oc-metricboard 0.1.9 → 0.1.10

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "oc-metricboard",
-  "version": "0.1.9",
+  "version": "0.1.10",
   "description": "OpenCode TUI sidebar plugin showing real-time LLM metrics (TPS, TTFT, tokens) with sub-agent × model breakdown",
   "keywords": [
     "opencode",

package/src/collector.ts

@@ -303,14 +303,28 @@ export function createCollector(
     return { rootID: requestedSessionID, ids: requestedIDs }
   }
 
-  function aggregateByModel(ids: readonly string[], now: number, scope: MetricsScope): ModelMetrics[] {
+  function aggregateByModel(
+    ids: readonly string[],
+    now: number,
+    scope: MetricsScope,
+    rootID: string,
+    foregroundTurnStart: number,
+  ): ModelMetrics[] {
     // Group requests by (modelID, providerID) — one row per distinct model
-    // across all sub-agent sessions, with a ×N session count.
+    // across all sub-agent sessions, with a ×N session count. Apply the same
+    // foreground-turn membership filter as the token totals: a session whose
+    // last turn predates the foreground turn (e.g. a model the user switched
+    // away from hours ago) would otherwise linger as a stale zombie row that
+    // is absent from the aggregate totals.
     const modelGroups = new Map<string, RequestMetrics[]>()
 
     for (const id of ids) {
       const request = state.requests.get(id)
       if (!request) continue
+      const turn = state.turns.get(id)
+      const belongsToForegroundTurn = id === rootID
+        || Boolean(turn && (turn.turnStartTime >= foregroundTurnStart || !turn.isComplete))
+      if (!belongsToForegroundTurn) continue
 
       // Key: modelID|providerID (model-grouped, not per-session)
       const key = `${request.modelID}|${request.providerID}`
@@ -557,7 +571,7 @@ export function createCollector(
       }
 
       // NEW: Build per-model breakdown for tree scope
-      const modelBreakdown = scope === "tree" ? aggregateByModel(ids, now, scope) : []
+      const modelBreakdown = scope === "tree" ? aggregateByModel(ids, now, scope, rootID, foregroundTurnStart) : []
       // Turn-level TTFT: opencode stamps step.started at the first token, so
       // intra-turn steps cannot measure their own request start (they would
       // collapse to ~0). Anchor TTFT to the turn's user message instead —

package/src/session-hydration.ts

@@ -235,6 +235,7 @@ export async function hydrateSession(input: HydrateSessionInput): Promise<boolea
     await input.api.client.session.status({ id: input.sessionID }),
     input.sessionID,
   ))
+  log(`hydrate session=${input.sessionID.slice(0, 8)} status='${sessionStatus}'`)
   const shouldReuseExisting = existing?.messageID === assistant.messageID
 
   const current = shouldReuseExisting ? existing : createFreshMetrics(
@@ -287,7 +288,12 @@ export async function hydrateSession(input: HydrateSessionInput): Promise<boolea
     applyAssistantTokens(stepMetrics, step.tokens)
     retireRequestIntoTurn(turn, stepMetrics)
   }
-  if (sessionStatus === "idle") completeTurn(turn, completedTime ?? input.now)
+  // Turn completion: the session-status lookup is unreliable across host SDK
+  // dialects (may resolve to an unrecognizable shape → status ''). The last
+  // assistant message's own completed flag is authoritative: if it finished,
+  // its turn is over. An in-flight resumed message (completed=false) keeps
+  // the turn open.
+  if (sessionStatus === "idle" || current.isComplete) completeTurn(turn, completedTime ?? input.now)
   input.state.turns.set(input.sessionID, turn)
 
   // Write-point race guard: the checks above ran before several awaits. By