Summary

• harden iMessage echo suppression using scoped messageId matching with longer TTL and normalized text fallback
• suppress reasoning payloads at outbound normalization and route-delivery boundaries so hidden reasoning never reaches channels
• include message.target fallback in message-send suppression extraction (same core fix as fix(agents): support message.target in send extraction #25623/fix(message): include target in reply suppression tracking #25757)

Why this is the right fix

• addresses both sides of the runaway loop: outbound leakage + inbound re-ingestion
• uses defense-in-depth at shared outbound boundaries, not channel-specific hotfixes
• keeps iMessage dedupe robust under delayed polling by preferring provider IDs when available

Tests

• added iMessage echo-cache unit tests (text normalization, ID matching, ID TTL > text TTL)
• added inbound iMessage echo-by-messageId regression test
• added reasoning-suppression tests for outbound normalization and route delivery
• added message.target extraction regression test

Validation

• pnpm check ❌ (fails on pre-existing unrelated tsgo errors in extensions/matrix/src/matrix/send-queue.test.ts and src/auto-reply/reply/followup-runner.test.ts)
• pnpm build ✅
• pnpm test ❌ (single unrelated failure in src/process/exec.test.ts: no-output timer SIGKILL assertion)

Closes #25897
Closes #25383
Supersedes #25623
Supersedes #25757

Greptile Summary

This PR fixes iMessage reasoning echo feedback loops through a defense-in-depth approach that addresses both sides of the runaway loop: outbound reasoning leakage and inbound echo re-ingestion.

• iMessage echo cache hardened: SentMessageCache split into dual maps with separate TTLs — messageId lookups use a 60s TTL (vs 5s for text) to handle delayed polling. Text normalization (CRLF, whitespace) and placeholder ID filtering ("ok"/"unknown") improve matching accuracy.
• Reasoning suppression at shared outbound boundaries: normalizeReplyPayloadsForDelivery and routeReply both filter isReasoning payloads early, preventing hidden thinking text from reaching any channel.
• message.target fallback for send suppression: extractMessagingToolSend now accepts args.target as an alias for args.to, fixing message-send suppression extraction (same core fix as fix(agents): support message.target in send extraction #25623/fix(message): include target in reply suppression tracking #25757).
• Comprehensive tests added: Echo cache unit tests (text normalization, ID matching, TTL behavior), inbound echo-by-messageId regression test, reasoning suppression tests for both outbound paths, and message.target extraction test.

Confidence Score: 5/5

• This PR is safe to merge — changes are well-scoped defensive fixes with comprehensive test coverage at each layer.
• All changes are additive guards (early returns for isReasoning, upgraded cache with backward-compatible interfaces). The dual-TTL echo cache is a clean improvement over the previous single-map approach. Reasoning suppression is placed at shared outbound boundaries rather than channel-specific hotfixes. Tests cover key scenarios including TTL behavior, normalization, and regression cases. No breaking changes to public APIs.
• No files require special attention

_{Last reviewed commit: 90256e9}