baro-ai 0.70.7 → 0.70.8

package/dist/cli.mjs

@@ -47767,7 +47767,7 @@ exactly this shape:
 {"action":"split"|"prereq"|"rewire"|"skip"|"abort",
  "reason":"\u2026",
  "added":[ { "id":"S?","priority":N,"title":"\u2026","description":"\u2026",
-             "dependsOn":["\u2026"], "acceptance":["\u2026"], "model":"sonnet" } ],
+             "dependsOn":["\u2026"], "acceptance":["\u2026"] } ],
  "removed":["S?"],
  "modifiedDeps":[{"id":"S?","newDependsOn":["\u2026"]}]}
 
@@ -47777,18 +47777,22 @@ Rules:
 - "modifiedDeps" rewires a story's dependsOn \u2014 use to repoint dependents
   of a removed story to a replacement.
 - "abort" \u2192 empty added/removed/modifiedDeps arrays.
-- TIER every added story with "model" ("haiku" | "sonnet" | "opus"),
-  the same way the planner does \u2014 by blast radius, not raw difficulty:
-    * "haiku"  \u2192 mechanical, single-concern, nothing important breaks
-    * "sonnet" \u2192 one contained feature/module
-    * "opus"   \u2192 cross-cutting / schema / wiring / a DAG hub
-  ESCALATION: the failing story already burned its retries at the tier
-  shown ("Tier that just failed"). Its replacement(s) must NOT repeat
-  that tier unless you are splitting it into genuinely smaller pieces.
-  When you keep the same scope (rewire / prereq replacement), bump the
-  tier UP one step (haiku\u2192sonnet\u2192opus). When you split (action "split"),
-  each child gets the tier its own (smaller) blast radius warrants \u2014
-  often lower, but a still-complex child stays "opus".
+- MODEL: LEAVE "model" UNSET on the stories you add \u2014 they run on the
+  default (cheaper) model, which is exactly what split children want.
+  Do NOT use planner tier names ("haiku"/"sonnet"/"opus") \u2014 the story
+  model is not chosen by tier here; it is either the default or an
+  explicit escalation route (below).
+- ESCALATION vs SPLIT \u2014 the failing story already burned its retries on
+  the model shown ("Model that just failed"). Two ways to recover:
+    * SPLIT (preferred): if it was TOO BROAD \u2014 too many files/concerns
+      for one session \u2014 break it into smaller, focused stories and
+      leave their "model" unset (they stay on the cheaper model). A
+      smaller, sharper story is usually what a stuck run actually needs.
+    * ESCALATE (sparingly): if the story was already RIGHT-SIZED but
+      genuinely needs a more capable model, set that ONE story's "model"
+      to the exact ESCALATION ROUTE printed in the failure context
+      below. That runs it on the stronger model. Only escalate when the
+      scope is already tight \u2014 never as a reflex.
 - Output ONLY the JSON object, nothing else.`;
 var Surgeon = class extends BaseObserver {
   opts;
@@ -47803,7 +47807,8 @@ var Surgeon = class extends BaseObserver {
       claudeBin: opts.claudeBin ?? "claude",
       timeoutMs: opts.timeoutMs ?? 9e4,
       snapshot: opts.snapshot,
-      resolveRoute: opts.resolveRoute
+      resolveRoute: opts.resolveRoute,
+      escalationRoute: opts.escalationRoute
     };
   }
   /** Resolves once every in-flight LLM evaluation has completed. */
@@ -47842,7 +47847,7 @@ var Surgeon = class extends BaseObserver {
    */
   async evaluateWithLlm(failure) {
     const snap = this.opts.snapshot();
-    const prompt = buildSurgeonPrompt(snap, failure, this.opts.resolveRoute);
+    const prompt = buildSurgeonPrompt(snap, failure, this.opts.resolveRoute, this.opts.escalationRoute);
     try {
       const { stdout } = await execFileAsync3(
         this.opts.claudeBin,
@@ -47892,7 +47897,7 @@ var Surgeon = class extends BaseObserver {
     }
   }
 };
-function buildSurgeonPrompt(snap, failure, resolveRoute) {
+function buildSurgeonPrompt(snap, failure, resolveRoute, escalationRoute) {
   const storyLines = snap.stories.map(
     (s2) => `  - ${s2.id} ${s2.passes ? "[passed]" : "[pending]"} ${s2.model ? `<tier:${s2.model}> ` : ""}"${s2.title}" deps=${JSON.stringify(s2.dependsOn)}`
   ).join("\n");
@@ -47915,6 +47920,12 @@ function buildSurgeonPrompt(snap, failure, resolveRoute) {
     ] : [],
     `Attempts: ${failure.attempts}`,
     `Error: ${failure.error ?? "(no reason captured)"}`,
+    ...escalationRoute ? [
+      "",
+      `# Escalation route`,
+      `To ESCALATE a right-sized story onto the stronger model, set that story's "model" to EXACTLY: ${escalationRoute}`,
+      `Otherwise leave "model" unset \u2014 added stories run on the default (cheaper) model. Prefer splitting a too-broad story over escalating.`
+    ] : [],
     "",
     `# Decide`,
     `Output the replan JSON per the rules in your system prompt.`
@@ -47964,7 +47975,8 @@ var SurgeonCodex = class extends BaseObserver {
       codexBin: opts.codexBin ?? "codex",
       timeoutMs: opts.timeoutMs ?? 3e5,
       snapshot: opts.snapshot,
-      resolveRoute: opts.resolveRoute
+      resolveRoute: opts.resolveRoute,
+      escalationRoute: opts.escalationRoute
     };
   }
   async idle() {
@@ -47988,7 +48000,7 @@ var SurgeonCodex = class extends BaseObserver {
   }
   async evaluateWithLlm(failure) {
     const snap = this.opts.snapshot();
-    const userPrompt = buildSurgeonPrompt(snap, failure, this.opts.resolveRoute);
+    const userPrompt = buildSurgeonPrompt(snap, failure, this.opts.resolveRoute, this.opts.escalationRoute);
     const prompt = `${SURGEON_SYSTEM_PROMPT}
 
