mixdog 0.9.91 → 0.9.93

package/src/tui/engine/turn.mjs

@@ -7,10 +7,12 @@ import { pickVerb, pickDoneVerb, compactEventLabel, compactEventDetail } from '.
 import { toolResultText, toolErrorDisplay } from './tool-result-text.mjs';
 import { toolCallId, toolResultCallId, toolCallName, toolCallArgs } from './tool-call-fields.mjs';
 import { promptDisplayText, STEERING_SUPPRESSED_DISPLAY } from './queue-helpers.mjs';
-import { yieldToRenderer } from './render-timing.mjs';
+import { TUI_FRAME_MS, yieldToRenderer } from './render-timing.mjs';
 import { createTurnWatchdog } from './turn-watchdog.mjs';
 import { aggregateRawResult, aggregateBucketForCategory, aggregateSummaries, assignAggregateSummaryOrder, failureDetailText, toolCallOutcome } from './tool-result-status.mjs';
 
+export const STREAM_BATCH_INTERVAL_MS = TUI_FRAME_MS;
+
 export function createRunTurn(bag) {
   const {
     runtime, nextId, tuiDebug, LEAD_TURN_TIMEOUT_MS, flags, pending, itemIndexById, getState, set, flushEmit, flushEmitImmediate, pushItem, appendItems, patchItem, replaceItems, updateStreamingTail: updateStreamingTailFromStore, settleStreamingTail: settleStreamingTailFromStore, clearStreamingTail: clearStreamingTailFromStore, pushNotice, pushUserOrSyntheticItem, markToolCallActive, markToolCallDone, clearActiveToolSummary, agentStatusState, routeState, transcriptRouteMetadata, syncContextStats, denyAllToolApprovals, requestToolApproval, patchToolCardResult, flushToolResults, flushDeferredExecutionPendingResumeKick, drain, drainPendingSteering,
@@ -563,15 +565,12 @@ export function createRunTurn(bag) {
     // --- Streaming-delta batcher ---
     // onTextDelta and onReasoningDelta fire on every tiny chunk (often <10 chars).
     // Each call previously called set() → emit() → full React reconcile. We
-    // batch accumulated text and flush at most once per STREAM_BATCH_INTERVAL_MS
-    // (≈16ms / 60fps cap). A forced flush happens before any tool call,
+    // batch accumulated text and flush at most once per STREAM_BATCH_INTERVAL_MS.
+    // A forced flush happens before any tool call,
     // finalization, or error so those code paths see the correct text getState().
-    // Flush cadence for streamed text/thinking. 8ms (~120fps) matches the Ink
-    // render maxFps (index.jsx render({ maxFps: 120 })), so a queued batch is
-    // never held back waiting for the next Ink frame. 16ms (~60fps) left every
-    // other Ink frame idle, which made fast provider streams visibly land in
-    // coarse chunks ("10 chars at a time").
-    const STREAM_BATCH_INTERVAL_MS = 16;
+    // Share Ink's exact 120fps cadence with the frame-batched store. A separate
+    // 16ms timer left every other terminal frame idle, so completed script lines
+    // accumulated and then climbed into view in coarse two-step chunks.
     let _batchTimer = null;
     let _pendingTextFlush = false;   // true when a text/spinner update is queued
     let _pendingThinkFlush = false;  // true when a thinking update is queued

package/src/tui/engine.mjs

@@ -55,6 +55,7 @@ import {
   toolCallArgs,
 } from './engine/tool-call-fields.mjs';
 import {
+  buildExecutionResponseToolItem,
   parseBackgroundTaskEnvelope,
   parseSyntheticAgentMessage,
   toolResultStatus,
@@ -179,6 +180,7 @@ const LEAD_TURN_TIMEOUT_MS = (() => {
 // the alternate-screen render; enable with MIXDOG_TUI_DEBUG=1.
 import { tuiDebug, nextId, cleanupStaleTranscriptSpillDirs, createTranscriptSpillBuffer, refillTranscriptViewOverlap, replaceEngineItemsState, createEngineItemMutators, TRANSCRIPT_LIVE_ITEM_CAP, TRANSCRIPT_SPILL_CHUNK_ITEMS } from './engine/transcript-spill.mjs';
 export { cleanupStaleTranscriptSpillDirs, createTranscriptSpillBuffer, refillTranscriptViewOverlap, replaceEngineItemsState, createEngineItemMutators, TRANSCRIPT_LIVE_ITEM_CAP, TRANSCRIPT_SPILL_CHUNK_ITEMS } from './engine/transcript-spill.mjs';
+export { parseBackgroundTaskEnvelope } from './engine/agent-envelope.mjs';
 
