feat!: move v0 under verifiers.legacy (aliased), remove the v1 bridge, drop -v1 suffixes (#2303)

* feat!: remove the v0 legacy bridge from v1

Deletes verifiers/v1/legacy.py and verifiers/v1/configs/legacy.py and every
legacy affordance in the v1 surface:

- serve protocol: drop the info and run_group routes, task_idx addressing
  (task_data is now required), the serve_env/EnvServerPool legacy flag, and
  the LegacyEnvServer worker selection
- eval CLI: drop --legacy.id, the legacy in-process eval path, row-indexed
  resume (whole_task/key_of), and the group-scored concurrency split
- init CLI: drop --v0 scaffolding; gepa: drop the vf-gepa redirect
- EvalConfig: drop the legacy block, is_legacy, and the mixed-run validator
- v1.__init__: drop the LegacyEnvConfig export

v1 no longer imports the v0 tree: install_utils moves to
verifiers/v1/utils/install_utils.py, load_prime_config / CACHE_DIR /
format_cost_usd / msgpack_encoder get v1-local homes.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

* feat!: move all v0 code under verifiers.legacy

The non-v1 tree (envs, clients, serve, scripts, gepa, rubrics, parsers,
utils, types, decorators, errors) moves to verifiers/legacy/.
verifiers/__init__.py collapses to __version__ only — importing the
package root no longer pulls the v0 surface or runs its logging setup;
the old API lives on as `import verifiers.legacy as vf`.

verifiers/cli/ stays at the top level (the Prime CLI executes
verifiers.cli.commands.* by hard-coded module path); its internals and
the vf-* console scripts now point at verifiers.legacy.scripts.*.

Also deletes the dead publish-envs workflow (its publish-only-v0 filter
matches nothing and would mass-publish everything under stale names
after the env renames).

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

* feat!: drop the _v1 suffix from the example environments

Same rename as research-environments: the v0 envs the suffix coexisted
with are gone, so the example tasksets shed it — dirs, packages, project
names, taskset/judge ids, configs, docs, and the lock.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

* Revert "feat!: move all v0 code under verifiers.legacy"

This reverts commit a44506e0f6a56d0ff9de4ec9127f425df4696722.

Scope change: v0 code stays in place; this branch only severs the
v1<->v0 tie (the legacy bridge) and drops the -v1 env suffixes.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

* chore: delete the publish-envs workflow

Its publish-only-classic-envs filter (grep -vE '_v1$|^compact$')
matched nothing while every env carried the _v1 suffix; with the suffix
gone it would auto-publish all example tasksets to the Environments Hub
as if they were classic envs.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

* chore: simplify v1 CLI leftovers from the bridge era

Resume keys are always task_key hashes now that row-index addressing is
gone — collapse the Hashable TypeVar to str in cli/resume.py and
cli/eval/resume.py. Inline the pure config.rich alias in eval main
(it existed to force the dashboard off for bridged runs). Drop stale
docstring references: the retired-keys phrase in configs/cli/env.py,
the v0 framing on the env-level-harness guard, and -v1 example ids in
comments and error messages.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

* fix: env packages advertise python >=3.11 to match verifiers

requires-python = ">=3.10" claimed an interpreter the required
verifiers package (>=3.11,<3.14) can never resolve on.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

* feat!: drop on-demand Hub installs from v1 plugin resolution

A plugin id is the installed package's name: loaders import it directly
and the ID type rejects `org/name[@Version]` references instead of
installing them on demand. install_utils returns to the v0 tree
(vf-install and the v0 eval keep it); v1 no longer imports it. Harbor
dataset ids (their own registry) are unaffected.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

* feat: move v0 code under verifiers.legacy with a top-level alias

All non-v1 code (envs, clients, serve, scripts, gepa, rubrics, parsers,
utils, types, decorators, errors) moves to verifiers/legacy/. A meta-path
finder in the package root resolves every pre-move path to the same
module object — `import verifiers as vf`, `from verifiers.types import
State`, deep submodule imports, and mock.patch targets all keep working,
including the absolute imports inside the moved code itself, so nothing
is rewritten. `verifiers/cli` stays physical for the Prime CLI's
hard-coded `python -m verifiers.cli.commands.*` execution.

Importing the package root (or verifiers.v1) no longer loads the v0
surface or runs its logging setup — that now happens on first touch of a
legacy module or attribute.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

* chore: ID is a plain str; public names for the legacy alias

Drop the _validate_id hub-syntax rejection — an unknown id already fails
at plugin import with the real error. LEGACY_MODULES and
LegacyAliasFinder lose their underscore prefixes.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

* style: ruff 0.16.0 fixes

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

* fix: make the legacy move visible to static tooling

CI ruff and ty can't see through the runtime alias, so first-party code
now states real paths: the legacy tree and verifiers/cli import
verifiers.legacy.* directly (116 files), isort pins the verifiers
namespace first-party (path probing classified the aliased submodules
as third-party), and the package root imports importlib names directly
instead of attribute access ty can't resolve. The alias keeps serving
external consumers and the unchanged v0 tests.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

* fix: GEPA_SCHEMA_VERSION is a format id, not a module path

The import rewrite caught the string constant; the persisted artifact
schema stays verifiers.gepa.v1.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

* docs: move v0 docs to docs/legacy

Nav and cross-links follow; mint.json redirects keep the published
/v0/* URLs working.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

* feat: restore on-demand Hub installs, homed in v1

ensure_installed/env_name/env_module and the hub-format ID validator
return, with install_utils copied into verifiers/v1/utils/ — each stack
owns its copy, so the paths stay fully decoupled (v1 never imports the
legacy tree, the legacy scripts never import v1).

