Merge pull request #1106 from Hy4ri/fix/issue-1103-context-bar-model-switch

fix(chat): keep context meter aligned after model swaps (#1103)

app/src/main/java/com/m57/hermescontrol/data/model/ContextBreakdownResponse.kt

@@ -1,5 +1,11 @@
 package com.m57.hermescontrol.data.model
 
+import com.m57.hermescontrol.data.ws.toAny
+import kotlinx.serialization.json.JsonElement
+import kotlinx.serialization.json.JsonPrimitive
+import kotlinx.serialization.json.intOrNull
+import kotlinx.serialization.json.longOrNull
+
 /**
  * Live context-window occupancy from the `session.context_breakdown` WS RPC —
  * the same RPC the Hermes desktop app's status-bar meter uses.
@@ -30,20 +36,47 @@ data class ContextBreakdownResponse(
  * Parse the decoded JSON-RPC result of `session.context_breakdown`.
  *
  * The WS event parser decodes JSON numbers as [Double] (`JsonRpcModels.toAny`),
- * so every count is read through [Number]. Returns null when the payload is
+ * so every count is read through [Number]. Also supports raw [JsonObject] from
+ * [HermesWsClient.request]. Returns null when the payload is
  * not a map (error/malformed response) — callers treat that as unknown and
  * keep the last known meter values.
  */
 fun parseContextBreakdown(result: Any?): ContextBreakdownResponse? {
-    val map = result as? Map<*, *> ?: return null
+    val rawMap =
+        when (result) {
+            is JsonElement -> result.toAny() as? Map<*, *>
+            is Map<*, *> -> result
+            else -> null
+        } ?: return null
 
-    fun num(key: String): Long? = (map[key] as? Number)?.toLong()
+    fun num(key: String): Long? =
+        when (val value = rawMap[key]) {
+            is Number -> value.toLong()
+            is JsonPrimitive -> value.longOrNull ?: value.intOrNull?.toLong()
+            is String -> value.toLongOrNull()
+            else -> null
+        }
+
+    fun intNum(key: String): Int? =
+        when (val value = rawMap[key]) {
+            is Number -> value.toInt()
+            is JsonPrimitive -> value.intOrNull
+            is String -> value.toIntOrNull()
+            else -> null
+        }
+
+    fun str(key: String): String? =
+        when (val value = rawMap[key]) {
+            is String -> value
+            is JsonPrimitive -> value.content
+            else -> value?.toString()
+        }
 
     return ContextBreakdownResponse(
         contextUsed = num("context_used"),
         contextMax = num("context_max"),
-        contextPercent = (map["context_percent"] as? Number)?.toInt(),
+        contextPercent = intNum("context_percent"),
         estimatedTotal = num("estimated_total"),
-        model = map["model"] as? String,
+        model = str("model"),
     )
 }

app/src/main/java/com/m57/hermescontrol/data/model/UsageSnapshotResponse.kt

@@ -1,5 +1,12 @@
 package com.m57.hermescontrol.data.model
 
+import com.m57.hermescontrol.data.ws.toAny
+import kotlinx.serialization.json.JsonElement
+import kotlinx.serialization.json.JsonPrimitive
+import kotlinx.serialization.json.doubleOrNull
+import kotlinx.serialization.json.intOrNull
+import kotlinx.serialization.json.longOrNull
+
 /**
  * Minimal view of the `session.usage` WS RPC result or `session.usage` push event —
  * only the fields the context widget consumes today. The backend sends many more keys
@@ -29,17 +36,47 @@ data class UsageSnapshotResponse(
  *
  * Handles both direct usage dict (from RPC response) and nested `{ "usage": { ... } }`
  * (from push event payload). The WS event parser decodes JSON numbers as [Double]
- * (`JsonRpcModels.toAny`), so every count is read through [Number].
+ * (`JsonRpcModels.toAny`), so every count is read through [Number]. Also supports raw [JsonObject].
  * Returns null when the payload is not a map (error/malformed response) — callers keep the last known value.
  */
 fun parseUsageSnapshot(result: Any?): UsageSnapshotResponse? {
-    val map = result as? Map<*, *> ?: return null
-    val usageMap = (map["usage"] as? Map<*, *>) ?: map
+    val rawMap =
+        when (result) {
+            is JsonElement -> result.toAny() as? Map<*, *>
+            is Map<*, *> -> result
+            else -> null
+        } ?: return null
+    val usageMap = (rawMap["usage"] as? Map<*, *>) ?: rawMap
+
+    fun num(key: String): Long? =
+        when (val value = usageMap[key]) {
+            is Number -> value.toLong()
+            is JsonPrimitive -> value.longOrNull ?: value.intOrNull?.toLong()
+            is String -> value.toLongOrNull()
+            else -> null
+        }
+
+    fun intNum(key: String): Int? =
+        when (val value = usageMap[key]) {
+            is Number -> value.toInt()
+            is JsonPrimitive -> value.intOrNull
+            is String -> value.toIntOrNull()
+            else -> null
+        }
+
+    fun doubleNum(key: String): Double? =
+        when (val value = usageMap[key]) {
+            is Number -> value.toDouble()
+            is JsonPrimitive -> value.doubleOrNull ?: value.longOrNull?.toDouble()
+            is String -> value.toDoubleOrNull()
+            else -> null
+        }
+
     return UsageSnapshotResponse(
-        compressions = (usageMap["compressions"] as? Number)?.toInt(),
-        contextUsed = (usageMap["context_used"] as? Number)?.toLong(),
-        contextMax = (usageMap["context_max"] as? Number)?.toLong(),
-        totalTokens = (usageMap["total"] as? Number)?.toLong(),
-        avgTps = (usageMap["avg_tps"] as? Number)?.toDouble(),
+        compressions = intNum("compressions"),
+        contextUsed = num("context_used"),
+        contextMax = num("context_max"),
+        totalTokens = num("total"),
+        avgTps = doubleNum("avg_tps"),
     )
 }

