0.49.22 (omp 18.13.0): after an aborted turn, every Anthropic turn is refused with prepared turn structure … expectedRows > actualRows comparison=role (empty-text aborted assistant dropped by omp, kept by the fence) #794

Summary

On the Anthropic Messages lane, once a turn has been refused and aborted, every later turn in that session is refused:

stage=provider dispatch … refused (provider request does not match this pass's managed dispatch (prepared turn structure);
row=3 expectedRows=7 actualRows=5 expectedRole=assistant actualRole=user
expectedCalls=0 actualCalls=0 expectedResults=0 actualResults=0 comparison=role)

Retrying, or sending the session a new message, never helps. The wire always has fewer rows than the prepared structure, and the gap equals the number of aborted assistant turns whose content is a single empty text block.

Environment

• @cortexkit/pi-magic-context 0.49.22
• omp 18.13.0, provider api anthropic-messages (Claude)
• Seen in fresh subagent sessions after the first-turn refusal from bug(omp): first turn after a restart is refused — a cancelled tag-snapshot prewarm is treated as a hydration failure (0.49.22) #790, and in main sessions on 18.12.6

What happens

1. A turn is refused (in my case by bug(omp): first turn after a restart is refused — a cancelled tag-snapshot prewarm is treated as a hydration failure (0.49.22) #790's Pi tag hydration: cancelled). Magic Context calls ctx.abort(), and the host persists an assistant message with stopReason: "aborted" and content: [{ "type": "text", "text": "" }].
2. On the next pass, the prepared structure keeps that assistant row.
3. omp's Anthropic converter (@oh-my-pi/pi-ai providers/anthropic.ts, assistant branch of the message conversion) skips text blocks where text.trim().length === 0, and then skips the whole message when blocks.length === 0. The aborted turn is gone from the wire, so the next row is the following user/developer message: expectedRole=assistant actualRole=user.
4. That refusal aborts again. This time the persisted message has content: [], which both sides drop (it matches shouldDropTruncatedThinkingOnlyAssistant), so the gap stays constant and every pass refuses.

Two sessions, each pass in order (the delta stays at the number of ""-text aborted turns):

The same abort on the OpenAI Responses lane (same [{text:""}] content) is not affected, and those sessions recovered on the next served pass.

Suggested direction

• Model the Anthropic converter's drop of an assistant message whose blocks are all whitespace-only text (in addition to the existing truncated-stop rule), or
• don't persist an empty text block on the aborted message the refusal produces, or
• degrade this shape instead of refusing, as Design proposal: tier the dispatch verdict so unmodeled shapes degrade the check, not the turn #761 proposes.

Related: #789 (Anthropic Continue. repair row, where the wire has more rows) and #790 (the refusal that creates the first aborted turn here).