reeval-python 0.14.0__tar.gz → 0.15.1__tar.gz

{reeval_python-0.14.0 → reeval_python-0.15.1}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.5
 Name: reeval-python
-Version: 0.14.0
+Version: 0.15.1
 Summary: Python authoring SDK for reeval (*.eval.py eval files)
 Project-URL: Homepage, https://reeval.ai
 Project-URL: Repository, https://github.com/invertase/reeval
@@ -86,6 +86,35 @@ a retry loop can assert; one the escaping error carries (each failure in a `Task
 for quick checks. Keep `expect` for reeval matchers (`to_have_file`, `to_succeed`, soft checks,
 labels) — Playwright-style.
 
+#### What a caught failure counts as
+
+A failed `expect` is a check the moment it fails: reeval records it before
+`ReevalAssertionError` reaches the body. One the body catches still fails the attempt when the
+body ends, as `expect.soft` does, and its error says the eval caught it. A plain `assert` the body
+catches records nothing, as in Python.
+
+| The body catches | Recorded | Fails the attempt |
+|---|---|---|
+| a failed `expect(…)` | a failed check | yes, when the body ends |
+| a failed `assert` | nothing | no |
+| neither (it escapes) | a failed check | yes |
+
+So retry with `expect.poll`, which retries the matcher and records only its final outcome, or with
+a plain `assert` or `if` in the loop, never `expect` in `try`/`except`:
+
+```python
+# Retries the command until it succeeds, for up to 10 seconds: one check.
+expect.poll(lambda: c.exec("curl -sf localhost:3000"), timeout=10).to_succeed()
+
+# A plain assert in a retry loop: only the try that held is recorded.
+for _ in range(3):
+    try:
+        assert c.exec("test -f ready.txt").ok
+        break
+    except AssertionError:
+        time.sleep(1)
+```
+
 Matchers that take a pattern (`to_have_file`, `to_have_run`, `to_have_searched`, `to_have_fetched`,
 `to_have_called_tool`, `to_have_edited`, `to_have_changed`) take a `str` or a compiled `re.Pattern`,
 `str` or `bytes`, with any flags. Python matches it with `pattern.search`, so anything `re` accepts
@@ -101,7 +130,7 @@ Use snake_case on the context (`read_file`, `write_file`, `file`) and on matcher
 `to_pass`, `to_equal`, `to_contain`, …). Matcher failures raise `ReevalAssertionError`.
 
 Options are typed (`EvalOptions`, and `JudgeOptions` for `c.judge(…)`), so pyright refuses a
-misspelt key such as `runz`. They take reeval's keys in snake_case at any depth (`min_lift`,
+misspelled key such as `runz`. They take reeval's keys in snake_case at any depth (`min_lift`,
 `delay_ms`); the keys of `env`, `headers` and the other maps of your own keep their spelling. Every
 duration is seconds or a `timedelta`: `{"timeout": 90}`, `docker(start_timeout=timedelta(minutes=2))`,
 `expect.poll(…, timeout=10)`. `docker()`, `fake_harness()`, `fake_judge()` and `plugin.*` return

{reeval_python-0.14.0 → reeval_python-0.15.1}/README.md

@@ -62,6 +62,35 @@ a retry loop can assert; one the escaping error carries (each failure in a `Task
 for quick checks. Keep `expect` for reeval matchers (`to_have_file`, `to_succeed`, soft checks,
 labels) — Playwright-style.
 