app/src/main/java/com/m57/hermescontrol/ui/chat/ChatModelSwitchDelegate.kt

@@ -22,6 +22,7 @@ import java.util.concurrent.ConcurrentHashMap
 private data class ActiveModelSwitch(
     val spec: String,
     val previousModel: String?,
+    val sequence: Long,
 )
 
 private data class ActiveFastSwitch(
@@ -44,6 +45,7 @@ class ChatModelSwitchDelegate(
     private val addAssistantMessage: (text: String) -> Unit,
     private val handleSlashCommand: (command: String) -> Unit,
     private val fetchContextUsage: () -> Unit,
+    private val onModelSwitchInitiated: () -> Unit = {},
     private val getModelOptionsCall: suspend (
         refresh: Boolean,
     ) -> NetworkResult<com.m57.hermescontrol.data.model.ModelOptionsResponse> =
@@ -56,6 +58,9 @@ class ChatModelSwitchDelegate(
     private val fastRejectedModels = ConcurrentHashMap.newKeySet<String>()
     private var activeModelSwitchConfirmation: ActiveModelSwitch? = null
     private var optimisticPreviousModel: String? = null
+    private var lastConfirmedModel: String? = null
+    private var unconfirmedTargetModel: String? = null
+    private var modelSwitchSequence: Long = 0L
     private var cachedModelOptions: List<ModelProvider> = emptyList()
 
     fun isModelPickerCommand(command: String): Boolean {
@@ -142,20 +147,37 @@ class ChatModelSwitchDelegate(
         uiState.update { it.copy(modelPickerPinned = currentPinned) }
     }
 
+    fun consumeOptimisticPreviousModel(): String? = optimisticPreviousModel.also { optimisticPreviousModel = null }
+
     fun sendSlashModel(
         provider: String,
         model: String,
     ) {
-        optimisticPreviousModel = uiState.value.currentSessionModel
+        val targetModelLabel = "$provider/$model"
+        val isDifferent = !targetModelLabel.equals(uiState.value.currentSessionModel, ignoreCase = true)
+        if (!isDifferent) {
+            uiState.update {
+                it.copy(
+                    showModelPicker = false,
+                    modelPickerLoading = false,
+                )
+            }
+            return
+        }
+        if (optimisticPreviousModel == null) {
+            optimisticPreviousModel = uiState.value.currentSessionModel
+        }
+        unconfirmedTargetModel = targetModelLabel
         uiState.update {
             it.copy(
                 showModelPicker = false,
                 modelPickerLoading = false,
-                currentSessionModel = "$provider/$model",
+                currentSessionModel = targetModelLabel,
+                fullContextTokens = null,
             )
         }
+        onModelSwitchInitiated()
         syncCurrentModelCapabilities()
-        fetchContextUsage()
         handleSlashCommand("/model $model --provider $provider --session")
     }
 
@@ -174,8 +196,15 @@ class ChatModelSwitchDelegate(
             } else {
                 command.trim()
             }
-        val previousModel = optimisticPreviousModel ?: uiState.value.currentSessionModel
-        optimisticPreviousModel = null
+        val previousModel = lastConfirmedModel ?: optimisticPreviousModel ?: uiState.value.currentSessionModel
+        // Retain optimisticPreviousModel for consumeOptimisticPreviousModel() when SessionInfo arrives.
+        if (optimisticPreviousModel == null && !spec.equals(uiState.value.currentSessionModel, ignoreCase = true)) {
+            optimisticPreviousModel = uiState.value.currentSessionModel
+            uiState.update { it.copy(fullContextTokens = null) }
+            onModelSwitchInitiated()
+        }
+        unconfirmedTargetModel = spec
+        val switchSeq = ++modelSwitchSequence
         val params =
             mutableMapOf<String, Any>(
                 "key" to "model",
@@ -191,7 +220,7 @@ class ChatModelSwitchDelegate(
                 params,
             ) { id ->
                 trackRequest(id, WsMethods.CONFIG_SET)
-                pendingModelSwitchRequests[id] = ActiveModelSwitch(spec, previousModel)
+                pendingModelSwitchRequests[id] = ActiveModelSwitch(spec, previousModel, switchSeq)
             }
         }
     }
@@ -207,13 +236,15 @@ class ChatModelSwitchDelegate(
         if (key == "model") {
             val confirmRequired = map["confirm_required"] as? Boolean ?: false
             if (confirmRequired) {
-                val confirmMessage =
-                    (map["confirm_message"] as? String)
-                        ?: (map["warning"] as? String)
-                        ?: "This model requires confirmation to switch. Continue?"
-                activeModelSwitchConfirmation = pending
-                uiState.update {
-                    it.copy(modelSwitchConfirmMessage = confirmMessage)
+                if (pending != null && pending.sequence == modelSwitchSequence) {
+                    val confirmMessage =
+                        (map["confirm_message"] as? String)
+                            ?: (map["warning"] as? String)
+                            ?: "This model requires confirmation to switch. Continue?"
+                    activeModelSwitchConfirmation = pending
+                    uiState.update {
+                        it.copy(modelSwitchConfirmMessage = confirmMessage)
+                    }
                 }
             }
         } else if (key == "fast") {
@@ -236,13 +267,18 @@ class ChatModelSwitchDelegate(
     fun dismissModelSwitchConfirm() {
         val pending = activeModelSwitchConfirmation
         activeModelSwitchConfirmation = null
+        optimisticPreviousModel = null
+        unconfirmedTargetModel = null
+        val revertModel = lastConfirmedModel ?: pending?.previousModel
         uiState.update {
             it.copy(
                 modelSwitchConfirmMessage = null,
-                currentSessionModel = pending?.previousModel ?: it.currentSessionModel,
+                currentSessionModel = revertModel ?: it.currentSessionModel,
+                fullContextTokens = null,
             )
         }
         syncCurrentModelCapabilities()
+        onModelSwitchInitiated()
         fetchContextUsage()
     }
 
