fix(kernel,talkie,context): three instruction-corruption fixes from the 2026-08-26 context audit

Each of the three lost or rewrote an instruction before the model saw it.

1. Orphan corrective nudges. pushBack() and the final tools-disabled emission
   built their nudge as a toolResult with a fresh sourceId, so it had no matching
   assistantToolCall and reconcileToolPairs() deleted it — the orchestrator
   believed it had corrected the model while the correction never reached the
   prompt. Affected the invalid-emit_artifact, premature-stage_complete,
   missing-write, read-loop, rejection-loop and final-JSON nudges. They are now
   USER turns (sourceType orchestratorCorrection, REQUIRED bucket, STRUCTURED in
   ContextClassifier so pruning cannot shred them), appended last so the builder's
   positional ordinal puts them at the end of the transcript. A superseded nudge
   is dropped rather than stacking stale demands.

2. Steering laundered through the router. The SteeringNoteAddedEvent carried the
   router model's paraphrase of the operator's message, not the message —
   negations, filenames, constraints and priority could change before the
   orchestrator saw them. It now carries the raw input; the router turn is still
   produced and shown as conversational acknowledgement, it is just not the
   mandate. Resolves the ponytail: note at TalkieFacade.kt:220.

3. Journal compaction erased its own history. compactIfNeeded() summarized only
   state.records while the reducer overwrote summaryArtifactId and dropped covered
   records, so the second compaction lost everything the first had preserved — and
   a low-salience-only batch replaced it with the "(no high-salience decisions)"
   fallback. Compaction is cumulative now, and a blank or fallback-only rewrite
   never replaces real history.

Tests: rendered-prompt regression for (1) — verified to fail against the old
toolResult shape — updated steering expectations for (2), two cumulative-compaction
tests for (3). Full build green, 1831 tests, 0 failures.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>

core/context/src/main/kotlin/com/correx/core/context/compression/ContextClassifier.kt

@@ -35,8 +35,13 @@ class ContextClassifier {
         // shredded the JSON to a bare id + protected path, giving the model amnesia about what it
         // had already done → it re-issued the same call and looped. Structured = format-compress
         // ok, never pruned.
+        // "orchestratorCorrection" is an in-loop corrective USER turn authored by the
+        // orchestrator (invalid emit_artifact, premature stage_complete, read loop, missing write).
+        // Pruning it as freeform prose can shred the tool name or the negation that makes it
+        // actionable, which leaves the model with a vague complaint instead of an instruction.
         val STRUCTURED_SOURCES = setOf(
             "toolLog", "artifact", "config", "structured", "steeringNote", "assistantToolCall",
+            "orchestratorCorrection",
         )
     }
 }

core/kernel/src/main/kotlin/com/correx/core/kernel/orchestration/JournalCompactionService.kt

@@ -30,16 +30,38 @@ class JournalCompactionService(
         val highRecords = state.records.filter { it.kind.salience() == Salience.HIGH }
         val lowCount = state.records.count { it.kind.salience() == Salience.LOW }
 
-        val summaryText = if (highRecords.isEmpty()) {
+        // Compaction is CUMULATIVE. The reducer overwrites summaryArtifactId and drops every
+        // covered record, so a summary built from `state.records` alone erases the previous
+        // summary on the second compaction: the intent, approvals and steering it carried are
+        // neither an input here nor retained as a predecessor. Feed the prior summary back in.
+        val priorSummary = state.summaryArtifactId
+            ?.let { artifactStore.get(it) }
+            ?.toString(Charsets.UTF_8)
+            ?.takeIf { it.isNotBlank() }
+
+        val summaryText = if (highRecords.isEmpty() && priorSummary == null) {
             "(no high-salience decisions to summarize)"
+        } else if (highRecords.isEmpty()) {
+            // Nothing new worth keeping — carry the prior summary forward untouched rather than
+            // replacing it with the fallback text.
+            priorSummary!!
         } else {
             val prompt = buildString {
                 appendLine("Summarize the following key decisions concisely (≤200 words).")
                 appendLine("Preserve all user intent, approvals, steering, and failures.")
                 appendLine()
+                priorSummary?.let {
+                    appendLine("Summary of earlier decisions (already compacted — preserve its content):")
+                    appendLine(it)
+                    appendLine()
+                    appendLine("New decisions since then:")
+                }
                 highRecords.forEach { appendLine("- [${it.kind}] ${it.summary}") }
             }
-            summarize(prompt)
+            // A blank or fallback-only rewrite must never replace real history.
+            summarize(prompt).takeIf { it.isNotBlank() }
+                ?: priorSummary
+                ?: "(no high-salience decisions to summarize)"
         }
 
         val summaryArtifactId = artifactStore.put(summaryText.toByteArray(Charsets.UTF_8))

core/kernel/src/main/kotlin/com/correx/core/kernel/orchestration/SessionOrchestratorExecution.kt