+#### What a caught failure counts as
+
+A failed `expect` is a check the moment it fails: reeval records it before
+`ReevalAssertionError` reaches the body. One the body catches still fails the attempt when the
+body ends, as `expect.soft` does, and its error says the eval caught it. A plain `assert` the body
+catches records nothing, as in Python.
+
+| The body catches | Recorded | Fails the attempt |
+|---|---|---|
+| a failed `expect(…)` | a failed check | yes, when the body ends |
+| a failed `assert` | nothing | no |
+| neither (it escapes) | a failed check | yes |
+
+So retry with `expect.poll`, which retries the matcher and records only its final outcome, or with
+a plain `assert` or `if` in the loop, never `expect` in `try`/`except`:
+
+```python
+# Retries the command until it succeeds, for up to 10 seconds: one check.
+expect.poll(lambda: c.exec("curl -sf localhost:3000"), timeout=10).to_succeed()
+
+# A plain assert in a retry loop: only the try that held is recorded.
+for _ in range(3):
+    try:
+        assert c.exec("test -f ready.txt").ok
+        break
+    except AssertionError:
+        time.sleep(1)
+```
+
 Matchers that take a pattern (`to_have_file`, `to_have_run`, `to_have_searched`, `to_have_fetched`,
 `to_have_called_tool`, `to_have_edited`, `to_have_changed`) take a `str` or a compiled `re.Pattern`,
 `str` or `bytes`, with any flags. Python matches it with `pattern.search`, so anything `re` accepts
@@ -77,7 +106,7 @@ Use snake_case on the context (`read_file`, `write_file`, `file`) and on matcher
 `to_pass`, `to_equal`, `to_contain`, …). Matcher failures raise `ReevalAssertionError`.
 
 Options are typed (`EvalOptions`, and `JudgeOptions` for `c.judge(…)`), so pyright refuses a
-misspelt key such as `runz`. They take reeval's keys in snake_case at any depth (`min_lift`,
+misspelled key such as `runz`. They take reeval's keys in snake_case at any depth (`min_lift`,
 `delay_ms`); the keys of `env`, `headers` and the other maps of your own keep their spelling. Every
 duration is seconds or a `timedelta`: `{"timeout": 90}`, `docker(start_timeout=timedelta(minutes=2))`,
 `expect.poll(…, timeout=10)`. `docker()`, `fake_harness()`, `fake_judge()` and `plugin.*` return

{reeval_python-0.14.0 → reeval_python-0.15.1}/pyproject.toml

@@ -1,6 +1,6 @@
 [project]
 name = "reeval-python"
-version = "0.14.0"
+version = "0.15.1"
 description = "Python authoring SDK for reeval (*.eval.py eval files)"
 readme = "README.md"
 requires-python = ">=3.11"

{reeval_python-0.14.0 → reeval_python-0.15.1}/src/reeval_python/_wire.py

@@ -73,8 +73,8 @@ class AssertionRecord(TypedDict):
 class AttemptAbort(TypedDict):
     # reeval's sentence, e.g. `Eval timed out after 600s`.
     message: str
-    # `timeout`: the eval's `timeout` ran out. `cancelled`: the run was stopped.
-    reason: Literal["timeout", "cancelled"]
+    # `timeout`: the eval's `timeout` ran out. `canceled`: the run was stopped.
+    reason: Literal["timeout", "canceled"]
     type: Literal["abort"]
 
 
@@ -389,6 +389,8 @@ class CaptureCallData(TypedDict):
     instructions: NotRequired[list[CaptureInstruction]]
     # Overrides cascaded `capture.model` for this call.
     model: NotRequired[str]
+    # The `artifacts/<name>.png` file name: 1 to 100 of `a-z`, `0-9`, `_` and `-`, starting with
+    # a letter or digit. Unique within the attempt.
     name: NotRequired[str]
     path: NotRequired[str]
     timeout: NotRequired[Duration]
@@ -500,7 +502,7 @@ class CodexHarnessDescriptor(TypedDict):
 class ConfigJudgeOptions(TypedDict):
     # The gateway the judge's calls go through, by name or by its helper. It skips the config's
     # `gateway` fallback: the plan refuses a gateway that isn't listed or can't serve the model.
-    gateway: NotRequired[DirectGatewayDescriptor | VercelGatewayDescriptor | Literal["direct", "vercel"]]
+    gateway: NotRequired[DirectGatewayDescriptor | OpenRouterGatewayDescriptor | VercelGatewayDescriptor | VertexGatewayDescriptor | Literal["direct", "openrouter", "vercel", "vertex"]]
     # The most judge calls to this model's provider at once, across the run.
     maxConcurrent: NotRequired[float]
     # `'provider/model'`, or a fake judge.
@@ -557,6 +559,7 @@ class DeclaredEval(TypedDict):
     name: str
     # The eval's own options, before the cascade.
     options: EvalOptionsData
+    source: DeclaredSource
     type: Literal["eval"]
 