@@ -282,6 +318,33 @@ class ChatModelSwitchDelegate(
         id: String,
         error: Any?,
     ) {
+        val pendingModel = pendingModelSwitchRequests.remove(id)
+        if (pendingModel != null && pendingModel.sequence == modelSwitchSequence) {
+            val revertModel = lastConfirmedModel ?: pendingModel.previousModel
+            optimisticPreviousModel = null
+            unconfirmedTargetModel = null
+            activeModelSwitchConfirmation = null
+            if (revertModel != null) {
+                uiState.update {
+                    it.copy(
+                        currentSessionModel = revertModel,
+                        fullContextTokens = null,
+                        modelSwitchConfirmMessage = null,
+                    )
+                }
+                syncCurrentModelCapabilities()
+                onModelSwitchInitiated()
+                fetchContextUsage()
+            } else {
+                uiState.update {
+                    it.copy(
+                        fullContextTokens = null,
+                        modelSwitchConfirmMessage = null,
+                    )
+                }
+                onModelSwitchInitiated()
+            }
+        }
         val pendingFast = pendingFastSwitchRequests.remove(id)
         if (pendingFast != null) {
             val errorMsg =
@@ -300,6 +363,26 @@ class ChatModelSwitchDelegate(
         }
     }
 
+    fun isSwitchPending(): Boolean =
+        optimisticPreviousModel != null ||
+            unconfirmedTargetModel != null ||
+            activeModelSwitchConfirmation != null
+
+    fun onModelConfirmed(modelLabel: String) {
+        lastConfirmedModel = modelLabel
+        optimisticPreviousModel = null
+        unconfirmedTargetModel = null
+    }
+
+    fun reset() {
+        pendingModelSwitchRequests.clear()
+        pendingFastSwitchRequests.clear()
+        activeModelSwitchConfirmation = null
+        optimisticPreviousModel = null
+        lastConfirmedModel = null
+        unconfirmedTargetModel = null
+    }
+
     fun setReasoningLevel(level: String?) {
         uiState.update { it.copy(reasoningLevel = level) }
         val sessionId = runtimeSessionId() ?: return

app/src/main/java/com/m57/hermescontrol/ui/chat/ChatViewModel.kt

@@ -28,9 +28,11 @@ import com.m57.hermescontrol.data.ws.WsEvent
 import com.m57.hermescontrol.data.ws.WsMethods
 import com.m57.hermescontrol.data.ws.toJsonElement
 import com.m57.hermescontrol.ui.common.ActionProgressController
+import kotlinx.coroutines.CancellationException
 import kotlinx.coroutines.Dispatchers
 import kotlinx.coroutines.Job
 import kotlinx.coroutines.delay
+import kotlinx.coroutines.ensureActive
 import kotlinx.coroutines.flow.MutableStateFlow
 import kotlinx.coroutines.flow.SharingStarted
 import kotlinx.coroutines.flow.StateFlow
@@ -394,6 +396,18 @@ class ChatViewModel(
     /** Tracks the auto-clear coroutine for reaction animations. */
     private var reactionClearJob: Job? = null
 
+    /** Last confirmed model label pushed by SessionInfo or SessionResume from backend (issue #1103). */
+    private var lastConfirmedSessionModel: String? = null
+
+    /** Model generation counter to prevent stale context writes across switches (issue #1103). */
+    private var modelGeneration: Long = 0L
+
+    /** Request sequence counter for fetchContextUsage calls. */
+    private var contextFetchSequence: Long = 0L
+
+    /** Tracks the latest context-usage poll or refetch job. */
+    private var contextUsageJob: Job? = null
+
     private val wsClient = HermesWsClient
 
     // ── Session persistence ──────────────────────────────────────────────
@@ -445,6 +459,7 @@ class ChatViewModel(
             addAssistantMessage = { text -> addAssistantMessage(text) },
             handleSlashCommand = { cmd -> handleSlashCommand(cmd) },
             fetchContextUsage = { fetchContextUsage() },
+            onModelSwitchInitiated = { onModelSwitchInitiated() },
         )
 
