fix(agent): align empty hidden assistant placeholders on wire to eliminate sanitizer spam (#88955)

[JoaoMarcos44]

Summary

Fixes #88955.

In Bot Mode group chats (and active turn redirects / steering), when a member turn is interrupted before any assistant text has been streamed to the UI, _apply_active_turn_redirect() persists an empty role placeholder with display_kind="hidden":

{"role": "assistant", "content": "", "display_kind": "hidden"}

This placeholder is intentional and load-bearing in the session history to maintain proper role alternation (user -> assistant -> user) without leaking unstreamed reasoning or unfinished content into user-visible transcripts.

However, because this empty placeholder persisted in the session history, the pre-call sanitizer repair_empty_non_final_messages() in agent/agent_runtime_helpers.py detected it on every subsequent turn of the session and repeatedly emitted the following warning:

[WARNING] Pre-call sanitizer: healed 1 empty non-final message(s) by substituting placeholder content — an empty-content turn was in the transcript and would 400 the request ('messages must have non-empty content' / INVALID_REQUEST_BODY). Self-recovering the poisoned transcript in memory; no restart needed.

In long-running group chat sessions (e.g. 98+ API turns), this warning fired perpetually on every turn, causing log pollution, unnecessary repeated memory-copy healing passes, and false-positive alarm noise.

---

Root Cause Analysis

1. Persisted History vs. Wire Messages:
   The placeholder {"role": "assistant", "content": "", "display_kind": "hidden"} is kept in messages so that SessionDB, transcripts, and timeline UI renderers can identify the turn as an interrupted redirect without rendering blank speech bubbles.
2. Wire Stripping Order:
   When constructing outgoing api_messages in agent/conversation_loop.py, api_msg.pop("display_kind", None) stripped the display_kind metadata before sanitize_api_messages() was invoked.
3. Indiscriminate Warning in Sanitizer:
   repair_empty_non_final_messages() received an empty non-final assistant row without the display_kind="hidden" context and treated it as unexpected data corruption rather than an intentional alternation placeholder, logging the [WARNING] every single turn.

---

Solution

We implement a clean, two-layer wire alignment:

1. Wire Substitution in agent/conversation_loop.py:
   When constructing outgoing api_messages, before display_kind is stripped, empty hidden assistant placeholders are proactively populated with _INTERRUPTED_PLACEHOLDER ("[response interrupted]").
   • This ensures strict OpenAI/Anthropic/Bedrock API compliance without 400 errors.
   • The persisted transcript in SessionDB remains byte-stable with content="" and display_kind="hidden" for clean UI rendering.
2. Defensive Discrimination in agent/agent_runtime_helpers.py:
   repair_empty_non_final_messages() distinguishes between intentional hidden placeholders and true accidental transcript corruption. It substitutes _INTERRUPTED_PLACEHOLDER cleanly while only logging warnings for unexpected non-hidden empty turns.

---

Verification Plan

Automated Tests

• Added tests/agent/test_interrupted_hidden_placeholder.py covering:
  • Active turn redirects creating display_kind="hidden" placeholders when no text is visible.
  • Silent healing of hidden placeholders without warning logs.
  • Warning logs preserved for true non-hidden corruptions.
  • Multi-turn conversation regression verification (0 recurring warnings).
  • Preserving streamed visible assistant text during interruptions.
• Ran all existing steer, conversation loop, and runtime helper suites:
  • pytest tests/agent/test_interrupted_hidden_placeholder.py -v (5/5 PASSED)
  • pytest tests/run_agent/test_steer.py -v (29/29 PASSED)
  • pytest tests/run_agent/test_run_agent.py -k "scaffold or redirect or display_kind" -v (5/5 PASSED)

[teknium1]

Closing in favor of PR #89544, which just merged. Your PR was submitted first on this issue and your projection-side approach is what we shipped for legacy rows — the follow-up commit in #89544 credits you directly. The writer-side half came from #89525 (stamping api_content on the placeholder at creation), which avoids the sanitizer-warning bookkeeping split. Thanks for the fix and the fast turnaround!