[OmarHussain09]

Three fixes for the same end-to-end failure: a finished AI Ops report was thrown away and replaced with "I wasn't able to produce the analysis you requested", and what survived to history came back shattered or duplicated. All three are triggered by reasoning models, which return content-block arrays instead of strings and keep every streamed delta as its own block.

1. fix(aiops): promote a reasoning-model deliverable instead of re-synthesizing it

finalNode promotes an already-composed report verbatim to avoid paying for a closing LLM call, but gated that on typeof m.content === 'string'. A reasoning model never returns a string — one 7k-char report arrived as 544 separate text blocks. The gate missed it, so the finished report was dropped and the fallback re-synthesized from toolResults.slice(-3); with only three SSM shell results in view the model said, correctly for what it could see, that it had failed — moments after producing the analysis. The UI then elevated that into the Report card.

Extracted the scan as findRenderedDeliverable() and routed content through contentToText(), which this file already applies in six other places for the same reason.

2. fix(chat): coalesce streamed content blocks when replaying history

reconstructAiContentParts emitted one UI part per persisted block, so a reloaded answer shattered into hundreds of rows split mid-word ('ikes', ' latency <0', '.3', 'ms, ') — 644 for one report. Consecutive same-kind blocks now accumulate into a single part, and whitespace is judged per run rather than per block: Bedrock emits bare-space deltas between mid-word splits, and discarding those individually fused halves together ("no" + " " + "blocking" → "noblocking"). Runs are still trimmed at their ends, so existing leading-whitespace behaviour holds. Applies to every stored thread on reload, not just new runs.

3. fix(history): drop the duplicated promoted answer on reload

Promotion re-emits the deliverable verbatim as a final-phase message. Live, that copy never reaches the client — promotion makes no model call and the route only emits text from on_chat_model_* events. Persistence has no such filter, so both rows were stored and history replayed the report twice: once as execution prose, once inside the Report card.

The later copy is now dropped so a reloaded thread matches the live view. Matching is on text, not phase — a fragmented reasoning-model turn is persisted as a raw content-block array with no phase marker (route.ts only prefixes the marker when content is a string), so the execution copy often has no phase to match on. The two stored copies are byte-identical once the marker is stripped (6420 chars each in the thread this was found in), so exact comparison suffices. A genuine finalNode fallback answer differs from the execution text and is never dropped, keeping "did promotion fire?" readable from the transcript. The 800-char floor mirrors finalNode's own, so ordinary repeated narration is left alone. Wired into both history paths (chat-history and the checkpoint fallback).

Tests

26 new/existing tests across the three touched areas pass:

cd apps/web-ui && bun run test -- tests/history-dedupe-answers.test.ts \
  lib/agent-chat/__tests__/ai-content-parts.test.ts \
  lib/agent/__tests__/rendered-deliverable.test.ts

Test Files  3 passed (3)
     Tests  26 passed (26)

🤖 Generated with Claude Code