     private val credentialPromptsDelegate =
@@ -796,16 +811,25 @@ class ChatViewModel(
                         } else {
                             model
                         }
-                    // Issue #817: on a REAL model swap the meter's denominator
+                    // Issue #817 & #1103: on a REAL model swap the meter's denominator
                     // still belongs to the old model until the next fetch.
-                    // Blank it so the chip hides instead of flashing a stale
-                    // window under the new label, and re-fire the fetch so the
-                    // new window lands immediately (no 30s poll wait). Only a
-                    // change from a known label counts — the first SessionInfo
-                    // of a session just sets the label.
-                    val previousLabel = _uiState.value.currentSessionModel
+                    // Check against both lastConfirmedSessionModel and optimisticPreviousModel
+                    // so optimistic model updates in sendSlashModel don't defeat swap detection.
+                    val optimisticPrevious = modelSwitchDelegate.consumeOptimisticPreviousModel()
+                    val previousModel =
+                        lastConfirmedSessionModel
+                            ?: optimisticPrevious
+                            ?: _uiState.value.currentSessionModel
                     val modelSwapped =
-                        previousLabel != null && newModelLabel != null && newModelLabel != previousLabel
+                        previousModel != null &&
+                            newModelLabel != null &&
+                            !newModelLabel.equals(previousModel, ignoreCase = true)
+                    val initialHydration = previousModel == null && newModelLabel != null
+                    val meterEmpty = _uiState.value.fullContextTokens == null
+                    lastConfirmedSessionModel = newModelLabel ?: lastConfirmedSessionModel
+                    if (newModelLabel != null) {
+                        modelSwitchDelegate.onModelConfirmed(newModelLabel)
+                    }
                     _uiState.update { state ->
                         state.copy(
                             currentSessionModel = newModelLabel ?: state.currentSessionModel,
@@ -823,12 +847,19 @@ class ChatViewModel(
                     }
                     modelSwitchDelegate.syncCurrentModelCapabilities()
                     if (modelSwapped) {
-                        // Issue #817: after a swap the REST model/info window is
+                        modelGeneration++
+                        contextUsageJob?.cancel()
+                        contextUsageJob = null
+                        // Issue #817 & #1103: after a swap the REST model/info window is
                         // PROFILE-scoped and may describe the old model (e.g. a
                         // session-scoped swap) — the meter must not fall back to
                         // it. Wait for the RPC's live context_max instead; the
                         // chip stays hidden until the real window lands.
                         viewModelScope.launch { fetchContextUsage(skipRestFallback = true) }
+                    } else if ((initialHydration || meterEmpty) && newModelLabel != null &&
+                        contextUsageJob?.isActive != true
+                    ) {
+                        viewModelScope.launch { fetchContextUsage() }
                     }
                     // Session.info can carry `pending_approval` (reconnect
                     // reconciliation) — surface it unless already on screen.
@@ -1151,6 +1182,23 @@ class ChatViewModel(
                         terminalBackend = terminalBackend ?: it.terminalBackend,
                     )
                 }
+                val resumedModelLabel =
+                    if (model != null && provider != null) {
+                        "$provider/$model"
+                    } else {
+                        model
+                    }
+                if (resumedModelLabel != null) {
+                    if (lastConfirmedSessionModel != null &&
+                        !resumedModelLabel.equals(lastConfirmedSessionModel, ignoreCase = true)
+                    ) {
+                        modelGeneration++
+                        contextUsageJob?.cancel()
+                        contextUsageJob = null
+                        _uiState.update { it.copy(fullContextTokens = null) }
+                    }
+                    lastConfirmedSessionModel = resumedModelLabel
+                }
                 modelSwitchDelegate.syncCurrentModelCapabilities()
                 // Mirror the active runtime session id app-wide (issue #532).
                 ActiveSessionHolder.set(runtimeSessionId ?: sessionId, sessionId)
@@ -2558,6 +2606,11 @@ class ChatViewModel(
     ): Long {
         val generation = ++sessionGeneration
         cancelResumeRetry()
+        contextUsageJob?.cancel()
+        contextUsageJob = null
+        modelGeneration++
+        lastConfirmedSessionModel = null
+        modelSwitchDelegate.reset()
         resumedGeneration = -1L
         hydratedGeneration = -1L
         runtimeSessionId = null
@@ -3043,6 +3096,57 @@ class ChatViewModel(
         }
     }
 
