infinity-harness 2.8.8 → 2.8.9

package/CHANGELOG.md

@@ -4,6 +4,16 @@ All notable changes to this project are documented here.
 Format follows [Keep a Changelog](https://keepachangelog.com/en/1.1.0/); versions follow
 [Semantic Versioning](https://semver.org/spec/v2.0.0.html).
 
+## [2.8.9] — 2026-08-31
+
+Locals no longer die on xhigh; DEFINE ladder re-consults.
+
+### Fixed
+
+- **Locals cut off at 1 token on xhigh.** LM Studio (lm-studio/*) forwarded reasoning_effort xhigh/max as max_completion_tokens:1. Added provider-aware clamp: lm-studio tiers (B/C/D/X when local) never get xhigh/max/high — they are clamped to medium at the router, in daemon/index and in worker createWorker, with a [worker] trace in harness/daemon.log. master (muse-spark) still gets xhigh. Verified: no more length=1 on lfm/qwen/granite.
+
+- **DEFINE never re-consulted after RESEARCH spent the ladder.** perLevelEscalation[subtask] kept consultedCount/masterUsed across phase advance (research/r1 → define/d1), so consult/master appeared spent. decideNext now resets perLevelEscalation when retryLevel changes or when tree moves; escalation instruction for feature-criteria now explicitly says to use infinity_plan to add criteria.
+
 ## [2.8.8] — 2026-08-31
 
 Routing is one source; default honestly idle.

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "infinity-harness",
-  "version": "2.8.8",
+  "version": "2.8.9",
   "description": "A pi agent extension that runs a gated build pipeline unattended \u2014 enforces phases, validates with deterministic gates, and keeps working for hours or days without losing the plan.",
   "type": "module",
   "keywords": [

package/src/core/modelRouter.ts

@@ -13,6 +13,15 @@ import type { TierMap, TierSpec, HarnessConfig, FeatureList, Phase, Difficulty }
 export type ThinkingLevel = import("./types.ts").ThinkingLevel;
 export type TierId = import("./types.ts").TierId;
 
+export function clampThinkingForProvider(provider: string, level: string | "" | undefined): string | "" {
+  if (!level) return (level ?? "") as string | "";
+  const pp = String(provider || "").toLowerCase();
+  const isLocal = pp === "lm-studio" || pp.startsWith("lm-studio/") || pp === "lmstudio";
+  if (!isLocal) return level as string | "";
+  if (level === "xhigh" || level === "max" || level === "high") return "medium";
+  return level as string | "";
+}
+
 const DIFF_TO_TIER: Record<Difficulty, import("./types.ts").TierId> = {
   easy: "B",
   moderate: "C",

package/src/daemon/index.ts

@@ -26,7 +26,7 @@ import { saveSupervisor, appendActivity, type SupervisorWorker } from "./supervi
 import { runPreflight } from "./preflight.ts";
 import { addUsageForTier, isCapExceeded, hasXLeak, xLeakReason, type Tier } from "./budget.ts";
 import { createWorker, promptWorker, type TurnResult } from "./worker.ts";
-import { routeModel, effectiveDifficultyForTask as effectiveDifficulty } from "../core/modelRouter.ts";
+import { routeModel, effectiveDifficultyForTask as effectiveDifficulty, clampThinkingForProvider } from "../core/modelRouter.ts";
 import { dirname } from "node:path";
 import type { Server } from "node:http";
 
@@ -275,7 +275,7 @@ async function runLoop(targetDir: string): Promise<void> {
         model: askedModel,
         askedModel,
         servedModel: null,
-        thinking: (tierSpec as { thinkingLevel?: string } | null)?.thinkingLevel ?? "medium",
+        thinking: clampThinkingForProvider(routed.provider, ((tierSpec as { thinkingLevel?: string } | null)?.thinkingLevel as unknown as string) ?? "medium") || "medium",
         state: "starting",
         doing: "starting",
         startedAt: new Date().toISOString(),
@@ -298,7 +298,7 @@ async function runLoop(targetDir: string): Promise<void> {
       const workerFactory = await import("./worker.ts");
       workerHandle = await workerFactory.createWorker({
         cwd: targetDir,
-        modelSpec: { provider: routed.provider, id: routed.id, thinkingLevel: (tierSpec as { thinkingLevel?: string } | null)?.thinkingLevel ?? "medium" },
+        modelSpec: { provider: routed.provider, id: routed.id, thinkingLevel: clampThinkingForProvider(routed.provider, ((tierSpec as { thinkingLevel?: string } | null)?.thinkingLevel as unknown as string) ?? "medium") || "medium" },
         askedModel,
         sessionManagerDir: `harness/sessions/${workerLabel.replace(/[^a-z0-9._-]/gi, "-")}`,
         customTools: (await import("./isolation.ts")).harnessToolsForWorker() as unknown[],
@@ -317,7 +317,7 @@ async function runLoop(targetDir: string): Promise<void> {
           if (!workerHandle) {
             workerHandle = await workerFactory.createWorker({
               cwd: targetDir,
-              modelSpec: { provider: routed.provider, id: routed.id, thinkingLevel: (tierSpec as { thinkingLevel?: string } | null)?.thinkingLevel ?? "medium" },
+              modelSpec: { provider: routed.provider, id: routed.id, thinkingLevel: clampThinkingForProvider(routed.provider, ((tierSpec as { thinkingLevel?: string } | null)?.thinkingLevel as unknown as string) ?? "medium") || "medium" },
               askedModel,
               sessionManagerDir: `harness/sessions/${workerLabel.replace(/[^a-z0-9._-]/gi, "-")}`,
               customTools: (await import("./isolation.ts")).harnessToolsForWorker() as unknown[],

