!-- greptile_comment --> <h3>Greptile Summary</h3> Added configurable retry logic for compaction operations to handle transient provider failures (OAuth 401, timeouts). Pre-emptive compaction now defaults to 3 retry attempts (was hardcoded to 1), and overflow compaction uses the same configurable value. **Key changes:** - New config fields: `compaction.maxAttempts` (default: 3) and `compaction.retryDelayMs` (default: 1000ms) - Both pre-emptive and overflow compaction paths now use `retryAsync` with configured values - Retry warnings logged with `[compaction-retry]` prefix - TypeScript types and Zod schema validation added for new fields **Minor optimization opportunity:** - Setting `jitter: 0` in the retry config would clarify intent for fixed-delay retries (currently gets clamped anyway) <h3>Confidence Score: 5/5</h3> - Safe to merge - backward compatible config change with sensible defaults - Well-structured change with proper config validation, default values maintain backward compatibility (3 attempts is reasonable), and the retry logic integrates cleanly with existing infrastructure. The only issue is a minor style optimization for retry config clarity. - No files require special attention _{Last reviewed commit: 7024b81} <!-- greptile_other_comments_section --> <!-- /greptile_comment -->