v0.1.11.dev0

Verifiers v0.1.11.dev0 Release Notes

Date: 02/14/2026

Full Changelog: v0.1.10...v0.1.11.dev0

v0.1.10

Verifiers v0.1.10 Release Notes

Date: 02/10/2026

Full Changelog: v0.1.9...v0.1.10

Highlights since v0.1.9

• Expanded environment support with OpenEnv and BrowserEnv integrations, env worker plumbing, and continued improvements to CliAgentEnv/RLMEnv sandbox reliability and customization hooks.
• Upgraded evaluation ergonomics with resumed evals, improved TUI info/log presentation, better rollout/token tracking, and non-TUI overflow rendering fixes.
• Improved reliability across model/runtime boundaries with timeout hardening, safer sandbox lifecycle behavior, and richer error/metadata handling.
• Modernized workspace setup and contributor workflows (vf-setup endpoint/config updates, GEPA config support, Prime CLI refactor, skills scaffolding, and AGENTS guidance updates).
• Added opencode harbor enhancements, including TITO support, tunnel sync stop behavior, and terminal-bench task coverage.

Changes included in v0.1.10 (since v0.1.9)

Environment, rollout, and runtime improvements

• openenv integration (#829)
• Add Browser Env Integration (#732)
• resume evals (#803)
• add Client Pool (#815)
• RLM: Eager sandbox creation, conditional pip install (#834)
• RLM: Add RLMEnv sandbox hooks for safer customization (#849)
• RLM: Make FIFO IO non-blocking (#850)
• CliAgentEnv: add SandboxMixin, refactor InterceptionServer (#847)
• rlm: migrate sandbox executor to SandboxMixin (#875)
• env worker integration (#832)
• track vf + env version in metadata (#881)
• handle empty metrics (#855)
• move sanitize_metadata out of save_metadata (#852)
• improve env client timeouts (#872)
• Fix ty logger protocol typing in sandbox retry setup (#835)
• Fix vf-eval concurrent rollout label to use effective cap (#836)

Evaluation UX, logging, and metrics

• Add robust token usage tracking (#858)
• Tighten vf-tui info preview formatting and typing checks (#830)
• Add subtle --debug hint beneath Logs panel (#824)
• Fix vf-eval non-TUI live overflow rendering (#883)
• misc logging improvs (#882)

Setup, CLI, and configuration

• vf-setup: prefer endpoints.toml, rename configs/lab->configs/rl, add GEPA configs, deprecate --vf-rl (#859)
• Support long endpoint field names in TOML registries (#861)
• prime CLI refactor (vf) (#870)
• refactor: split RL trainer into optional in-repo verifiers-rl package (#843)
• move rlm secrets out of vf and into research-environments (#856)

Documentation, workflows, and skills

• Compile AGENTS docs from modular assets and make guidance concrete (#857)
• skills setup (#873)
• Strengthen lab AGENTS env-development guardrails (#876)
• Clarify MCPEnv is for global read-only MCP servers (#838)
• docs: remove parser-centric guidance from environment READMEs (#839)
• docs: remove parser field from env init README template (#840)
• chore: enforce ruff formatting and improve dev tooling docs (#845)

Integrations and environment packages

• openenv: default template proj/ path and simplify prompt renderer signatures (#853)
• remove vf pin in opencode_harbor (#844)
• opencode env: TITO support, tunnel sync stop, add terminal-bench (#874)
• ci: skip terminus_harbor in test-envs (#846)
• fix math rubric timeouts (#831)
• Fix for dir resolutions (#879)
• Update browse-environments freshness and quality priorities (#884)
• Clarify agent skill handling (#886)

v0.1.10.dev5

Verifiers v0.1.10.dev5 Release Notes

Date: 02/10/2026

Full Changelog: v0.1.10.dev4...v0.1.10.dev5

Highlights since v0.1.10.dev4

• Improved environment worker reliability and metadata tracking by migrating sandbox execution internals and recording verifiers/environment versions in rollout metadata.
• Polished evaluation UX with a fix for non-TUI live overflow rendering and broader logging improvements.
• Added and refined developer-facing guidance for skills and agent workflows, including clearer skill handling and browse-environment quality priorities.
• Extended opencode harbor support with TITO support, tunnel sync stop behavior, and a terminal-bench task addition.

