fix(context): preserve chronological turn order through pack assembly
Tool-calling loops were rendered as all-assistant-calls-then-all-tool-results with the original task last, causing the model to re-issue the same tool call indefinitely. Two stacked reorderings caused it: - DefaultContextPackBuilder grouped entries by sourceType for per-type compression, discarding a,t,a,t interleaving. - PromptRenderer forced L1 (the live user turn) to render last, pushing the task after the entire transcript. - SessionOrchestrator's tool loop re-fed currentContext.layers.values.flatten() (grouped by layer) each round, compounding the scramble. Add a chronological `ordinal` to ContextEntry, stamped by the builder from input order and restored after grouping/compression (the compressor preserves entry identity, so ordinals survive). PromptRenderer now orders non-system messages by ordinal, with the old L1-last layer priority kept only as a tiebreak so router chat (ordinal 0) is unchanged. The orchestrator keeps a running ordered accumulator instead of reading back the grouped pack. Adds builder + renderer ordering regression tests.
This commit is contained in:
@@ -1,5 +1,6 @@
|
||||
package com.correx.core.inference
|
||||
|
||||
import com.correx.core.context.model.ContextEntry
|
||||
import com.correx.core.context.model.ContextLayer
|
||||
import com.correx.core.context.model.ContextPack
|
||||
import com.correx.core.context.model.EntryRole
|
||||
@@ -12,12 +13,13 @@ data class ChatMessage(
|
||||
)
|
||||
|
||||
object PromptRenderer {
|
||||
// L1 (live conversation, ends with the current user turn) must render last so
|
||||
// the model template sees the user query as the final message. Background/memory
|
||||
// layers (L2, L3, L4) are injected before it.
|
||||
private val nonL0RenderOrder: Comparator<ContextLayer> = Comparator { a, b ->
|
||||
val priority = { layer: ContextLayer -> if (layer == ContextLayer.L1) Int.MAX_VALUE else layer.ordinal }
|
||||
priority(a) - priority(b)
|
||||
// Tiebreak only: when entries carry no chronological ordinal (all 0 — e.g. router
|
||||
// chat, which assembles its pack directly), fall back to the old layer priority that
|
||||
// renders L1 (the live user turn) last so the template sees a user query at the end.
|
||||
// When ordinals are present (orchestrator stage path), they dominate and produce true
|
||||
// chronological order — a tool loop reads task → assistant → tool → assistant → …
|
||||
private val layerPriority: (ContextLayer) -> Int = { layer ->
|
||||
if (layer == ContextLayer.L1) Int.MAX_VALUE else layer.ordinal
|
||||
}
|
||||
|
||||
fun render(contextPack: ContextPack): List<ChatMessage> {
|
||||
@@ -29,24 +31,23 @@ object PromptRenderer {
|
||||
.takeIf { it.isNotBlank() }
|
||||
val conversationMessages = contextPack.layers.entries
|
||||
.filter { it.key != ContextLayer.L0 }
|
||||
.sortedWith(Comparator.comparing({ it.key }, nonL0RenderOrder))
|
||||
.flatMap { (_, entries) ->
|
||||
entries.map { entry ->
|
||||
ChatMessage(
|
||||
role = when (entry.role) {
|
||||
EntryRole.SYSTEM -> "system"
|
||||
EntryRole.ASSISTANT -> "assistant"
|
||||
EntryRole.TOOL -> "tool"
|
||||
EntryRole.USER -> "user"
|
||||
},
|
||||
content = entry.content,
|
||||
)
|
||||
}
|
||||
}
|
||||
.flatMap { (layer, entries) -> entries.map { layer to it } }
|
||||
.sortedWith(compareBy({ it.second.ordinal }, { layerPriority(it.first) }))
|
||||
.map { (_, entry) -> entry.toChatMessage() }
|
||||
val messages = buildList {
|
||||
systemContent?.let { add(ChatMessage("system", it)) }
|
||||
addAll(conversationMessages)
|
||||
}
|
||||
return messages.ifEmpty { listOf(ChatMessage("user", "")) }
|
||||
}
|
||||
|
||||
private fun ContextEntry.toChatMessage(): ChatMessage = ChatMessage(
|
||||
role = when (role) {
|
||||
EntryRole.SYSTEM -> "system"
|
||||
EntryRole.ASSISTANT -> "assistant"
|
||||
EntryRole.TOOL -> "tool"
|
||||
EntryRole.USER -> "user"
|
||||
},
|
||||
content = content,
|
||||
)
|
||||
}
|
||||
|
||||
Reference in New Issue
Block a user