fix(slackbot): keep streamed answer when a plan block is finalized (Codex answers dropped)

[jalagrange]

Problem

When the Codex harness produces a final answer, the Slack reply shows only the "Thinking" widget and no answer text. The agent completes successfully and the answer is streamed, but it never appears in the closed message. The Claude Code harness is unaffected.

Concretely, the bot message ends up with blocks [rich_text, plan] (header + Thinking) and the answer is gone.

Root cause

Slack streaming is startStream → appendStream (chunks) → stopStream (final blocks). The stopStream blocks replace the streamed message body; they do not append to the accumulated chunks.

In closeTextStream() (services/slackbot/src/slack/agent-session.ts), the answer markdown was only included in the final blocks when it was not streamed live:

...(!streamedTextLive && answerMarkdown ? renderMarkdownBlocks(answerMarkdown) : [])

That is correct only when stopStream is called with no blocks (the accumulated stream chunks then stand on their own). But plan-emitting harnesses (Codex always renders a "Thinking" plan block) cause stopStream to be called with a plan block, which replaces the streamed body and drops the live-streamed answer.

Claude Code does not emit a plan block, so stopStream is called with no blocks and its streamed answer survives — which is why only Codex is affected.

Fix

When we emit any composed blocks (e.g. a plan block), also re-include the answer markdown, since those blocks replace the streamed chunks. Only omit the answer blocks when sending no blocks at all and the answer was already streamed live.

const planBlocks = tasks.length && !segment.planStarted ? [planBlock(...)] : []
const includeAnswerBlocks =
  Boolean(answerMarkdown) && (planBlocks.length > 0 || !streamedTextLive)

Verification

Verified end-to-end on a live deployment. A Codex Slack reply that previously rendered [rich_text, plan] (no answer) now renders [rich_text, plan, rich_text, table] with the answer table present.

All existing agent-session and codex-session tests pass (45 across the two suites). The existing "does not duplicate live plan or streamed answer markdown on finalize" test still passes — its scenario has the plan already streamed (planStarted=true), so no plan block is emitted at finalize and the answer is correctly left as the streamed chunks.

I attempted a focused unit regression test, but the exact trigger state (a plan/Thinking block emitted at finalize while the answer streamed live) depends on reasoning-summary publish ordering that was non-trivial to reproduce synthetically at the unit level. Happy to add one with guidance on the preferred test shape.