@@ -2,6 +2,7 @@ package com.correx.core.kernel.orchestration
 
 import com.correx.core.context.builder.RequiredContextOverflowException
 import com.correx.core.context.model.ContextEntry
+import com.correx.core.context.model.ContextBucket
 import com.correx.core.context.model.ContextLayer
 import com.correx.core.context.model.TokenBudget
 import com.correx.core.context.model.EntryRole
@@ -351,18 +352,19 @@ internal suspend fun SessionOrchestrator.executeStage(
         return completedIds.isEmpty()
     }
 
-    // Append a corrective tool-result and re-run inference (bounded by MAX_TOOL_ROUNDS).
+    // Append a corrective USER turn and re-run inference (bounded by MAX_TOOL_ROUNDS).
     // Returns the new result rather than mutating inferenceResult, to preserve smart casts.
+    //
+    // NOT a toolResult: these nudges are orchestrator-authored, so they have no matching
+    // assistantToolCall, and reconcileToolPairs() drops every tool result whose call ID is absent
+    // (see its doc). The orchestrator believed it had corrected the model while the correction
+    // never reached the prompt. A USER turn is what this actually is — the operator side of the
+    // loop telling the model what to do next — and it survives reconciliation. Appended last so
+    // the builder's positional ordinal stamp puts it at the end of the transcript, and only one
+    // correction is ever live: a superseded nudge is dropped rather than stacking stale demands.
     suspend fun pushBack(nudge: String, forceWriteOnly: Boolean = false): InferenceResult {
-        accumulatedEntries = accumulatedEntries + ContextEntry(
-            id = ContextEntryId(UUID.randomUUID().toString()),
-            layer = ContextLayer.L2,
-            sourceType = "toolResult",
-            sourceId = UUID.randomUUID().toString(),
-            content = nudge,
-            tokenEstimate = estimateTokens(nudge),
-            role = EntryRole.TOOL,
-        )
+        accumulatedEntries = accumulatedEntries.filterNot { it.sourceType == CORRECTION_SOURCE_TYPE } +
+            correctionEntry(nudge)
         currentContext = contextPackBuilder.build(
             id = ContextPackId(UUID.randomUUID().toString()),
             sessionId = sessionId,
@@ -582,15 +584,8 @@ internal suspend fun SessionOrchestrator.executeStage(
         if (needsCleanEmission && !isCancelled(sessionId)) {
             val nudge = "Stop calling tools. Output the required '${llmEmittedSlots.first().name.value}' " +
                 "artifact now as a single JSON object matching the schema — no tool calls, no commentary."
-            accumulatedEntries = accumulatedEntries + ContextEntry(
-                id = ContextEntryId(UUID.randomUUID().toString()),
-                layer = ContextLayer.L2,
-                sourceType = "toolResult",
-                sourceId = UUID.randomUUID().toString(),
-                content = nudge,
-                tokenEstimate = estimateTokens(nudge),
-                role = EntryRole.TOOL,
-            )
+            accumulatedEntries = accumulatedEntries.filterNot { it.sourceType == CORRECTION_SOURCE_TYPE } +
+                correctionEntry(nudge)
             currentContext = contextPackBuilder.build(
                 id = ContextPackId(UUID.randomUUID().toString()),
                 sessionId = sessionId,
@@ -687,3 +682,19 @@ internal suspend fun SessionOrchestrator.executeStage(
         }
     }
 }
+
+// Orchestrator-authored corrections. STRUCTURED in ContextClassifier (never token-pruned) and
+// REQUIRED so a correction is never traded away for budget — the whole point of a nudge is that
+// the next inference sees it.
+internal const val CORRECTION_SOURCE_TYPE = "orchestratorCorrection"
+
+private suspend fun SessionOrchestrator.correctionEntry(nudge: String) = ContextEntry(
+    id = ContextEntryId(UUID.randomUUID().toString()),
+    layer = ContextLayer.L2,
+    sourceType = CORRECTION_SOURCE_TYPE,
+    sourceId = UUID.randomUUID().toString(),
+    content = nudge,
+    tokenEstimate = estimateTokens(nudge),
+    role = EntryRole.USER,
+    bucket = ContextBucket.REQUIRED,
+)

core/kernel/src/test/kotlin/com/correx/core/kernel/orchestration/JournalCompactionServiceTest.kt

