infinity-harness 2.8.9 → 2.8.11

package/CHANGELOG.md

@@ -4,6 +4,24 @@ 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).
 
+## [2.8.11] — 2026-08-31
+
+Node 26 fix — daemon ran as raw .ts under node_modules and crashed; widget never updated.
+
+### Fixed
+
+- **Daemon crashed on Node 26 — widget never updated and nothing ran.** Node 26 forbids `—experimental-strip-types` for any file under `node_modules` (`ERR_UNSUPPORTED_NODE_MODULES_TYPE_STRIPPING`). `2.8.10` spawned `src/daemon/index.ts` under `~/.pi/agent/npm/node_modules/\u2026/node_modules/infinity-harness` so the detached daemon died twice on every arm; no `daemon.json`/`supervisor.json`, no local or nemotron turns. Added a real `npm run build` (`tsc -p tsconfig.build.json \u2192 dist/`, `rewriteRelativeImportExtensions`) with `prepare` so the package now ships `dist/` (both `dist/src/daemon/index.js` and the legacy `dist/daemon/index.js` alias). Both `init` and `run` daemon spawns now prefer `dist/src/daemon/index.js` before any `src/daemon/index.ts` fallback, so Node 26 never hits the ban. Extension `pi.extensions` now declares both `dist/extensions` and `extensions` so dev checkouts without a build still work; `.gitignore` keeps `dist/` uncommitted but `files` in `package.json` ships it. Also reaffirms `tiers.A` (Default/nemotron) routing for `RESEARCH` general work — the daemon crash masked the `2.8.10` thinking fix, not the router.
+
+## [2.8.10] — 2026-08-31
+
+Thinking levels from pi, not inventoried; locals inherit correctly.
+
+### Fixed
+
+- **Thinking menu was hard-coded, not model-aware.** Wizard now filters the thinking-level picker to exactly what pi reports for the picked model (“pi truth”): local reasoning-off models (lfm2.5/qwen3.8/granite) only offer “(inherit)” + “off”; reasoning locals offer off…high; full providers offer off…max. “/infinity:config” thinking settings do the same per-tier. Previously all 7 were always offered, so locals could be set to xhigh.
+- **Locals died on “inherit” even with thinking blank.** Worker forced “inherit” (“”) to “medium” before “createAgentSession”, so even blank B/C/D sent reasoning_effort → max_completion_tokens:1. Changed to leave “” as undefined so pi uses its per-model default then clamps via model.thinkingLevelMap (off for reasoning:false locals). Fixed in daemon/worker, daemon/index, and core/modelRouter clampThinkingForProvider (now model-aware with provider-string fallback).
+- **Provider-string clamp was model-blind.** ”lm-studio → medium” broke xhigh-capable locals and still sent medium to 2b QAD models. Replaced with pi-ai getSupportedThinkingLevels semantics: reasoning:false → only off; reasoning:true + empty map → off…high (no xhigh/max); map null → filtered. Config file stays on pi’s THINKING_LEVELS definition, nothing invented.
+
 ## [2.8.9] — 2026-08-31
 
 Locals no longer die on xhigh; DEFINE ladder re-consults.