Problem

NewBodyPair (backend/pkg/cast/chain_ast.go) prunes ToolCall parts that have a nil FunctionCall from an AI message. It first collects their indices into partsToDelete (ascending), then deletes them in a forward loop:

for _, id := range partsToDelete {
    aiMsg.Parts = append(aiMsg.Parts[:id], aiMsg.Parts[id+1:]...)
}

Each append-based deletion shifts every later element left by one, but the loop keeps using the original, now-stale indices. With two or more collected indices, the second (and later) deletions remove the wrong elements.

Concrete failure

Input parts [ToolCall{FC:nil}, ToolCall{FC:nil}, TextContent{"keep me"}] (two leading invalid tool calls, then real text):

• Expected after pruning: ["keep me"]
• Actual: [ToolCall{FC:nil}] — the legitimate text is dropped and an invalid nil-FunctionCall tool call survives into the message chain, feeding malformed messages to downstream summarization / LLM code.

(The scan continues over nil-FC tool calls and only breaks on a valid tool call, so 2+ leading nil-FC tool calls are collected — the trigger.)

Fix

Delete the collected indices in reverse, so removing a later element never invalidates the indices still pending:

for i := len(partsToDelete) - 1; i >= 0; i-- {
    id := partsToDelete[i]
    aiMsg.Parts = append(aiMsg.Parts[:id], aiMsg.Parts[id+1:]...)
}

Tests

Added TestNewBodyPair_DropsMultipleNilFunctionCallToolCalls in backend/pkg/cast/newbodypair_reg_test.go.

• Before the fix it fails (surviving part is a ToolCall, text lost).
• After the fix it passes.

go test ./pkg/cast/ ./pkg/csum/ passes (no regressions) and go vet ./pkg/cast/ is clean.