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

---------

Co-authored-by: Claude Fable 5 <noreply@anthropic.com>

Author: mikasenghaas

.github/workflows/publish-envs.yml

@@ -25,4 +25,4 @@
 ## Testing
 
 - **Prefer e2e tests over unit tests**: v1's end-to-end tests are sufficient — extra unit tests clog the repo. Editing existing tests is fine; to check your own work, write a temporary script instead of committing new tests.
-- **Run the contributor checks**: run `uv run pre-commit install` once, then for touched areas `uv run ruff check --fix .`, `uv run pytest tests/`, and `uv run pre-commit run --all-files` (see `docs/v0/development.md`).
+- **Run the contributor checks**: run `uv run pre-commit install` once, then for touched areas `uv run ruff check --fix .`, `uv run pytest tests/`, and `uv run pre-commit run --all-files` (see `docs/legacy/development.md`).

AGENTS.md

@@ -4,7 +4,7 @@ num_tasks = 2
 id = "agentic-judge"
 
 [env.taskset]
-id = "gsm8k-v1"
+id = "gsm8k"
 
 [env.solver.harness]
 id = "bash"

configs/agentic_judge.toml

@@ -1,11 +1,11 @@
-# alphabet-sort-v1 — multi-turn alphabetical sorting, driven by a colocated user simulator.
+# alphabet-sort — multi-turn alphabetical sorting, driven by a colocated user simulator.
 #
 #   uv run eval @ configs/alphabet_sort.toml --model <small-instruct-model>
 num_tasks = 5
 num_rollouts = 2
 
 [env.taskset]
-id = "alphabet-sort-v1"
+id = "alphabet-sort"
 min_turns = 2
 max_turns = 2
 

configs/alphabet_sort.toml

@@ -1,10 +1,10 @@
-# code-golf-v1 — the taskset ships a recipe env: one episode runs
+# code-golf — the taskset ships a recipe env: one episode runs
 # `--env.attempts` independent attempts (default 2) and scores them against
 # each other (shortest wins), so one rollout per task is a full comparison.
 #
-#   uv run eval code-golf-v1 @ configs/code_golf.toml
+#   uv run eval code-golf @ configs/code_golf.toml
 num_tasks = 1
 num_rollouts = 1
 
 [env.taskset]
-id = "code-golf-v1"
+id = "code-golf"

configs/code_golf.toml

@@ -1,7 +1,7 @@
-# deepwiki-v1 — an existing remote MCP server, reached by URL (nothing to run ourselves).
+# deepwiki — an existing remote MCP server, reached by URL (nothing to run ourselves).
 #
 #   uv run eval @ configs/deepwiki.toml
 num_tasks = 1
 
 [env.taskset]
-id = "deepwiki-v1"
+id = "deepwiki"

configs/deepwiki.toml

@@ -1,7 +1,7 @@
-# glossary-v1 — a colocated MCP tool server (runs inside the harness's own runtime).
+# glossary — a colocated MCP tool server (runs inside the harness's own runtime).
 #
-#   uv run eval glossary-v1 @ configs/glossary.toml
+#   uv run eval glossary @ configs/glossary.toml
 num_tasks = 1
 
 [env.taskset]
-id = "glossary-v1"
+id = "glossary"

configs/glossary.toml

@@ -1,8 +1,8 @@
-# gsm8k-v1 — grade-school math, bash harness; a few tasks with several rollouts each.
+# gsm8k — grade-school math, bash harness; a few tasks with several rollouts each.
 #
-#   uv run eval gsm8k-v1 @ configs/gsm8k.toml
+#   uv run eval gsm8k @ configs/gsm8k.toml
 num_tasks = 5
 num_rollouts = 3
 
 [env.taskset]
-id = "gsm8k-v1"
+id = "gsm8k"

configs/gsm8k.toml

@@ -1,4 +1,4 @@
-# The gsm8k-v1 taskset run with the rlm harness, configured entirely from TOML.
+# The gsm8k taskset run with the rlm harness, configured entirely from TOML.
 #
 #   uv run eval @ configs/gsm8k_rlm.toml
 #
@@ -10,7 +10,7 @@
 num_tasks = 1
 
 [env.taskset]
-id = "gsm8k-v1"
+id = "gsm8k"
 
 [env.agent.harness]
 id = "rlm"

configs/gsm8k_rlm.toml

@@ -1,7 +1,7 @@
-# wiki-search-v1 — a shared MCP tool server (the corpus is built once, reused across rollouts).
+# wiki-search — a shared MCP tool server (the corpus is built once, reused across rollouts).
 #
 #   uv run eval @ configs/wiki_search.toml
 num_tasks = 1
 
 [env.taskset]
-id = "wiki-search-v1"
+id = "wiki-search"