Incremental changes since v0.1.10.dev4

• env worker integration (#832)
• track vf + env version in metadata (#881)
• misc logging improvs (#882)
• rlm: migrate sandbox executor to SandboxMixin (#875)
• opencode env: TITO support, tunnel sync stop, add terminal-bench (#874)
• Fix vf-eval non-TUI live overflow rendering (#883)
• Fix for dir resolutions (#879)
• Update browse-environments freshness and quality priorities (#884)
• Clarify agent skill handling (#886)

v0.1.10.dev4

Verifiers v0.1.10.dev4 Release Notes

Date: 02/09/2026

Full Changelog: v0.1.10.dev3...v0.1.10.dev4

Highlights since v0.1.10.dev3

• Refactored the vf CLI command surface around Prime CLI integration and plugin wiring.
• Improved environment client timeout handling across worker and ZMQ clients, with regression coverage in timeout tests.
• Added skills setup scaffolding and related docs updates for Prime Lab setup and workflow guidance.
• Tightened AGENTS guidance for environment development guardrails in Lab-facing docs.

Incremental changes since v0.1.10.dev3

• improve env client timeouts (#872)
• skills setup (#873)
• Strengthen lab AGENTS env-development guardrails (#876)
• prime CLI refactor (vf) (#870)

v0.1.10.dev3

Verifiers v0.1.10.dev3 Release Notes

Date: 02/08/2026

Full Changelog: v0.1.9...v0.1.10.dev3

Highlights since v0.1.9

• Added new environment capabilities, including OpenEnv integration, BrowserEnv integration, and env server support for more flexible tool and environment workflows.
• Expanded evaluation UX with eval TUI, copy mode, improved logs/debug display, rollout token usage tracking, and richer saved-output rendering for tool calls.
• Introduced and iterated on RLMEnv improvements: tool partitioning (tools, root_tools, sub_tools), better stop/error propagation, prompt/verbosity controls, safer sandbox lifecycle handling, and new sandbox hooks for customization.
• Improved reliability across execution and infrastructure paths via retries for infrastructure and model-response errors, better auth/overlong prompt handling for OpenRouter, and cleanup fixes to avoid task/sandbox leakage.
• Modernized setup and training ergonomics with vf-setup config changes (endpoints.toml, configs/rl, GEPA configs), support for long TOML endpoint field names, and an optional in-repo verifiers-rl package split.
• Hardened runtime internals with CliAgentEnv sandbox/interception refactors, client pooling, non-blocking FIFO IO for RLM, and metadata/metrics handling fixes.
• Added broader OpenEnv ecosystem support and examples (e.g., openenv_echo, openenv_textarena, opencode_harbor) with updated version requirements.