 ${userPrompt}`;
@@ -48061,7 +48073,8 @@ var SurgeonOpenAI = class extends BaseObserver {
       maxReplans: opts.maxReplans ?? Infinity,
       model: opts.model ?? "gpt-5.5",
       snapshot: opts.snapshot,
-      resolveRoute: opts.resolveRoute
+      resolveRoute: opts.resolveRoute,
+      escalationRoute: opts.escalationRoute
     };
     this.model = pickModel3(this.opts.model);
   }
@@ -48093,7 +48106,7 @@ var SurgeonOpenAI = class extends BaseObserver {
    */
   async evaluate(failure) {
     const snap = this.opts.snapshot();
-    const userPrompt = buildSurgeonPrompt(snap, failure, this.opts.resolveRoute);
+    const userPrompt = buildSurgeonPrompt(snap, failure, this.opts.resolveRoute, this.opts.escalationRoute);
     const context = ModelContext.create("surgeon").addContextItem(SystemMessageItem.create(SURGEON_SYSTEM_PROMPT)).addContextItem(UserMessageItem.create(userPrompt));
     try {
       const round = await runInferenceRound(context, this.model);
@@ -48151,7 +48164,8 @@ var SurgeonOpenCode = class extends BaseObserver {
       opencodeBin: opts.opencodeBin ?? "opencode",
       timeoutMs: opts.timeoutMs ?? 3e5,
       snapshot: opts.snapshot,
-      resolveRoute: opts.resolveRoute
+      resolveRoute: opts.resolveRoute,
+      escalationRoute: opts.escalationRoute
     };
   }
   async idle() {
@@ -48175,7 +48189,7 @@ var SurgeonOpenCode = class extends BaseObserver {
   }
   async evaluateWithLlm(failure) {
     const snap = this.opts.snapshot();
-    const userPrompt = buildSurgeonPrompt(snap, failure, this.opts.resolveRoute);
+    const userPrompt = buildSurgeonPrompt(snap, failure, this.opts.resolveRoute, this.opts.escalationRoute);
     const prompt = `${SURGEON_SYSTEM_PROMPT}
 
 ${userPrompt}`;
@@ -48231,7 +48245,8 @@ var SurgeonPi = class extends BaseObserver {
       piBin: opts.piBin ?? "pi",
       timeoutMs: opts.timeoutMs ?? 3e5,
       snapshot: opts.snapshot,
-      resolveRoute: opts.resolveRoute
+      resolveRoute: opts.resolveRoute,
+      escalationRoute: opts.escalationRoute
     };
   }
   async idle() {
@@ -48255,7 +48270,7 @@ var SurgeonPi = class extends BaseObserver {
   }
   async evaluateWithLlm(failure) {
     const snap = this.opts.snapshot();
-    const userPrompt = buildSurgeonPrompt(snap, failure, this.opts.resolveRoute);
+    const userPrompt = buildSurgeonPrompt(snap, failure, this.opts.resolveRoute, this.opts.escalationRoute);
     const prompt = `${SURGEON_SYSTEM_PROMPT}
 
 ${userPrompt}`;
@@ -48552,16 +48567,20 @@ async function orchestrate(config) {
         return null;
       }
     } : void 0;
+    const surgeonEscalationModel = config.surgeonModel ?? (surgeonLlm === "openai" ? "gpt-5.5" : surgeonLlm === "claude" ? "opus" : void 0);
+    const escalationRoute = surgeonEscalationModel && !config.storyModel ? `${surgeonLlm}:${surgeonEscalationModel}` : void 0;
     if (surgeonLlm === "openai") {
       surgeon = new SurgeonOpenAI({
         snapshot,
         resolveRoute,
+        escalationRoute,
         model: config.surgeonModel ?? "gpt-5.5"
       });
     } else if (surgeonLlm === "codex") {
       surgeon = new SurgeonCodex({
         snapshot,
         resolveRoute,
+        escalationRoute,
         useLlm: config.surgeonUseLlm ?? true,
         model: config.surgeonModel
       });
@@ -48569,6 +48588,7 @@ async function orchestrate(config) {
       surgeon = new SurgeonOpenCode({
         snapshot,
         resolveRoute,
+        escalationRoute,
         useLlm: config.surgeonUseLlm ?? true,
         model: config.surgeonModel
       });
@@ -48576,6 +48596,7 @@ async function orchestrate(config) {
       surgeon = new SurgeonPi({
         snapshot,
         resolveRoute,
+        escalationRoute,
         useLlm: config.surgeonUseLlm ?? true,
         model: config.surgeonModel
       });
@@ -48583,6 +48604,7 @@ async function orchestrate(config) {
       surgeon = new Surgeon({
         snapshot,
         resolveRoute,
+        escalationRoute,
         useLlm: config.surgeonUseLlm ?? false,
         model: config.surgeonModel ?? "opus"
       });