+    internal fun onModelSwitchInitiated() {
+        modelGeneration++
+        contextUsageJob?.cancel()
+        contextUsageJob = null
+        _uiState.update { it.copy(fullContextTokens = null) }
+    }
+
+    private fun isCurrentContextFetch(
+        sessionId: String,
+        targetSessionGeneration: Long,
+        targetModelGeneration: Long,
+        targetRequestSequence: Long,
+    ): Boolean =
+        targetSessionGeneration == sessionGeneration &&
+            targetModelGeneration == modelGeneration &&
+            targetRequestSequence == contextFetchSequence &&
+            _uiState.value.currentSessionId == sessionId
+
+    internal fun isMatchingModel(
+        currentModel: String?,
+        info: com.m57.hermescontrol.data.model.ModelInfoResponse?,
+    ): Boolean {
+        if (currentModel.isNullOrBlank() || info == null || info.model.isNullOrBlank()) {
+            return false
+        }
+        val restModel = info.model
+        val restProvider = info.provider
+        return if (currentModel.contains('/')) {
+            val curProvider = currentModel.substringBefore('/')
+            val curModel = currentModel.substringAfter('/')
+            if (!restProvider.isNullOrBlank()) {
+                curProvider.equals(restProvider, ignoreCase = true) &&
+                    curModel.equals(restModel, ignoreCase = true)
+            } else {
+                curModel.equals(restModel, ignoreCase = true)
+            }
+        } else {
+            currentModel.equals(restModel, ignoreCase = true)
+        }
+    }
+
+    internal fun isMatchingRpcModel(
+        currentModel: String?,
+        rpcModel: String?,
+    ): Boolean {
+        if (currentModel.isNullOrBlank() || rpcModel.isNullOrBlank()) return false
+        if (currentModel.equals(rpcModel, ignoreCase = true)) return true
+        return currentModel.contains('/') &&
+            currentModel.substringAfter('/').equals(rpcModel, ignoreCase = true)
+    }
+
     /**
      * Refresh the context meter: used / full tokens for the current session.
      *
@@ -3068,105 +3172,211 @@ class ChatViewModel(
     fun fetchContextUsage(skipRestFallback: Boolean = false) {
         val sessionId = _uiState.value.currentSessionId ?: return
         val profile = AuthManager.activeProfileId.value
-        viewModelScope.launch(ioDispatcher) {
-            // Denominator fallback: full context window (cheap, public, rarely
-            // changes). The RPC's context_max below overrides it when present.
-            // Kept as a local (not a state write) so both sources resolve
-            // before ONE atomic update below (issue #817 — two independent
-            // writes let a stale pre-swap value override a fresh one mid-swap).
-            val fullResult =
-                safeApiCall { ApiClient.hermesApi.getModelInfo() }
-            val restFull =
-                if (fullResult is NetworkResult.Success) {
-                    fullResult.data.effective_context_length
-                        ?: fullResult.data.auto_context_length
-                        ?: fullResult.data.config_context_length
-                } else {
-                    null
+        val targetSessionGeneration = sessionGeneration
+        val targetModelGeneration = modelGeneration
+        val targetRequestSequence = ++contextFetchSequence
+        val isSwitchPending = modelSwitchDelegate.isSwitchPending()
+
+        contextUsageJob?.cancel()
+        contextUsageJob =
+            viewModelScope.launch(ioDispatcher) {
+                // Denominator fallback: full context window (cheap, public, rarely
+                // changes). The RPC's context_max below overrides it when present.
+                // Kept as a local (not a state write) so both sources resolve
+                // before ONE atomic update below (issue #817 — two independent
+                // writes let a stale pre-swap value override a fresh one mid-swap).
+                val fullResult =
+                    safeApiCall { ApiClient.hermesApi.getModelInfo() }
+                coroutineContext.ensureActive()
+                if (!isCurrentContextFetch(
+                        sessionId,
+                        targetSessionGeneration,
+                        targetModelGeneration,
+                        targetRequestSequence,
+                    )
+                ) {
+                    return@launch
                 }
-            // Numerator: live context occupancy from the gateway's live agent,
-            // via the same RPC the desktop meter uses. `context_used` is the
-            // real current prompt size (drops after compression); `context_max`
-            // is the compressor's actual window. Any failure keeps the last
-            // known values — never blank the meter over a transient RPC error.
-            //
-            // These RPCs resolve the session against the gateway's LIVE runtime
-            // registry (_sess_nowait) — the storage session id 4001s "session
-            // not found" until session.resume has registered it. ChatScreen's
-            // sync effect fires this immediately on session switch, before the
-            // resume result lands, so skip the RPCs until resume confirms the
-            // runtime id. The REST parts below stay live (they key on the
-            // storage id).
-            var rpcUsed: Long? = null
-            var rpcMax: Long? = null
-            val rpcSessionId = runtimeSessionId
-            if (rpcSessionId != null) {
-                try {
-                    val result =
-                        sendRpcAndAwait(
-                            WsMethods.SESSION_CONTEXT_BREAKDOWN,
-                            mapOf("session_id" to rpcSessionId),
-                        )
-                    val ctx = parseContextBreakdown(result)
-                    if (ctx != null) {
-                        rpcUsed = ctx.contextUsed?.takeIf { it > 0L }
-                        rpcMax = ctx.contextMax?.takeIf { it > 0L }
+                val restFull =
+                    if (fullResult is NetworkResult.Success && !isSwitchPending) {
+                        val info = fullResult.data
+                        val currentModel = _uiState.value.currentSessionModel
+                        // Issue #1103: only use profile REST model/info as fallback if it
+                        // matches the current session model identity. If the session was switched to
+                        // another model (e.g. Solar, Gemini), the profile-level model/info describes
+                        // a different model and must never poison the session's context window.
+                        if (isMatchingModel(currentModel, info)) {
+                            info.effective_context_length
+                                ?: info.auto_context_length
+                                ?: info.config_context_length
+                        } else {
+                            null
+                        }
+                    } else {
+                        null
                     }
-                } catch (_: Exception) {
-                    // Best-effort: RPC error/timeout/disconnect — keep last values.
-                }
-                // Compression count: how many times this session has been compacted
-                // (session.usage → compressions). Feeds the "compressed ×N" badge —
-                // the same usage snapshot the desktop status bar reads.
-                try {
-                    val usage =
-                        sendRpcAndAwait(
-                            WsMethods.SESSION_USAGE,
-                            mapOf("session_id" to rpcSessionId),
+                // Numerator: live context occupancy from the gateway's live agent,
+                // via the same RPC the desktop meter uses. `context_used` is the
+                // real current prompt size (drops after compression); `context_max`
+                // is the compressor's actual window. Any failure keeps the last
+                // known values — never blank the meter over a transient RPC error.
+                //
+                // These RPCs resolve the session against the gateway's LIVE runtime
+                // registry (_sess_nowait) — the storage session id 4001s "session
+                // not found" until session.resume has registered it. ChatScreen's
+                // sync effect fires this immediately on session switch, before the
+                // resume result lands, so skip the RPCs until resume confirms the
+                // runtime id. The REST parts below stay live (they key on the
+                // storage id).
+                var rpcUsed: Long? = null
+                var rpcMax: Long? = null
+                val rpcSessionId = runtimeSessionId
+                if (rpcSessionId != null) {
+                    try {
+                        val result =
+                            sendRpcAndAwait(
+                                WsMethods.SESSION_CONTEXT_BREAKDOWN,
+                                mapOf("session_id" to rpcSessionId),
+                            )
+                        coroutineContext.ensureActive()
+                        val ctx = parseContextBreakdown(result)
+                        if (ctx != null) {
+                            val currentModel = _uiState.value.currentSessionModel
+                            val modelMatches =
+                                if (ctx.model != null) {
+                                    isMatchingRpcModel(currentModel, ctx.model)
+                                } else {
+                                    !isSwitchPending
+                                }
+                            if (modelMatches) {
+                                rpcUsed = ctx.contextUsed?.takeIf { it > 0L }
+                                rpcMax = ctx.contextMax?.takeIf { it > 0L }
+                            }
+                        }
+                    } catch (e: CancellationException) {
+                        throw e
+                    } catch (_: Exception) {
+                        // Best-effort: RPC error/timeout/disconnect — keep last values.
+                    }
+                    if (!isCurrentContextFetch(
+                            sessionId,
+                            targetSessionGeneration,
+                            targetModelGeneration,
+                            targetRequestSequence,
                         )
-                    val snapshot = parseUsageSnapshot(usage)
-                    if (snapshot != null && snapshot.compressions != null) {
-                        _uiState.update { it.copy(compressionCount = snapshot.compressions) }
+                    ) {
+                        return@launch
+                    }
+                    // Compression count: how many times this session has been compacted
+                    // (session.usage → compressions). Feeds the "compressed ×N" badge —
+                    // the same usage snapshot the desktop status bar reads.
+                    if (!isSwitchPending) {
+                        try {
+                            val usage =
+                                sendRpcAndAwait(
+                                    WsMethods.SESSION_USAGE,
+                                    mapOf("session_id" to rpcSessionId),
+                                )
+                            coroutineContext.ensureActive()
+                            val snapshot = parseUsageSnapshot(usage)
+                            if (snapshot != null && snapshot.compressions != null) {
+                                _uiState.update { current ->
+                                    if (!isCurrentContextFetch(
+                                            sessionId,
+                                            targetSessionGeneration,
+                                            targetModelGeneration,
+                                            targetRequestSequence,
+                                        )
+                                    ) {
+                                        current
+                                    } else {
+                                        current.copy(compressionCount = snapshot.compressions)
+                                    }
+                                }
+                            }
+                        } catch (e: CancellationException) {
+                            throw e
+                        } catch (_: Exception) {
+                            // Best-effort: keep the last known badge value.
+                        }
                     }
-                } catch (_: Exception) {
-                    // Best-effort: keep the last known badge value.
                 }
-            }
-            // Issue #817: single atomic denominator write. The RPC's live
-            // context_max wins, REST model/info is the fallback, and a stale
-            // value from a pre-swap fetch can never overwrite a fresh one —
-            // the meter always shows ONE coherent window.
-            _uiState.update { current ->
-                val fallbackFull =
-                    if (skipRestFallback) current.fullContextTokens else restFull ?: current.fullContextTokens
-                current.copy(
-                    usedContextTokens = rpcUsed ?: current.usedContextTokens,
-                    fullContextTokens = rpcMax ?: fallbackFull,
-                )
-            }
-            // Detail-sheet accounting (cumulative REST counters, informational).
-            val usedResult =
-                safeApiCall { ApiClient.hermesApi.getSessionDetail(sessionId, profile) }
-            if (usedResult is NetworkResult.Success) {
-                val d = usedResult.data
-                val used = d.input_tokens
-                if (used != null) {
-                    _uiState.update {
-                        it.copy(
-                            contextBreakdown =
-                                ContextBreakdown(
-                                    inputTokens = used,
-                                    outputTokens = d.output_tokens ?: 0L,
-                                    cacheReadTokens = d.cache_read_tokens ?: 0L,
-                                    cacheWriteTokens = d.cache_write_tokens ?: 0L,
-                                    reasoningTokens = d.reasoning_tokens ?: 0L,
-                                    messageCount = d.message_count ?: 0,
-                                ),
+                coroutineContext.ensureActive()
+                if (!isCurrentContextFetch(
+                        sessionId,
+                        targetSessionGeneration,
+                        targetModelGeneration,
+                        targetRequestSequence,
+                    )
+                ) {
+                    return@launch
+                }
+                // Issue #817 & #1103: single atomic denominator write. The RPC's live
+                // context_max wins, REST model/info is the fallback, and a stale
+                // value from a pre-swap fetch can never overwrite a fresh one —
+                // the meter always shows ONE coherent window.
+                _uiState.update { current ->
+                    if (!isCurrentContextFetch(
+                            sessionId,
+                            targetSessionGeneration,
+                            targetModelGeneration,
+                            targetRequestSequence,
+                        )
+                    ) {
+                        current
+                    } else {
+                        val fallbackFull =
+                            if (skipRestFallback) current.fullContextTokens else restFull ?: current.fullContextTokens
+                        current.copy(
+                            usedContextTokens = rpcUsed ?: current.usedContextTokens,
+                            fullContextTokens = rpcMax ?: fallbackFull,
                         )
                     }
                 }
+                // Detail-sheet accounting (cumulative REST counters, informational).
+                coroutineContext.ensureActive()
+                if (!isCurrentContextFetch(
+                        sessionId,
+                        targetSessionGeneration,
+                        targetModelGeneration,
+                        targetRequestSequence,
+                    )
+                ) {
+                    return@launch
+                }
+                val usedResult =
+                    safeApiCall { ApiClient.hermesApi.getSessionDetail(sessionId, profile) }
+                coroutineContext.ensureActive()
+                if (usedResult is NetworkResult.Success) {
+                    val d = usedResult.data
+                    val used = d.input_tokens
+                    if (used != null) {
+                        _uiState.update { current ->
+                            if (!isCurrentContextFetch(
+                                    sessionId,
+                                    targetSessionGeneration,
+                                    targetModelGeneration,
+                                    targetRequestSequence,
+                                )
+                            ) {
+                                current
+                            } else {
+                                current.copy(
+                                    contextBreakdown =
+                                        ContextBreakdown(
+                                            inputTokens = used,
+                                            outputTokens = d.output_tokens ?: 0L,
+                                            cacheReadTokens = d.cache_read_tokens ?: 0L,
+                                            cacheWriteTokens = d.cache_write_tokens ?: 0L,
+                                            reasoningTokens = d.reasoning_tokens ?: 0L,
+                                            messageCount = d.message_count ?: 0,
+                                        ),
+                                )
+                            }
+                        }
+                    }
+                }
             }
-        }
     }
 