Incremental changes since v0.1.10.dev2

• Compile AGENTS docs from modular assets and make guidance concrete (#857)
• vf-setup: prefer endpoints.toml, rename configs/lab->configs/rl, add GEPA configs, deprecate --vf-rl (#859)
• Support long endpoint field names in TOML registries (#861)
• Add robust token usage tracking (#858)
• move rlm secrets out of vf and into research-environments (#856)
• CliAgentEnv: add SandboxMixin, refactor InterceptionServer (#847)
• handle empty metrics (#855)
• move sanitize_metadata out of save_metadata (#852)
• openenv: default template proj/ path and simplify prompt renderer signatures (#853)
• refactor: split RL trainer into optional in-repo verifiers-rl package (#843)
• add Client Pool (#815)
• chore: enforce ruff formatting and improve dev tooling docs (#845)
• RLM: Make FIFO IO non-blocking (#850)
• RLM: Add RLMEnv sandbox hooks for safer customization (#849)
• RLM: Eager sandbox creation, conditional pip install (#834)
• ci: skip terminus_harbor in test-envs (#846)
• resume evals (#803)
• remove vf pin in opencode_harbor (#844)
• fix math rubric timeouts (#831)
• docs: remove parser-centric guidance from environment READMEs (#839)
• openenv integration (#829)
• Fix ty logger protocol typing in sandbox retry setup (#835)
• docs: remove parser field from env init README template (#840)
• Clarify MCPEnv is for global read-only MCP servers (#838)
• Fix vf-eval concurrent rollout label to use effective cap (#836)
• Tighten vf-tui info preview formatting and typing checks (#830)
• Add subtle --debug hint beneath Logs panel (#824)

v0.1.10.dev2

Verifiers v0.1.10.dev1 Release Notes

Date: 02/04/2026

Full Changelog: v0.1.10.dev0...v0.1.10.dev1

Changes since v0.1.10.dev0

• info oaitools fix (#821)
• Capture stdout/stderr for live display (#819)
• track token usage in eval (#816)
• RLM: show full user message (#818)
• remove filesystem info from rlm system prompts (#817)
• RLM: add prompt verbosity parameters (#814)
• re-raise auth errs + fix overlong prompt err for openrouter (#813)
• add default sandbox_labels to rlm-secrets (#810)
• Improve vf-eval display (#809)
• Increase Sandbox Default Thread Worker Count (#807)
• Tool content validation (#806)
• env server (#799)
• Add Browser Env Integration (#732)
• adjust CliAgentEnv sandbox creation timeout + remove DummyHarborEnv (#804)
• verifiers: fix tool call rendering from saved outputs (#802)
• RLM: remove code jail -> simplify code (#800)
• Add sync bulk sandbox teardown for RLM env (#798)
• overhaul saving outputs (#774)
• Propagate RLM stop errors from root and sub tools (#797)
• clean up on task cancelation to avoid resource leakage (#795)
• CliAgentEnv: teardown sandboxes via bulk delete (#796)
• RLM: Fix trajectory collision (#786)
• lazy import datasets (#794)
• add rLLM integration to docs
• cancel outstanding tasks if one task raises in generate (#793)
• revert wiki-search
• update environments/README.md (#790)
• util for enforcing env vars are set (#789)
• return last result if retries exhausted (#782)
• warning log in RLTrainer (#783)
• RLM: Simplify code (#781)
• hello world tasks for TerminusHarborEnv and OpenCodeHarborEnv (#775)
• fix retry for invalid model response errors (#778)
• Make local RLM REPL concurrency configurable (#777)
• RLM: re-enable Sandboxes for both Python and Bash (#776)
• fix alphabet-sort
• fix alphabet-sort
• raise on empty response error (openrouter) to trigger retries (#772)
• mirror cli in toml config (#773)
• fix double save results in vf-eval (#771)
• remove log file for cli agent env (#770)
• eval --debug mode to skip Rich (#769)
• tools eval example
• Harbor examples (#766)
• Feature: Add tools metadata for eval viewer (#767)
• expose sandbox labels in SandboxEnv and CliAgentEnv (#768)
• RLM env stop condition fix (#757)
• lazy init locked chromadb instance in wiki-search (#765)
• created the rlm_secrets environment (#763)
• Move RLM system prompt into first user prompt (#764)
• gepa dep
• integrated gepa training, ui to track (#747)
• prime tunnel in cliagentenv (#746)
• RLM: make bash REPL default, keep Python REPL optional (#758)
• Sebastian/rlm file system 2026 01 20 (#756)
• eval tui (#735)
• multi-env evals config (#734)
• Add retry support for infrastructure errors in vf-eval (#750)
• RLM: tools, sub_tools, root_tools (#749)
• optional DatasetBuilder pattern (#739)
• Add copy mode to vf-tui (#745)
• RLMEnv: Make sub-LLM calls work for training (#738)