@@ -33,6 +33,65 @@ class JournalCompactionServiceTest {
     private fun makeRecord(seq: Long, kind: DecisionKind) =
         DecisionRecord(sequence = seq, kind = kind, summary = "summary of $kind")
 
+    // A store that really round-trips, so a second compaction can read back the first summary.
+    private fun recordingArtifactStore(): Pair<ArtifactStore, MutableMap<TypeId, ByteArray>> {
+        val blobs = mutableMapOf<TypeId, ByteArray>()
+        val store = object : ArtifactStore {
+            override suspend fun put(bytes: ByteArray): TypeId {
+                val id = TypeId("artifact-${blobs.size + 1}")
+                blobs[id] = bytes
+                return id
+            }
+            override suspend fun get(id: TypeId): ByteArray? = blobs[id]
+            override suspend fun flushBefore(commit: suspend () -> Unit) = commit()
+        }
+        return store to blobs
+    }
+
+    @Test
+    fun `second compaction feeds the prior summary back into the prompt`(): Unit = runBlocking {
+        val (store, blobs) = recordingArtifactStore()
+        val prompts = mutableListOf<String>()
+        val svc = JournalCompactionService(store, { prompts += it; "SUMMARY-${prompts.size}" }, { 100 })
+
+        val first = stateWithRecords(makeRecord(1, DecisionKind.INTENT))
+        val emitted = mutableListOf<EventPayload>()
+        assertTrue(svc.compactIfNeeded(SessionId("s1"), first, 500) { emitted += it })
+        val firstId = (emitted.single() as JournalCompactedEvent).summaryArtifactId
+
+        // Reducer semantics: covered records are gone, summaryArtifactId now points at SUMMARY-1.
+        val second = DecisionJournalState(
+            records = listOf(makeRecord(2, DecisionKind.INTENT)),
+            compactedThroughSequence = 1,
+            summaryArtifactId = firstId,
+        )
+        emitted.clear()
+        assertTrue(svc.compactIfNeeded(SessionId("s1"), second, 500) { emitted += it })
+
+        assertTrue(prompts[1].contains("SUMMARY-1"), "prior summary must be an input: ${prompts[1]}")
+        val kept = blobs[(emitted.single() as JournalCompactedEvent).summaryArtifactId]!!
+        assertEquals("SUMMARY-2", kept.toString(Charsets.UTF_8))
+    }
+
+    @Test
+    fun `a low-salience-only batch carries the prior summary forward instead of erasing it`():
+        Unit = runBlocking {
+        val (store, blobs) = recordingArtifactStore()
+        val priorId = store.put("EARLIER DECISIONS".toByteArray(Charsets.UTF_8))
+        val svc = JournalCompactionService(store, { "should not be called" }, { 100 })
+
+        val state = DecisionJournalState(
+            records = listOf(makeRecord(2, DecisionKind.TRANSITION)),
+            compactedThroughSequence = 1,
+            summaryArtifactId = priorId,
+        )
+        val emitted = mutableListOf<EventPayload>()
+        assertTrue(svc.compactIfNeeded(SessionId("s1"), state, 500) { emitted += it })
+
+        val kept = blobs[(emitted.single() as JournalCompactedEvent).summaryArtifactId]!!
+        assertEquals("EARLIER DECISIONS", kept.toString(Charsets.UTF_8))
+    }
+
     @Test
     fun `returns false when token estimate is below threshold`(): Unit = runBlocking {
         val svc = JournalCompactionService(fakeArtifactStore(), { it }, tokenThreshold = { 2000 })

core/talkie/src/main/kotlin/com/correx/core/talkie/TalkieFacade.kt

@@ -217,16 +217,16 @@ class DefaultTalkieFacade(
             emitIdeasCaptured(sessionId, ideas.ideas)
         }
 
-        // ponytail: STEERING launders the user's text through the router LLM (the `content` above)
-        // before injecting it as a note. For a clear instruction that's an extra inference that can
-        // distort intent; the reformulation only earns its cost when the input is terse/context-
-        // dependent. Upgrade path: inject the raw (validated) input directly and skip the rewrite
-        // unless a heuristic flags the message as too short/ambiguous to stand alone.
-        val steeringEmitted = mode == ChatMode.STEERING && rawContent.isNotBlank()
+        // The steering note carries the operator's OWN text, not the router's paraphrase of it.
+        // Routing it through inference first let a lossy rewrite acquire the authority of a user
+        // directive — negations, filenames, constraints and priority could all change before the
+        // orchestrator saw them. The router turn above is still produced and shown to the operator
+        // as conversational acknowledgement; it is just not the mandate.
+        val steeringEmitted = mode == ChatMode.STEERING && input.isNotBlank()
         if (steeringEmitted) {
-            val validationError = validateSteering?.invoke(content)
+            val validationError = validateSteering?.invoke(input)
             if (validationError == null) {
-                emitSteeringNote(sessionId, content, effectiveStageId)
+                emitSteeringNote(sessionId, input, effectiveStageId)
             }
         }
 