 export async function createEngineSession({
   provider: providerName,
@@ -647,16 +649,15 @@ export async function createEngineSession({
     });
   };
   const pushAsyncAgentResponse = (text, id = nextId(), origin = 'injected', metadata = {}) => {
-    const synthetic = parseSyntheticAgentMessage(text);
-    const isAgent = synthetic?.name === 'agent';
-    if (!isAgent) return pushUserOrSyntheticItem(text, id, origin);
-    const responseHasBody = /\n\s*\n[\s\S]*\S/.test(String(text || ''));
-    const rawResult = synthetic.rawResult ?? text;
-    const args = {
-      ...(synthetic.args && typeof synthetic.args === 'object' ? synthetic.args : {}),
-      type: 'result',
-    };
-    const responseKey = String(metadata.responseKey || metadata.executionId || args.task_id || '').trim();
+    const responseItem = buildExecutionResponseToolItem(text, {
+      id,
+      responseKey: metadata.responseKey || metadata.executionId,
+    });
+    if (!responseItem) return pushUserOrSyntheticItem(text, id, origin);
+    if (responseItem.name !== 'agent') {
+      pushItem(responseItem);
+      return true;
+    }
     const previous = state.items.at(-1);
     // Tail-only aggregation prevents a later completion from mutating a card
     // above any outbound tool, assistant, user, or preview/body boundary.
@@ -665,43 +666,19 @@ export async function createEngineSession({
       && previous.agentDirection === 'inbound'
     ) {
       const patch = appendAgentResponseTail(previous, {
-        key: responseKey,
-        args,
-        result: synthetic.result,
-        rawResult,
-        hasBody: responseHasBody,
-        isError: synthetic.isError === true,
+        key: responseItem.agentResponseKey,
+        args: responseItem.args,
+        result: responseItem.result,
+        rawResult: responseItem.rawResult,
+        hasBody: responseItem.agentResponseHasBody,
+        isError: responseItem.isError,
       });
       if (patch) {
         patchItem(previous.id, patch);
         return true;
       }
     }
-    pushItem({
-      kind: 'tool',
-      id,
-      name: 'agent',
-      args,
-      result: synthetic.result,
-      rawResult,
-      isError: synthetic.isError === true,
-      expanded: false,
-      count: 1,
-      completedCount: 1,
-      startedAt: Date.now(),
-      completedAt: Date.now(),
-      agentDirection: 'inbound',
-      agentResponseKey: responseKey,
-      agentResponseHasBody: responseHasBody,
-      agentResponseAggregate: false,
-      agentResponseEntries: [{
-        key: responseKey,
-        raw: String(rawResult ?? '').trim(),
-        result: synthetic.result,
-        hasBody: responseHasBody,
-        isError: synthetic.isError === true,
-      }],
-    });
+    pushItem(responseItem);
     return true;
   };
   const pushToast = (text, tone = 'info', ttlMs = 3000) => {

package/src/workflows/default/WORKFLOW.md

@@ -2,7 +2,7 @@
 id: default
 name: Cowork
 description: "Parallel delegation."
-agents: worker, heavy-worker, reviewer, debugger, maintainer
+agents: worker, heavy-worker, reviewer, debugger
 ---
 
 # Cowork
@@ -13,23 +13,12 @@ investigation and planning — no edits, no state mutation, no delegation.
 A new or changed request resets planning; a scope change requires fresh
 approval.
 
-On approval, fan out at maximum width: one agent per independent scope, all
-spawned in one turn; only a scope that depends on another's output waits.
-Split the plan into as many scopes as possible: disjoint file/module sets
-are independent; merge only on a true output dependency. Prefer parallel
-scopes over sequential slices in one agent. Brief each agent per the Lead
-Brief contract.
-
-Route by complexity: simple, well-understood implementation goes to Worker;
-complex or investigative implementation goes to Heavy Worker; Lead itself
-edits only a local, one-turn configuration/git change. Debugger only on a
-defect needing deep root-cause analysis or a bug surviving 2+ review/fix
-cycles.
-
-Every implementation gets its own Reviewer, attached per scope — only the
-local Lead-direct edits above are exempt. Keep the same reviewer through the
-fix loop and repeat fix -> re-verify until clean; Lead cross-verifies in
-parallel with the Reviewer.
+On approval, delegate maximally: one agent per independent scope, fit to the
+situation, all spawned in one turn; only a scope that depends on another's
+output waits. Split the plan into as many scopes as possible: disjoint
+file/module sets are independent; merge only on a true output dependency.
+Prefer parallel scopes over sequential slices in one agent. Brief each agent
+per the Lead Brief contract.
 
 Report the verified result against the approved plan. Build, deploy, commit,
 and push happen only on an explicit user request.

package/src/workflows/solo/WORKFLOW.md

@@ -15,12 +15,6 @@ changed request resets planning; a scope change requires fresh approval.
 On approval, Lead executes all work itself — never spawn, send, or delegate
 to agents. Complete in-scope fixes without reapproval.
 
-Verification is single-pass and risk-proportional: exactly one check that
-most directly proves the deliverable — a direct result check for trivial or
-read-only work, full test/build runs only for risky or behavior-changing
-work. A pass is final. Iterate only on a failing check, re-running just that
-check after each fix, or report the blocker.
-
 Report the result against the approved plan. Build, deploy, commit, and push
 happen only on an explicit user request.
 

package/src/workflows/solo-bench/WORKFLOW.md

@@ -0,0 +1,17 @@
+---
+id: solo-bench
+name: Solo Bench
+description: "Benchmark-only Solo without the planning approval gate."
+agents:
+hidden: true
+---
+
+# Solo Bench
+
+Lead executes all work itself — never spawn, send, or delegate to agents.
+Complete in-scope fixes without reapproval.
+
+Report the result. Build, deploy, commit, and push happen only on an explicit
+user request.
+
+On direction change, pause and re-consult the user.