     private suspend fun fetchServerMessageCount(

app/src/test/java/com/m57/hermescontrol/data/model/ContextBreakdownResponseTest.kt

@@ -84,4 +84,24 @@ class ContextBreakdownResponseTest {
         assertEquals(262144L, parsed?.contextMax)
         assertEquals(0, parsed?.contextPercent)
     }
+
+    @Test
+    fun `JsonObject and JsonPrimitive payloads parse successfully`() {
+        val json =
+            kotlinx.serialization.json.buildJsonObject {
+                put("context_used", kotlinx.serialization.json.JsonPrimitive(52428))
+                put("context_max", kotlinx.serialization.json.JsonPrimitive(524288))
+                put("context_percent", kotlinx.serialization.json.JsonPrimitive(10))
+                put("estimated_total", kotlinx.serialization.json.JsonPrimitive(52428))
+                put("model", kotlinx.serialization.json.JsonPrimitive("upstage/solar-pro4:free"))
+            }
+
+        val parsed = parseContextBreakdown(json)
+
+        assertEquals(52428L, parsed?.contextUsed)
+        assertEquals(524288L, parsed?.contextMax)
+        assertEquals(10, parsed?.contextPercent)
+        assertEquals(52428L, parsed?.estimatedTotal)
+        assertEquals("upstage/solar-pro4:free", parsed?.model)
+    }
 }