v0.1.10.dev1

Verifiers v0.1.10.dev1 Release Notes

Date: 02/04/2026

Full Changelog: v0.1.10.dev0...v0.1.10.dev1

Changes since v0.1.10.dev0

• info oaitools fix (#821)
• Capture stdout/stderr for live display (#819)
• track token usage in eval (#816)
• RLM: show full user message (#818)
• remove filesystem info from rlm system prompts (#817)
• RLM: add prompt verbosity parameters (#814)
• re-raise auth errs + fix overlong prompt err for openrouter (#813)
• add default sandbox_labels to rlm-secrets (#810)
• Improve vf-eval display (#809)
• Increase Sandbox Default Thread Worker Count (#807)
• Tool content validation (#806)
• env server (#799)
• Add Browser Env Integration (#732)
• adjust CliAgentEnv sandbox creation timeout + remove DummyHarborEnv (#804)
• verifiers: fix tool call rendering from saved outputs (#802)
• RLM: remove code jail -> simplify code (#800)
• Add sync bulk sandbox teardown for RLM env (#798)
• overhaul saving outputs (#774)
• Propagate RLM stop errors from root and sub tools (#797)
• clean up on task cancelation to avoid resource leakage (#795)
• CliAgentEnv: teardown sandboxes via bulk delete (#796)
• RLM: Fix trajectory collision (#786)
• lazy import datasets (#794)
• add rLLM integration to docs
• cancel outstanding tasks if one task raises in generate (#793)
• revert wiki-search
• update environments/README.md (#790)
• util for enforcing env vars are set (#789)
• return last result if retries exhausted (#782)
• warning log in RLTrainer (#783)
• RLM: Simplify code (#781)
• hello world tasks for TerminusHarborEnv and OpenCodeHarborEnv (#775)
• fix retry for invalid model response errors (#778)
• Make local RLM REPL concurrency configurable (#777)
• RLM: re-enable Sandboxes for both Python and Bash (#776)
• fix alphabet-sort
• fix alphabet-sort
• raise on empty response error (openrouter) to trigger retries (#772)
• mirror cli in toml config (#773)
• fix double save results in vf-eval (#771)
• remove log file for cli agent env (#770)
• eval --debug mode to skip Rich (#769)
• tools eval example
• Harbor examples (#766)
• Feature: Add tools metadata for eval viewer (#767)
• expose sandbox labels in SandboxEnv and CliAgentEnv (#768)
• RLM env stop condition fix (#757)
• lazy init locked chromadb instance in wiki-search (#765)
• created the rlm_secrets environment (#763)
• Move RLM system prompt into first user prompt (#764)
• gepa dep
• integrated gepa training, ui to track (#747)
• prime tunnel in cliagentenv (#746)
• RLM: make bash REPL default, keep Python REPL optional (#758)
• Sebastian/rlm file system 2026 01 20 (#756)
• eval tui (#735)
• multi-env evals config (#734)
• Add retry support for infrastructure errors in vf-eval (#750)
• RLM: tools, sub_tools, root_tools (#749)
• optional DatasetBuilder pattern (#739)
• Add copy mode to vf-tui (#745)
• RLMEnv: Make sub-LLM calls work for training (#738)

v0.1.10.dev0

Verifiers v0.1.10.dev0 Release Notes

Date: 01/17/2026

Full Changelog: v0.1.9.post3...v0.1.10.dev0

v0.1.9.post3

Verifiers v0.1.9 Release Notes

Date: 01/08/2026

Verifiers v0.1.9 introduces several new experimental environments, monitor rubrics for automatic metrics collection, improved error handling, and documentation overhaul.

