Understudy Desktop v0.3.41

Understudy Desktop 0.3.41 bundles CLI 0.6.38 and brings the benchmark lab into the app.

Highlights

• Experiment lineage cards — every result in the app now traces back to the run, dataset, and prompt that produced it (understudy.experiment.v1 sidecar).
• Upload approval gate — the drag-drop training flow asks before anything leaves your machine; nothing uploads without an explicit yes.
• Benchmark linkage pane — see which benchmark and split each experiment was scored against, right next to the result.
• Pi chat benchmark tools — the embedded Pi chat is wired to the benchmark/experiment spine, and the new operate-benchmark-lab skill drives the whole lab conversationally.
• Sidebar cleanup — New chat unnested to the top level; clipping, overlap, and leftover artifacts fixed.
• Smart defaults in the bundled CLI — a one-time trust posture replaces per-call approval prompts.
• The full benchmark CLI spine — benchmarks from-dataset turns labeled datasets into first-class benchmarks; the partner report produces the honest benchmark-and-savings deliverable; trivial-agent calibration arms, majority-class floors, and the ABC rigor report keep scores honest; GEPA-style prompt evolution runs over prompt-override experiment arms.

Existing installs update in place via the Tauri updater (latest.json on this release).

What's Changed

• Add local trace benchmark foundry by @lluisinthedesert in #314
• Benchmark hub: benchmark.v1 spine, local hub/leaderboard viewer, real dogfood data, correctness + hardening + tests by @lluisinthedesert in #315
• Benchmark hub: full foundry unification — lifecycle-aware viewer, one review store, one styling system, designed empty states by @lluisinthedesert in #317
• Complete trace foundry runtime and environment workflow by @lluisinthedesert in #316
• Add reusable local trace viewer by @aamir in #318
• Foundry fixes: cut Prime push (privacy), batch truncation, semantic gold, promote verb, schema-collision resolution by @lluisinthedesert in #319
• Roadmap wave: trace-aware grouping, judgeable contracts, run queue + executor, narrative layout by @lluisinthedesert in #320
• Automation-focus round: accepted-task runs, statement-first task page, regenerate-env + environment fixes by @lluisinthedesert in #321
• Live rollout watching + fair anchor scoring + world validation + judgeability guarantee by @lluisinthedesert in #322
• Fix live-journal newlines (empty live view during real rollouts) by @lluisinthedesert in #323
• Observation-tightened world-model validation by @lluisinthedesert in #324
• Unify benchmark artifact producer/consumer contracts behind one shared codec by @lluisinthedesert in #326
• Add ABC benchmark-rigor checkpoints to eval skills by @lluisinthedesert in #330
• docs: benchmark-hub → desktop migration scoping by @lluisinthedesert in #329
• Add instrument skill: zero-friction trace-capture on-ramp by @lluisinthedesert in #331
• Incumbent baseline: record the capture-producing model, rerun it as a labeled arm, gate tasks on calibration by @lluisinthedesert in #325
• fix(skills): add missing Resolve CLI section to instrument skill (unbreak main gates) by @lluisinthedesert in #332
• Sentinel gates against the silent-zero failure class by @lluisinthedesert in #328
• Agent-operator MCP surface: understudy benchmarks mcp by @lluisinthedesert in #327
• Trivial-agent calibration arms (null/spam) + ABC rigor report by @lluisinthedesert in #333
• Statistical honesty + environment integrity: bootstrap CIs, gold-leakage audit, rollout isolation pin by @lluisinthedesert in #334
• Exception-based benchmark review: auto-accept policy + conversational task edit by @lluisinthedesert in #335
• Fuzzy tier for the gold-leakage audit (shingles + fingerprints) by @lluisinthedesert in #336
• Surface follow-ups: floors rendering, MCP policy tools, configurable auto-accept by @lluisinthedesert in #337
• Oracle full-contract coverage: response obligations verified against stored gold by @lluisinthedesert in #338
• Harden benchmark run queue/executor: claims + capability gate, per-rollout timeout, structural output isolation by @lluisinthedesert in #339
• Prompt-override experiment arms (run-scoped SOP suffix, capability-gated) by @lluisinthedesert in #340
• Fixtures-state-split: stop serving gold post-state to candidates by @lluisinthedesert in #342
• App-replay arm (tier 1): run the user's own app on frozen benchmark tasks by @lluisinthedesert in #341
• benchmark hub: multi-objective (Pareto) trade-offs view by @lluisinthedesert in #343
• Skills: operate the full benchmark-lab lifecycle (MCP + CLI + daemons + desktop distribution) by @lluisinthedesert in #344
• Local trained-artifact arms + majority-class floor for the benchmark run executor by @lluisinthedesert in #345
• GEPA-style prompt evolution driving prompt_overrides arms (benchmarks evolve) by @lluisinthedesert in #347
• Rejection guidance as an optimizable surface: guidance.json, recovery metric, --guidance override by @lluisinthedesert in #348
• Experiment-lineage layer: understudy.experiment.v1 sidecar + CLI/MCP surfaces by @lluisinthedesert in #346
• Fix benchmark-hub visual-QA findings (hydration, floors, feedback ids, pills, dev preflight) by @lluisinthedesert in #350
• Born-accepted task inbox: proposed review flow + page redesign by @lluisinthedesert in #349
• OpenAI response-shaped workload fixes (warp-domain-identification generality validation) by @lluisinthedesert in #351
• Merge-train tooling: gated PR train, wave-ownership checker, retro doc by @lluisinthedesert in #352
• Wire the embedded Pi chat to the benchmark/experiment spine by @lluisinthedesert in #353
• benchmarks from-dataset: labeled datasets become first-class benchmarks by @lluisinthedesert in #354
• Desktop app → benchmark/experiment spine bridge (lineage, approval gate, benchmark linkage) by @lluisinthedesert in #355
• Partner report: the honest benchmark-and-savings deliverable by @lluisinthedesert in #356
• Smart defaults: one-time trust posture replaces per-call approval gates by @lluisinthedesert in #357
• Desktop sidebar cleanup: unnest New chat, fix clipping, remove artifacts by @lluisinthedesert in #358
• release: Desktop 0.3.41 / CLI 0.6.38 by @lluisinthedesert in #359

Full Changelog: desktop-v0.3.40-mvp...desktop-v0.3.41-mvp