experimental: codex / ai integration

TODO.md

@@ -6,6 +6,25 @@ annotations; cached read-only snapshots; persistent unread state; contextual
 keybindings; thread replies and resolution changes; and pull-request metadata
 editing are already implemented.
 
+## Experimental AI follow-up
+
+- Add Claude Code and OpenRouter adapters behind the existing provider
+  interface. Direct API adapters must require no-training and zero-data-
+  retention routing and must never silently fall back to another provider or
+  model.
+- Add local-only proposed reply drafts that the user can inspect and explicitly
+  publish under their own GitHub identity. Publishing is intentionally absent
+  from the initial AI implementation.
+- Add a review-history browser with per-run scope, model, head SHA, exclusions,
+  redaction counts, and deletion/export controls without persisting raw diffs.
+- Improve semantic near-duplicate detection across existing GitHub comments
+  and local findings, while retaining deterministic exact fingerprints.
+- Add configurable sensitive-path policy sets and a per-run file picker before
+  confirmation.
+- Add recorded provider event fixtures and adversarial prompt-injection,
+  ANSI/OSC, path traversal, stale-head, oversized-output, and tool-attempt
+  integration tests.
+
 ## Completed resilience work
 
 - Refreshes render core data before annotations and conflict analysis. A