Post-release update:

• Tweaks to setup script (post1).
• Fix for exporting setup script (post0).
• Fix for gitignore section in setup script (post2).
• Tweaks to setup script and endpoint defaults (post3).

Highlights

• RLMEnv (Experimental): New environment implementing the Recursive Language Model (RLM) inference strategy, where language models decompose and recursively interact with input data through sandboxed REPL environments. Supports sub-LLM calls via llm_batch() function intercepted through HTTP proxy. See RLM paper.

• GymEnv (Experimental): Universal Gym-compatible environment runner for standard RL gymnasium environments. Enables training on classic control tasks and custom Gym environments.

• CliAgentEnv & HarborEnv (Experimental): New environments for running custom agent code in sandboxes (CliAgentEnv) and loading Harbor-format tasks (HarborEnv).

• MCPEnv (Experimental): Environment for Model Context Protocol (MCP) server integration.

• Monitor Rubrics: Each environment now automatically includes a monitor rubric that tracks environment-specific metrics without affecting rewards (weight=0). For example:

  • MultiTurnEnv: num_turns
  • ToolEnv: tool_call_count
  • SandboxEnv: sandbox_call_count, sandbox_total_time_seconds, sandbox_mean_time_seconds
  • PythonEnv: repl_call_count, repl_total_time_seconds, repl_mean_time_seconds
  • RLMEnv: Sub-LLM metrics and more

• Improved Error Handling: New error chain helpers, better error propagation through rollouts, and abort_on_code_timeout support for sandbox environments.

• Documentation Overhaul: Complete reorganization of documentation with new Mintlify-based docs, improved examples, and automatic docs sync workflow.

New Features

Environments

• Add get_sandbox_request hook for per-rollout sandbox customization (#699)
• Expose render_completion and add_trajectory_step methods with private/final guardrails (#679)
• Add final_messages pattern for cleaner message handling (#677)
• Support for token-in vLLM endpoint (#626)
• Static make_dataset function for environments (#683)
• Add alphabet-sort example environment (#695)
• system_prompt is now prepended to existing prompts that don't already start with a system message

Evaluation & Training

• Optionally enable independent per-rollout scoring: run and score rollouts independently rather than only in groups (#694)
• vf-tui improvements: regex search modal and run details panel (#705)
• Log eventloop lag during vf-eval (#687)
• Log timings in vf-eval (#686)
• Show rolling average as tqdm postfix (#693)
• Option to bypass scoring for faster iteration (#645)
• Add trajectory_id to TrajectoryStep (#675)

Rubrics

• Add RLM monitor rubric for sub-LLM metrics (#698)
• Improvements to math rubric with better timeout handling (#657)
• JudgeRubric now accepts optional state argument (#684)

Error Handling

• Helpers for error chains (#649)
• Better error handling with abort_on_code_timeout (#659)
• Handle all truncation cases (#637)
• Raise ModelError when response.choices is None (#640)
• Apply stop_errors pattern to StatefulToolEnv for parse/call errors (#618)
• Normalize messages from sub-LLM calls to prevent errors (#664)

Bug Fixes

• Fix tool duplication when calling add_tool on ToolEnv with shared list reference
• Fix args_to_skip validation failure for dict type parameters in StatefulToolEnv (#674)
• Fix empty slice handling (#701)
• Fix wiki-search environment (#697)
• Fix tool test environment (#692)
• Fix PythonEnv deadlock (#652)
• Fix auto-format dataset for message_type=completions (#624)
• Fix math verify timeout (#620)
• Fix sub-LLM metrics and context warnings
• pip_install_packages="" no longer breaks sandbox (#633)
• Remove prompt logprobs to reduce memory usage (#666)
• Warn when ignoring system prompt/few-shot with prompt present (#668)
• Handle empty completions in parse_answer (#672)