 
@@ -589,6 +592,16 @@ DeclaredMode: TypeAlias = "Literal[\"run\", \"skip\", \"only\"]"
 DeclaredNode: TypeAlias = "DeclaredEval | DeclaredGroup | DeclaredTodo"
 
 
+# An eval's own source, hashed by the SDK that declared it: sha256 (hex) of its body's source
+# text, and of an `evaluate.each` case's data as canonical JSON (object keys sorted, no
+# whitespace, non-ASCII characters as themselves).
+class DeclaredSource(TypedDict):
+    # sha256 of the body function's source text.
+    body: str
+    # sha256 of the case's data; set on each case of an `evaluate.each`.
+    case: NotRequired[str]
+
+
 # `evaluate.todo` or `group.todo`.
 class DeclaredTodo(TypedDict):
     # The name passed to `.todo`.
@@ -601,8 +614,12 @@ class Diff(TypedDict):
     patch: str
 
 
-# Each provider's own key, as without `gateway`: made by `direct()` in `reeval/gateway`.
+# Each provider's own key: made by `direct()` in `reeval/gateway`. Every `gateway` list ends
+# with it unless it's listed elsewhere, where it's tried in that place.
 class DirectGatewayDescriptor(TypedDict):
+    # `false` turns off the fallback to each provider's own key, so a cell no other gateway in
+    # the list serves is refused at plan time. Default `true`.
+    enabled: NotRequired[bool]
     kind: Literal["direct"]
 
 
@@ -628,7 +645,7 @@ DockerSandboxDescriptor = TypedDict(
 )
 
 
-# More than zero: `90s`, `15m`, `1h`, or a number of milliseconds.
+# More than zero and at most 596h: `90s`, `15m`, `1h`, or a number of milliseconds.
 Duration: TypeAlias = "float | str"
 
 
@@ -755,7 +772,13 @@ class EvalOptionsDataCompare(TypedDict):
     # all end; one without is asked of each attempt as it ends. Replaces the inherited list, so
     # `[]` asks none.
     questions: NotRequired[list[Question]]
-    # Attempts per cell.
+    # The eval's version, part of its fingerprint: bump it when a change the eval file doesn't
+    # show (a helper, a fixture, data it reads) changes what it measures. A group's applies to
+    # the evals inside it that set none.
+    revision: NotRequired[float | str]
+    # Attempts per cell. Unset at every level, a compare with variants runs 4 a side, the fewest
+    # whose lift can reach a verdict, and anything else runs once, a pooled `evaluate.each` case
+    # included.
     runs: NotRequired[float]
     # The default machine when the resolved environment omits `sandbox`. Cascades like
     # `harness`. Not an environment: it does not invent skills, files or a name.
@@ -824,7 +847,13 @@ class EvalOptionsDataCompare2(TypedDict):
     # all end; one without is asked of each attempt as it ends. Replaces the inherited list, so
     # `[]` asks none.
     questions: NotRequired[list[Question]]
-    # Attempts per cell.
+    # The eval's version, part of its fingerprint: bump it when a change the eval file doesn't
+    # show (a helper, a fixture, data it reads) changes what it measures. A group's applies to
+    # the evals inside it that set none.
+    revision: NotRequired[float | str]
+    # Attempts per cell. Unset at every level, a compare with variants runs 4 a side, the fewest
+    # whose lift can reach a verdict, and anything else runs once, a pooled `evaluate.each` case
+    # included.
     runs: NotRequired[float]
     # The default machine when the resolved environment omits `sandbox`. Cascades like
     # `harness`. Not an environment: it does not invent skills, files or a name.
@@ -877,7 +906,13 @@ class EvalOptionsData3(TypedDict):
     # all end; one without is asked of each attempt as it ends. Replaces the inherited list, so
     # `[]` asks none.
     questions: NotRequired[list[Question]]
-    # Attempts per cell.
+    # The eval's version, part of its fingerprint: bump it when a change the eval file doesn't
+    # show (a helper, a fixture, data it reads) changes what it measures. A group's applies to
+    # the evals inside it that set none.
+    revision: NotRequired[float | str]
+    # Attempts per cell. Unset at every level, a compare with variants runs 4 a side, the fewest
+    # whose lift can reach a verdict, and anything else runs once, a pooled `evaluate.each` case
+    # included.
     runs: NotRequired[float]
     # The default machine when the resolved environment omits `sandbox`. Cascades like
     # `harness`. Not an environment: it does not invent skills, files or a name.
@@ -915,7 +950,7 @@ class FakeHarnessDescriptorWithSkill(TypedDict):
 