app/src/test/java/com/m57/hermescontrol/data/model/UsageSnapshotResponseTest.kt

@@ -30,4 +30,24 @@ class UsageSnapshotResponseTest {
         assertNull(parseUsageSnapshot("oops"))
         assertNull(parseUsageSnapshot(listOf(1.0)))
     }
+
+    @Test
+    fun `JsonObject and JsonPrimitive payloads parse successfully`() {
+        val json =
+            kotlinx.serialization.json.buildJsonObject {
+                put("compressions", kotlinx.serialization.json.JsonPrimitive(3))
+                put("context_used", kotlinx.serialization.json.JsonPrimitive(12000))
+                put("context_max", kotlinx.serialization.json.JsonPrimitive(200000))
+                put("total", kotlinx.serialization.json.JsonPrimitive(50000))
+                put("avg_tps", kotlinx.serialization.json.JsonPrimitive(45.5))
+            }
+
+        val parsed = parseUsageSnapshot(json)
+
+        assertEquals(3, parsed?.compressions)
+        assertEquals(12000L, parsed?.contextUsed)
+        assertEquals(200000L, parsed?.contextMax)
+        assertEquals(50000L, parsed?.totalTokens)
+        assertEquals(45.5, parsed?.avgTps ?: 0.0, 0.001)
+    }
 }

app/src/test/java/com/m57/hermescontrol/ui/chat/ChatModelSwitchDelegateTest.kt

@@ -30,6 +30,7 @@ class ChatModelSwitchDelegateTest {
     private val slashCommands = mutableListOf<String>()
     private val assistantMessages = mutableListOf<String>()
     private var contextRefetched = 0
+    private var modelSwitchInitiated = 0
     private var pinnedList = mutableListOf(PinnedModel("anthropic", "claude-3"))
 
     private val fakeResponse =
@@ -60,6 +61,7 @@ class ChatModelSwitchDelegateTest {
             addAssistantMessage = { assistantMessages.add(it) },
             handleSlashCommand = { slashCommands.add(it) },
             fetchContextUsage = { contextRefetched++ },
+            onModelSwitchInitiated = { modelSwitchInitiated++ },
             getModelOptionsCall = { NetworkResult.Success(fakeResponse) },
             getPinnedModels = { pinnedList },
             savePinnedModels = { pinnedList = it.toMutableList() },
@@ -92,7 +94,24 @@ class ChatModelSwitchDelegateTest {
         assertEquals("openai/gpt-4o", uiState.value.currentSessionModel)
         assertFalse(uiState.value.showModelPicker)
         assertEquals(listOf("/model gpt-4o --provider openai --session"), slashCommands)
-        assertEquals(1, contextRefetched)
+        assertEquals(1, modelSwitchInitiated)
+    }
+
+    @Test
+    fun sendSlashModel_withSameModel_doesNotInitiateSwitchOrBlankTokens() {
+        uiState.value =
+            uiState.value.copy(
+                currentSessionModel = "openai/gpt-4o",
+                fullContextTokens = 128_000L,
+                showModelPicker = true,
+            )
+        val initialInitiated = modelSwitchInitiated
+        delegate.sendSlashModel("openai", "gpt-4o")
+        assertEquals("openai/gpt-4o", uiState.value.currentSessionModel)
+        assertEquals(128_000L, uiState.value.fullContextTokens)
+        assertEquals(initialInitiated, modelSwitchInitiated)
+        assertEquals(emptyList<String>(), slashCommands)
+        assertFalse(uiState.value.showModelPicker)
     }
 
     @Test
@@ -139,6 +158,21 @@ class ChatModelSwitchDelegateTest {
             assertEquals("anthropic/claude-3", uiState.value.currentSessionModel)
         }
 
+    @Test
+    fun handleConfigSetError_rollsBackToPreviousModel_onlyWhenMatchingCurrentSequence() =
+        testScope.runTest {
+            delegate.sendSlashModel("openai", "gpt-4o")
+            advanceUntilIdle()
+            delegate.handleModelSwitch("/model gpt-4o --provider openai --session")
+            advanceUntilIdle()
+            assertEquals("openai/gpt-4o", uiState.value.currentSessionModel)
+
+            // Current switch error rolls back
+            delegate.handleConfigSetError("req-1", "Model unavailable")
+            assertEquals("anthropic/claude-3", uiState.value.currentSessionModel)
+            assertNull(uiState.value.fullContextTokens)
+        }
+
     @Test
     fun toggleFastMode_sendsConfigSetFast_andUpdatesOnAck() =
         testScope.runTest {
@@ -207,4 +241,53 @@ class ChatModelSwitchDelegateTest {
             advanceUntilIdle()
             assertEquals(sentCount, sentMethods.size)
         }
+
+    @Test
+    fun handleConfigSetResult_staleSequenceConfirmRequired_isIgnored() =
+        testScope.runTest {
+            delegate.sendSlashModel("openai", "gpt-4o")
+            advanceUntilIdle()
+            delegate.handleModelSwitch("/model gpt-4o --provider openai --session")
+            advanceUntilIdle()
+
+            // A second switch arrives before first confirms
+            delegate.sendSlashModel("anthropic", "claude-3-5-sonnet")
+            advanceUntilIdle()
+            delegate.handleModelSwitch("/model claude-3-5-sonnet --provider anthropic --session")
+            advanceUntilIdle()
+
+            // First switch response arrives with confirm_required
+            delegate.handleConfigSetResult(
+                id = "req-1",
+                result = mapOf("key" to "model", "confirm_required" to true),
+            )
+            // Guarded: sequence 1 != current sequence 2, so confirmation is ignored
+            assertNull(uiState.value.modelSwitchConfirmMessage)
+        }
+
+    @Test
+    fun rapidPicks_preservesOriginalConfirmedModelForRollback() =
+        testScope.runTest {
+            // Initially confirmed model is anthropic/claude-3
+            delegate.onModelConfirmed("anthropic/claude-3")
+            uiState.value = uiState.value.copy(currentSessionModel = "anthropic/claude-3")
+
+            // Pick 1: switch to gpt-4o
+            delegate.sendSlashModel("openai", "gpt-4o")
+            advanceUntilIdle()
+            delegate.handleModelSwitch("/model gpt-4o --provider openai --session")
+            advanceUntilIdle()
+            assertEquals("openai/gpt-4o", uiState.value.currentSessionModel)
+
+            // Pick 2: rapidly pick solar without gpt-4o ever confirming
+            delegate.sendSlashModel("nous", "solar")
+            advanceUntilIdle()
+            delegate.handleModelSwitch("/model solar --provider nous --session")
+            advanceUntilIdle()
+            assertEquals("nous/solar", uiState.value.currentSessionModel)
+
+            // Error on second switch (req-2) should roll back to confirmed model, not optimistic gpt-4o
+            delegate.handleConfigSetError("req-2", "Solar failed")
+            assertEquals("anthropic/claude-3", uiState.value.currentSessionModel)
+        }
 }