Infrastructure & Documentation

• Ensure integrations can be installed via full path (#704)
• Reorganize third-party env integrations (TextArena, ReasoningGym, etc.) (#682)
• Experimental folder structure for newer environments (#643)
• Overhaul docs with example configs (#700)
• Update docs for v0.1.8 API (#670)
• Add automatic docs sync workflow (#628)
• Redirect RTD to shared Mintlify docs (#654)
• Dynamic logger names (#639)
• Use threading for sandbox client (#638)
• Bump prime-sandboxes>=2.7.0 (#660)

vf-setup Command

The vf-setup command bootstraps a verifiers training workspace:

Default behavior (no flags):

• Creates configs/ and environments/ directories
• Downloads AGENTS.md, CLAUDE.md, and environments/AGENTS.md for AI coding assistants
• Downloads configs/endpoints.py (API endpoint configuration)
• Downloads lab configs for quick experimentation (configs/lab/*.toml)

With --prime-rl:

• Installs prime-rl and syncs dependencies
• Installs all environments from environments/ into the prime-rl workspace
• Downloads prime-rl-specific configs to configs/prime-rl/

With --vf-rl:

• Downloads configs/zero3.yaml (DeepSpeed config)
• Downloads vf-rl configs to configs/vf-rl/

With --skip-agents-md:

• Skips downloading AGENTS.md, CLAUDE.md, and environments/AGENTS.md

Migration Notes

• Environments now automatically include monitor rubrics which track default class-specific metrics. If you were manually adding metrics for num_turns, tool_call_count, etc., these are now provided automatically.
• Third-party integrations (TextArena, ReasoningGym) have been moved to verifiers.envs.integrations.
• Experimental environments (GymEnv, MCPEnv, CliAgentEnv, HarborEnv, RLMEnv) are now in verifiers.envs.experimental and require explicit imports or verifiers[all] installation.

Full Changelog: v0.1.8.post2...v0.1.9

v0.1.9.post2

Verifiers v0.1.9 Release Notes

Date: 01/08/2026

Verifiers v0.1.9 introduces several new experimental environments, monitor rubrics for automatic metrics collection, improved error handling, and documentation overhaul.

Post-release update:

• Tweaks to setup script (post1).
• Fix for exporting setup script (post0).
• Fix for gitignore section in setup script (post2).

Highlights

• RLMEnv (Experimental): New environment implementing the Recursive Language Model (RLM) inference strategy, where language models decompose and recursively interact with input data through sandboxed REPL environments. Supports sub-LLM calls via llm_batch() function intercepted through HTTP proxy. See RLM paper.

• GymEnv (Experimental): Universal Gym-compatible environment runner for standard RL gymnasium environments. Enables training on classic control tasks and custom Gym environments.

• CliAgentEnv & HarborEnv (Experimental): New environments for running custom agent code in sandboxes (CliAgentEnv) and loading Harbor-format tasks (HarborEnv).

• MCPEnv (Experimental): Environment for Model Context Protocol (MCP) server integration.

• Monitor Rubrics: Each environment now automatically includes a monitor rubric that tracks environment-specific metrics without affecting rewards (weight=0). For example:

  • MultiTurnEnv: num_turns
  • ToolEnv: tool_call_count
  • SandboxEnv: sandbox_call_count, sandbox_total_time_seconds, sandbox_mean_time_seconds
  • PythonEnv: repl_call_count, repl_total_time_seconds, repl_mean_time_seconds
  • RLMEnv: Sub-LLM metrics and more

• Improved Error Handling: New error chain helpers, better error propagation through rollouts, and abort_on_code_timeout support for sandbox environments.

• Documentation Overhaul: Complete reorganization of documentation with new Mintlify-based docs, improved examples, and automatic docs sync workflow.