 # A canned agent for tests and dry runs, made by `fakeHarness()` in `reeval/harness`.
 class FakeHarnessDescriptor(TypedDict):
-    # Simulated turn length, honours the abort signal.
+    # Simulated turn length, honors the abort signal.
     delayMs: NotRequired[float]
     # Requested reasoning effort, recorded as `effort` on the cell. Use `servedModel` for what
     # the turn reports (`fake-1[effort=medium]`). Fast is not this.
@@ -1026,20 +1061,27 @@ class FileChange(TypedDict):
 
 
 # fx through `@reeval/harness-fx` (`fx acp`), made by `fx()`. It runs on Vercel AI Gateway or
-# OpenAI's API by the model id, and reports each turn's tokens.
+# OpenAI's API by the model id, or on OpenRouter or Vertex AI, and reports each turn's tokens.
 class FxHarnessDescriptor(TypedDict):
+    # Reasoning effort, sent as `reasoning_effort` on each model call; only on `vertex()`, since
+    # fx sends none itself. Values are what the model takes there, e.g. `low`, `medium` or
+    # `high`.
+    effort: NotRequired[str]
     # The gateway this harness's model calls go through, by name. It skips the config's
     # `gateway` fallback: the plan refuses a gateway that isn't listed or can't serve this
     # harness.
     gateway: NotRequired[GatewayName]
-    # Cell id; defaults to `fx:<model>`.
+    # Cell id; defaults to `fx:<model>`, or `fx:<model>[effort=<value>]` when `effort` is set.
     id: NotRequired[str]
     kind: Literal["fx"]
     # The most fx attempts at once, across the run, e.g. for the key's rate limits.
     maxConcurrent: NotRequired[float]
     # A Vercel AI Gateway model id, `<provider>/<model>`, run on fx's `gateway` provider with
     # `AI_GATEWAY_API_KEY`, e.g. `anthropic/claude-sonnet-5`; or a bare OpenAI model id, run on
-    # OpenAI's Chat Completions API with `OPENAI_API_KEY`, e.g. `gpt-5.5`.
+    # OpenAI's Chat Completions API with `OPENAI_API_KEY`, e.g. `gpt-5.5`. On `openrouter()`,
+    # either runs on OpenRouter's Chat Completions API with `OPENROUTER_API_KEY`, a bare id as
+    # `openai/<id>`. On `vertex()`, a Gemini model id, `google/<id>` or a bare `gemini-…` id,
+    # run on Vertex AI's Chat Completions endpoint, e.g. `google/gemini-3.8-flash`.
     model: str
     # How long the harness's session may take to become ready (its bootstrap and first session)
     # before the attempt is an infra error, after one retry; default 90s.
@@ -1047,8 +1089,9 @@ class FxHarnessDescriptor(TypedDict):
 
 
 # A gateway the config lists by name: `direct` is each provider's own key, as without `gateway`;
-# `vercel` is Vercel AI Gateway.
-GatewayName: TypeAlias = "Literal[\"direct\", \"vercel\"]"
+# `vercel` is Vercel AI Gateway; `openrouter` is OpenRouter; `vertex` is Google Cloud's Vertex
+# AI.
+GatewayName: TypeAlias = "Literal[\"direct\", \"vercel\", \"openrouter\", \"vertex\"]"
 
 
 # A git repository (or one directory of it) the workspace starts as, from the `git` helper.
@@ -1149,8 +1192,12 @@ class JudgeArtifactRef(TypedDict):
     sha256: str
 
 
-# Each provider's own key, as without `gateway`: made by `direct()` in `reeval/gateway`.
+# Each provider's own key: made by `direct()` in `reeval/gateway`. Every `gateway` list ends
+# with it unless it's listed elsewhere, where it's tried in that place.
 class JudgeCallDataGatewayDirect(TypedDict):
+    # `false` turns off the fallback to each provider's own key, so a cell no other gateway in
+    # the list serves is refused at plan time. Default `true`.
+    enabled: NotRequired[bool]
     kind: Literal["direct"]
 
