world-model-optimizer v0.2.0

First release under the world-model-optimizer name, published to PyPI as
world-model-optimizer.

pip install world-model-optimizer
wmo --help

Breaking

The project was renamed from world-model-harness, and the rename goes all the way down. If you
have an existing install, note:

	before	after
Package	world-model-harness	world-model-optimizer
Command	wmh	wmo
Import	import wmh	import wmo
State dir	.wmh/	.wmo/
Environment	WMH_* (46 vars)	WMO_*
OTel attributes	wmh.*	wmo.*

There is no compatibility shim. Uninstall the old distribution first — both bundle the same
llm_waterfall module, so side-by-side installs conflict:

pip uninstall world-model-harness
pip install world-model-optimizer

Also renamed: renderer registry keys, HuggingFace dataset repo ids, the E2B template alias,
W&B project names, and harbor agent import paths. Existing harbor job directories are not
resumable across this change, and run configs naming the old renderer keys need updating —
both fail loudly rather than silently.

world-model-harness 0.1.0 remains on PyPI and is not being updated.

world-model-harness v0.1.0

What's Changed

• tau-bench RL: pinned shared train/eval scenario sets (all six arms) by @SilenNaihin in #65
• cli: show help when invoked without a command by @SilenNaihin in #64
• ICL arm: policy agent + critique memory vs the tau WM; OpenAI provider custom endpoints by @SilenNaihin in #67
• skills: ready-for-merge never merges — hands the PR back to the user by @SilenNaihin in #68
• Repo layout conformance: allowlist rule 5, .agents/ workspace, executable enforcement by @SilenNaihin in #60
• skills: ready-for-merge scales code-review effort to the PR by @SilenNaihin in #76
• cli: invalid model/example names get friendly errors, not tracebacks by @SilenNaihin in #75
• cli: dash-join whitespace in model names instead of rejecting by @SilenNaihin in #77
• skills: ready-for-merge polls reviewers and scopes the test gate by @SilenNaihin in #80
• cli: arrow-key pickers + credential-aware provider selection by @SilenNaihin in #79
• cli: provider SDKs install with plain uv sync; wizard drops openai_responses by @SilenNaihin in #84
• providers: reasoning-safe verify ping budget by @SilenNaihin in #85
• wizard: live-verify provider and embedder inline, right after the model pick by @SilenNaihin in #86
• build: honest GEPA progress totals + capped selection valset by @SilenNaihin in #88
• build reporter: the GEPA bar never pins at 100% while calls keep landing by @SilenNaihin in #89
• build reporter: bar snaps to 100% on the completion event, stays visible by @SilenNaihin in #90
• wizard: creds annotation names the env var it found by @SilenNaihin in #91
• build: live GEPA activity window + default budget 50 -> 10 by @SilenNaihin in #92
• gepa: stream per-step activity immediately + parallel batch evaluation by @SilenNaihin in #93
• build: separate GEPA judge model with cheap per-provider defaults by @SilenNaihin in #94
• gepa: narrate every phase of the loop in the activity window by @SilenNaihin in #95
• build: stop orphaned activity-window headers; cap GEPA valset at 16 by @SilenNaihin in #102
• The monorepo PR by @SilenNaihin in #100
• wizard: collapsed pickers (2 visible + '… N more') + batched-key handling by @SilenNaihin in #103
• build: width-safe activity lines; true train/val/test split by @SilenNaihin in #104
• cli: demo replays sampled scenarios open-loop; play suggests real actions by @SilenNaihin in #106
• cli: narrated provider retries + switch-provider flow in demo by @SilenNaihin in #107
• cli: instant first retry, deduped hiccups, inline countdown + spinner by @SilenNaihin in #108
• demo: stream every step live, show the env prompt, resume after switch by @SilenNaihin in #109
• Pluggable trace ingestion: observability providers + chat/tool-call conversion by @SilenNaihin in #33
• Monorepo members move under packages/ by @SilenNaihin in #115
• environment-capture: benchmark adapter contract + 4 real-benchmark trace corpora + WM-replacement seam (WS-B2) by @SilenNaihin in #56
• environment-capture: PyPI-ready packaging + pure-package README by @SilenNaihin in #122
• Scenario-set construction from traces: Clio facets -> cluster -> select -> synthesize -> verify by @kfallah in #81
• Follow-up to #81: remove dupe-file slop, restore research results, package wmh/scenarios, README for the scenarios CLI by @kfallah in #125
• Closed-loop eval: wmh eval --mode closed-loop + the wmh/evals package by @kfallah in #111
• Harness artifacts: versioned HarnessDoc store under .wmh/harnesses// by @kfallah in #112
• wmh harness create: delta search over harness-space against the world model by @kfallah in #113
• llm-waterfall: preserve jitter at the backoff cap; fix stale azure comment by @SilenNaihin in #105
• providers: WaterfallProvider backed by llm-waterfall (failover chains) by @SilenNaihin in #51
• Adopt llm-waterfall classification in research scripts; classify terminal-capture retries by @SilenNaihin in #129
• Harden corpus hygiene against credential leaks in capture by @devin-ai-integration[bot] in #130
• Vendored pi + RunnerLink frame transport (no E2B) by @kfallah in #134
• wmh login/logout/status + push/pull against the hosted platform by @kfallah in #132
• judge: collapse to a single judge + principled overhaul, proven by a judge-quality meta-eval by @SilenNaihin in #83
• Website: world-model gallery + playground + build-your-own (WS-B1) by @SilenNaihin in #52
• Platform registry client: org-scoped workspaces (lockstep with platform #276) by @kfallah in #138
• Run the real pi harness in parallel E2B sandboxes against the simulated world model by @kfallah in #135
• Standardize provider-independent model types by @kfallah in #144
• Live agent sessions: interactive multi-turn pi runner + host engine by @kfallah in #140
• Optimizer: 16k proposer replies + narrate every skipped iteration by @kfallah in #143
• Support Azure models in pi harness runs by @kfallah in #142
• Fix live_session for the provider-agnostic worker API (unbreak main) by @kfallah in #145
• Never send AZURE_OPENAI_API_KEY to a config-controlled endpoint by @devin-ai-integration[bot] in #146
• Restrict server builds to managed trace uploads by @devin-ai-integration[bot] in #147
• Run platform agents in E2B with local workspace sync by @kfallah in #150
• Preserve tool calls through provider retries by @kfallah in #153
• Retry E2B transport timeouts safely by @kfallah in #151
• Use canonical production domain for login by @kfallah in #156
• Resolve chat token parameters by model by @kfallah in #155
• Concurrency scaling law: WM-vs-sandbox differential (rebuilt on wmh eval) by @SilenNaihin in #41
• Trace scaling law + RAG optimization (RAG-only): findings, figures, reproducible methodology by @SilenNaihin in #72
• GEPA scaling law: fidelity vs iterations × traces, + judge-sensitivity ablation by @SilenNaihin in #97
• World model as an agent: editable knowledge base, reasoning pass, web grounding (WS-A3, roadmap f) by @SilenNaihin in #55
• Verbalized confidence: contract field + gated verify (WS-A6, stacked on #55) by @SilenNaihin in #120
• Benchmark grid: 5 models × 4 conditions × 4 benchmarks (wmh eval grid + ...