New Features

Environments

• Add get_sandbox_request hook for per-rollout sandbox customization (#699)
• Expose render_completion and add_trajectory_step methods with private/final guardrails (#679)
• Add final_messages pattern for cleaner message handling (#677)
• Support for token-in vLLM endpoint (#626)
• Static make_dataset function for environments (#683)
• Add alphabet-sort example environment (#695)
• system_prompt is now prepended to existing prompts that don't already start with a system message

Evaluation & Training

• Optionally enable independent per-rollout scoring: run and score rollouts independently rather than only in groups (#694)
• vf-tui improvements: regex search modal and run details panel (#705)
• Log eventloop lag during vf-eval (#687)
• Log timings in vf-eval (#686)
• Show rolling average as tqdm postfix (#693)
• Option to bypass scoring for faster iteration (#645)
• Add trajectory_id to TrajectoryStep (#675)

Rubrics

• Add RLM monitor rubric for sub-LLM metrics (#698)
• Improvements to math rubric with better timeout handling (#657)
• JudgeRubric now accepts optional state argument (#684)

Error Handling

• Helpers for error chains (#649)
• Better error handling with abort_on_code_timeout (#659)
• Handle all truncation cases (#637)
• Raise ModelError when response.choices is None (#640)
• Apply stop_errors pattern to StatefulToolEnv for parse/call errors (#618)
• Normalize messages from sub-LLM calls to prevent errors (#664)

Bug Fixes

• Fix tool duplication when calling add_tool on ToolEnv with shared list reference
• Fix args_to_skip validation failure for dict type parameters in StatefulToolEnv (#674)
• Fix empty slice handling (#701)
• Fix wiki-search environment (#697)
• Fix tool test environment (#692)
• Fix PythonEnv deadlock (#652)
• Fix auto-format dataset for message_type=completions (#624)
• Fix math verify timeout (#620)
• Fix sub-LLM metrics and context warnings
• pip_install_packages="" no longer breaks sandbox (#633)
• Remove prompt logprobs to reduce memory usage (#666)
• Warn when ignoring system prompt/few-shot with prompt present (#668)
• Handle empty completions in parse_answer (#672)

Infrastructure & Documentation

• Ensure integrations can be installed via full path (#704)
• Reorganize third-party env integrations (TextArena, ReasoningGym, etc.) (#682)
• Experimental folder structure for newer environments (#643)
• Overhaul docs with example configs (#700)
• Update docs for v0.1.8 API (#670)
• Add automatic docs sync workflow (#628)
• Redirect RTD to shared Mintlify docs (#654)
• Dynamic logger names (#639)
• Use threading for sandbox client (#638)
• Bump prime-sandboxes>=2.7.0 (#660)

vf-setup Command

The vf-setup command bootstraps a verifiers training workspace:

Default behavior (no flags):

• Creates configs/ and environments/ directories
• Downloads AGENTS.md, CLAUDE.md, and environments/AGENTS.md for AI coding assistants
• Downloads configs/endpoints.py (API endpoint configuration)
• Downloads lab configs for quick experimentation (configs/lab/*.toml)

With --prime-rl:

• Installs prime-rl and syncs dependencies
• Installs all environments from environments/ into the prime-rl workspace
• Downloads prime-rl-specific configs to configs/prime-rl/

With --vf-rl:

• Downloads configs/zero3.yaml (DeepSpeed config)
• Downloads vf-rl configs to configs/vf-rl/

With --skip-agents-md:

• Skips downloading AGENTS.md, CLAUDE.md, and environments/AGENTS.md

Migration Notes

• Environments now automatically include monitor rubrics which track default class-specific metrics. If you were manually adding metrics for num_turns, tool_call_count, etc., these are now provided automatically.
• Third-party integrations (TextArena, ReasoningGym) have been moved to verifiers.envs.integrations.
• Experimental environments (GymEnv, MCPEnv, CliAgentEnv, HarborEnv, RLMEnv) are now in verifiers.envs.experimental and require explicit imports or verifiers[all] installation.

Full Changelog: v0.1.8.post2...v0.1.9