app/src/test/java/com/m57/hermescontrol/ui/chat/ChatViewModelTest.kt

@@ -2916,6 +2916,113 @@ class ChatViewModelTest {
             assertEquals(262_144L, viewModel.uiState.value.fullContextTokens)
         }
 
+    @Test
+    fun testPickerModelSwitch_blanksMeterAndResolvesNewContextWindow_issue1103() =
+        runTest {
+            stubEmptySessionRests("session-a")
+            val (viewModel, _) = createViewModelWithSession()
+            advanceUntilIdle()
+
+            // Profile default model is 900k
+            coEvery { ApiClient.hermesApi.getModelInfo() } returns
+                retrofit2.Response.success(
+                    com.m57.hermescontrol.data.model.ModelInfoResponse(
+                        model = "gpt-5.6-luna-900k",
+                        provider = "openai-codex",
+                        effective_context_length = 900_000L,
+                    ),
+                )
+
+            // Initial session.info for default model
+            mockEventsFlow.emit(
+                WsEvent.SessionInfo(
+                    mapOf("model" to "gpt-5.6-luna-900k", "provider" to "openai-codex"),
+                ),
+            )
+            advanceUntilIdle()
+            assertEquals("openai-codex/gpt-5.6-luna-900k", viewModel.uiState.value.currentSessionModel)
+
+            var breakdownResult: Any =
+                mapOf("context_max" to 900_000L, "context_used" to 10_000L)
+            every { HermesWsClient.request(WsMethods.SESSION_CONTEXT_BREAKDOWN, any(), any()) } answers {
+                CompletableDeferred<Any?>(breakdownResult)
+            }
+            viewModel.fetchContextUsage()
+            advanceUntilIdle()
+            assertEquals(900_000L, viewModel.uiState.value.fullContextTokens)
+
+            // User picks Solar 200k via model picker (sendSlashModel)
+            viewModel.sendSlashModel("nous", "upstage/solar-pro4:free")
+            advanceUntilIdle()
+            assertEquals("nous/upstage/solar-pro4:free", viewModel.uiState.value.currentSessionModel)
+            assertNull("Picker model switch must immediately blank meter", viewModel.uiState.value.fullContextTokens)
+
+            // Backend replies with session.info confirming Solar
+            // Breakdown RPC returns JsonObject (from wire) with 524288 context_max
+            breakdownResult =
+                kotlinx.serialization.json.buildJsonObject {
+                    put("context_max", kotlinx.serialization.json.JsonPrimitive(524288))
+                    put("context_used", kotlinx.serialization.json.JsonPrimitive(20000))
+                }
+            mockEventsFlow.emit(
+                WsEvent.SessionInfo(
+                    mapOf("model" to "upstage/solar-pro4:free", "provider" to "nous"),
+                ),
+            )
+            advanceUntilIdle()
+
+            // Verified: meter resolves to Solar's 524288 window, NOT profile's 900k
+            assertEquals(524288L, viewModel.uiState.value.fullContextTokens)
+
+            // Periodic background sync with skipRestFallback=false must NOT overwrite with profile's 900k
+            viewModel.fetchContextUsage(skipRestFallback = false)
+            advanceUntilIdle()
+            assertEquals(524288L, viewModel.uiState.value.fullContextTokens)
+        }
+
+    @Test
+    fun testIsMatchingRpcModel_namespacedAndProviderMatching() =
+        runTest {
+            stubEmptySessionRests("session-a")
+            val (viewModel, _) = createViewModelWithSession()
+
+            // Current session model has provider prefix, rpc model retains internal namespace
+            assertTrue(
+                viewModel.isMatchingRpcModel(
+                    "nous/upstage/solar-pro4:free",
+                    "upstage/solar-pro4:free",
+                ),
+            )
+            // Exact match
+            assertTrue(
+                viewModel.isMatchingRpcModel(
+                    "upstage/solar-pro4:free",
+                    "upstage/solar-pro4:free",
+                ),
+            )
+            // Provider prefix with non-namespaced model
+            assertTrue(
+                viewModel.isMatchingRpcModel(
+                    "openai/gpt-4o",
+                    "gpt-4o",
+                ),
+            )
+            // Never blindly strip model's own namespace: rpcModel missing internal namespace must not match
+            assertFalse(
+                viewModel.isMatchingRpcModel(
+                    "nous/upstage/solar-pro4:free",
+                    "solar-pro4:free",
+                ),
+            )
+            // Mismatch
+            assertFalse(
+                viewModel.isMatchingRpcModel(
+                    "nous/upstage/solar-pro4:free",
+                    "claude-3",
+                ),
+            )
+        }
+
     @Test
     fun testSessionInfoSameModel_keepsMeterUntouched() =
         runTest {