0.49.23 (omp 18.12.6): prepared turn structure ... comparison=role refusal: dropping a tool call can leave two adjacent assistants, and omp's Anthropic Continue. repair row isn't modeled #789

Summary

On a resumed omp session (Anthropic Messages API, dispatch fence on, side_contexts=0), every turn is refused:

refused (provider request does not match this pass's managed dispatch (prepared turn structure);
row=22 expectedRows=611 actualRows=616 expectedRole=assistant actualRole=user
expectedCalls=1 actualCalls=0 expectedResults=0 actualResults=0 comparison=role)

Retrying never helps, because the wire always has exactly 5 more rows than the prepared structure.

Environment

• @cortexkit/pi-magic-context 0.49.23
• omp 18.12.6, provider api anthropic-messages (claude-sonnet-5-5)
• Session of ~1,840 entries; it contains 95 todo tool calls

Cause (reproduced)

I reproduced it in an isolated HOME (copied context.db, fake Anthropic endpoint) and instrumented the copy so it dumps active.preparedStructure together with the actual payload at the fence:

1. MC's transform output keeps only 19 of the session's 95 todo tool calls. It removes the others together with their results.
2. Five assistant messages consisted of one short text block plus a single todo call (e.g. "Refreshing the todo list, dispatching next." + todo). With the call removed, each becomes a text-only assistant, and the next row in MC's output is also an assistant.
3. omp's Anthropic converter (@oh-my-pi/pi-ai providers/anthropic.ts, the "Dropped empty user/developer turns can leave two assistant params adjacent" loop) repairs adjacent assistant params by inserting { role: "user", content: "Continue." } between them.
4. The prepared structure doesn't model that repair. The payload has a user "Continue." row at index 16, 181, 238, 312 and 426, which is 5 extra rows. The first one shifts the comparison so that it fails at row 22 with expectedRole=assistant actualRole=user.

Every pass reproduces it, since the adjacency is produced by MC's own output.

Suggested direction

Either of these would fix it:

• When removing tool calls leaves an assistant message adjacent to another assistant, don't emit that shape: merge the two, drop the remnant, or keep the call. Or
• Model the provider-side adjacent-assistant repair in the prepared Anthropic structure, the same way other host-inserted rows are modeled (compare 0.49.23 (omp 18.12.6): every turn refused on resumed sessions: prepared turn structure ... expectedRole=missing actualRole=system comparison=count (trailing omp tool-control system message isn't modeled) #785 for the trailing tool-control system row), or degrade it as Design proposal: tier the dispatch verdict so unmodeled shapes degrade the check, not the turn #761 proposes.