package/src/daemon/worker.ts

@@ -110,7 +110,15 @@ export async function createWorker(opts: CreateWorkerOpts): Promise<{
     ? SessionManager.create(opts.cwd, opts.sessionManagerDir)
     : SessionManager.inMemory(opts.cwd);
 
-  const thinkingLevel = (opts.thinkingLevel ?? opts.modelSpec.thinkingLevel ?? "medium") as unknown as string;
+  let thinkingLevel = (opts.thinkingLevel ?? opts.modelSpec.thinkingLevel ?? "medium") as unknown as string;
+  // A: clamp for lm-studio (xhigh/max not supported, caused max_completion_tokens:1)
+  try {
+    const p = String(opts.modelSpec.provider || "").toLowerCase();
+    const isLocal = p === "lm-studio" || p.startsWith("lm-studio/") || p === "lmstudio";
+    if (isLocal && (thinkingLevel === "xhigh" || thinkingLevel === "max" || thinkingLevel === "high")) thinkingLevel = "medium";
+  } catch {}
+  // B: debug trace of routed model/thinking (readable in harness/daemon.log)
+  try { console.log("[worker] model=" + opts.modelSpec.provider + "/" + opts.modelSpec.id + " thinking=" + thinkingLevel + " askedModel=" + opts.askedModel); } catch {}
 
   const agentTools = opts.customTools as unknown as import("@earendil-works/pi-coding-agent").ToolDefinition[] | undefined;
 

package/src/loop.ts

@@ -486,11 +486,18 @@ export async function decideNext(options: DecideOptions): Promise<{ decision: Lo
     // budgets in rework.json and replan.json still bound the run across
     // stalls, but a rung spent on a problem that resolved should not be
     // missing when a different problem appears.
-    state.escalation = { ...state.escalation, tried: [] };
-    const lev = retryLevel
-      ? escalationStateForLevel(state, retryLevel)
-      : null;
-    if (lev) setEscalationForLevel(state, retryLevel, { ...lev, tried: [] });
+    // C: also reset per-level escalation when the active unit changes (define/d1 vs research/r1),
+    // otherwise consult/master stays marked spent across unrelated tasks.
+    const lastLev = state.escalations.length ? state.escalations[state.escalations.length-1]?.level as string | undefined : undefined;
+    if (lastLev && lastLev !== (retryLevel ?? "task")) {
+      setEscalationForLevel(state, retryLevel ?? "task", { consultedCount: 0, masterUsed: false, lastUnstuckAt: null, fingerprints: [], tried: [] } as unknown as import("./escalate.ts").EscalationState);
+    } else {
+      state.escalation = { ...state.escalation, tried: [] };
+      const lev = retryLevel
+        ? escalationStateForLevel(state, retryLevel)
+        : null;
+      if (lev) setEscalationForLevel(state, retryLevel, { ...lev, tried: [] });
+    }
   } else {
     state.noProgressStreak += 1;
   }
@@ -597,6 +604,13 @@ export async function decideNext(options: DecideOptions): Promise<{ decision: Lo
   const task = nextActionableTask(list);
   const focus = task ? `\nCurrent task: ${task.compositeKey} — ${task.description}` : "";
 
+  // C: feature-criteria guard — when DEFINE stalls on feature-criteria, the plan is wrong, not the model.
+  if (escalation && escalation.strategy === "consult" && gate && (gate.failures as string[]).some((f: string) => String(f).toLowerCase().includes("feature-criteria"))) {
+    try {
+      const msg = "The " + phase.toUpperCase() + " gate is blocked on feature-criteria: every feature needs acceptance criteria. This is a plan gap — use infinity_plan to add criteria, not just a model switch. Escalating to " + (escalation.model ?? "master") + " for help writing them.";
+      escalation = { ...escalation, instruction: msg + "\n" + (escalation.instruction ?? "") } as unknown as typeof escalation;
+    } catch {}
+  }
   // An escalation replaces the standard "fix these" nudge, because repeating
   // that nudge is exactly what the ladder exists to interrupt.
   const head = standingRejection