testing/deterministic/src/test/kotlin/DefaultContextPackBuilderTest.kt

@@ -5,8 +5,10 @@ import com.correx.core.context.builder.RequiredContextOverflowException
 import com.correx.core.context.model.ContextBucket
 import com.correx.core.context.model.ContextEntry
 import com.correx.core.context.model.ContextLayer
+import com.correx.core.context.model.EntryRole
 import com.correx.core.context.model.TokenBudget
 import com.correx.core.events.types.ContextEntryId
+import com.correx.core.inference.PromptRenderer
 import com.correx.core.events.types.ContextPackId
 import com.correx.core.events.types.SessionId
 import com.correx.core.events.types.StageId
@@ -42,6 +44,55 @@ class DefaultContextPackBuilderTest {
         tokenEstimate = tokens
     )
 
+    @Test
+    fun `an orchestrator correction reaches the rendered prompt as the last user turn`() =
+        kotlinx.coroutines.runBlocking {
+            // The nudge used to be built as a toolResult with a fresh sourceId, so it had no
+            // matching assistantToolCall and reconcileToolPairs() deleted it: the orchestrator
+            // "corrected" the model and the correction never reached the provider. Assert on the
+            // RENDERED prompt, not the entry list — that is where the defect was invisible.
+            val entries = listOf(
+                typedEntry("sys", ContextLayer.L0, 100, "systemPrompt"),
+                typedEntry("task", ContextLayer.L1, 50, "agentPrompt"),
+                ContextEntry(
+                    id = ContextEntryId("call1"),
+                    layer = ContextLayer.L2,
+                    content = "{\"tool\":\"file_read\"}",
+                    sourceType = "assistantToolCall",
+                    sourceId = "inv-1",
+                    tokenEstimate = 40,
+                    role = EntryRole.ASSISTANT,
+                ),
+                ContextEntry(
+                    id = ContextEntryId("res1"),
+                    layer = ContextLayer.L2,
+                    content = "file contents",
+                    sourceType = "toolResult",
+                    sourceId = "inv-1",
+                    tokenEstimate = 40,
+                    role = EntryRole.TOOL,
+                ),
+                ContextEntry(
+                    id = ContextEntryId("nudge"),
+                    layer = ContextLayer.L2,
+                    content = "STOP reading. You MUST now call file_write.",
+                    sourceType = "orchestratorCorrection",
+                    sourceId = "corr-1",
+                    tokenEstimate = 20,
+                    role = EntryRole.USER,
+                    bucket = ContextBucket.REQUIRED,
+                ),
+            )
+            val pack = builder.build(packId, sessionId, stageId, entries, TokenBudget(limit = 4096))
+            val messages = PromptRenderer.render(pack)
+            val last = messages.last()
+            assertEquals("user", last.role)
+            assertTrue(
+                last.content.contains("STOP reading"),
+                "correction must survive to the prompt; got: " + messages.map { it.role to it.content },
+            )
+        }
+
     @Test
     fun `oversized tool results from a stage tool loop are trimmed to budget`() = kotlinx.coroutines.runBlocking {
         // Live repro (2026-06-11): analyst stage with three file_read results of ~5.6k/11.9k/10.7k

testing/deterministic/src/test/kotlin/TalkieFacadeTest.kt

@@ -129,7 +129,8 @@ class TalkieFacadeTest {
         assertEquals(com.correx.core.events.events.ChatTurnRole.USER, chatEvents[0].role)
         assertEquals("inference response", chatEvents[1].content)
         assertEquals(com.correx.core.events.events.ChatTurnRole.ROUTER, chatEvents[1].role)
-        assertEquals("inference response", steeringEvents[0].content)
+        // The steering note is the operator's raw text, never the router's paraphrase of it.
+        assertEquals("steer this way", steeringEvents[0].content)
     }
 
     @Test
@@ -158,7 +159,7 @@ class TalkieFacadeTest {
         val payloads = mockStore.appendedEvents.map { it.payload }
         val steeringEvents = payloads.filterIsInstance<com.correx.core.events.events.SteeringNoteAddedEvent>()
         assertEquals(1, steeringEvents.size)
-        assertEquals("steering response", steeringEvents[0].content)
+        assertEquals("Hello!", steeringEvents[0].content)
     }
 
     // --------------------------------------------------------------------------