 
@@ -1160,6 +1207,30 @@ class JudgeCallDataGatewayVercel(TypedDict):
     kind: Literal["vercel"]
 
 
+# OpenRouter, with `OPENROUTER_API_KEY`: made by `openrouter()` in `reeval/gateway`. It serves
+# Claude Code, Codex, Goose, OpenCode, fx, Grok Build and pi.
+class JudgeCallDataGatewayOpenrouter(TypedDict):
+    # Let OpenRouter fail over to another upstream provider when the one it picked can't serve a
+    # call. Off by default, so reeval's own calls (the judge, analysis, compare questions and
+    # captures) send `allow_fallbacks: false`. The harnesses' calls take OpenRouter's own
+    # routing either way.
+    fallbacks: NotRequired[bool]
+    kind: Literal["openrouter"]
+
+
+# Google Cloud's Vertex AI, on Application Default Credentials: made by `vertex()` in
+# `reeval/gateway`.
+class JudgeCallDataGatewayVertex(TypedDict):
+    kind: Literal["vertex"]
+    # The region the models are called in, e.g. `us-east5`; defaults to `global`.
+    location: NotRequired[str]
+    # The Google Cloud project the calls are billed to; defaults to `GOOGLE_CLOUD_PROJECT`.
+    project: NotRequired[str]
+    # A region per publisher, where it differs from `location`, e.g. `{ anthropic: 'us-east5'
+    # }`.
+    regions: NotRequired[dict[str, str]]
+
+
 class JudgeCallDataOnTurn(TypedDict):
     # The turn's index, from 0.
     turn: int
@@ -1179,7 +1250,7 @@ class JudgeCallData(TypedDict):
     artifacts: NotRequired[Literal[True] | list[str | Capture]]
     # The gateway the judge's calls go through, by name or by its helper. It skips the config's
     # `gateway` fallback: the plan refuses a gateway that isn't listed or can't serve the model.
-    gateway: NotRequired[Literal["direct", "vercel"] | JudgeCallDataGatewayDirect | JudgeCallDataGatewayVercel]
+    gateway: NotRequired[Literal["direct", "vercel", "openrouter", "vertex"] | JudgeCallDataGatewayDirect | JudgeCallDataGatewayVercel | JudgeCallDataGatewayOpenrouter | JudgeCallDataGatewayVertex]
     # The most judge calls to this model's provider at once, across the run.
     maxConcurrent: NotRequired[int]
     # `'provider/model'`, or a fake judge with a numeric score.
@@ -1203,7 +1274,7 @@ class JudgeFailure(TypedDict):
     artifacts: NotRequired[list[JudgeArtifactRef]]
     error: str
     # The gateway from the config's `gateway` list the call went through; unset without one.
-    gateway: NotRequired[Literal["direct", "vercel"]]
+    gateway: NotRequired[Literal["direct", "vercel", "openrouter", "vertex"]]
     model: str
     question: str
     score: None
@@ -1326,6 +1397,17 @@ class OpenCodeHarnessDescriptor(TypedDict):
     startTimeout: NotRequired[Duration]
 
 
+# OpenRouter, with `OPENROUTER_API_KEY`: made by `openrouter()` in `reeval/gateway`. It serves
+# Claude Code, Codex, Goose, OpenCode, fx, Grok Build and pi.
+class OpenRouterGatewayDescriptor(TypedDict):
+    # Let OpenRouter fail over to another upstream provider when the one it picked can't serve a
+    # call. Off by default, so reeval's own calls (the judge, analysis, compare questions and
+    # captures) send `allow_fallbacks: false`. The harnesses' calls take OpenRouter's own
+    # routing either way.
+    fallbacks: NotRequired[bool]
+    kind: Literal["openrouter"]
+
+
 # A matrix factor a pairwise question varies; the others are held.
 PairFactor: TypeAlias = "Literal[\"environment\", \"harness\"]"
 
@@ -1679,7 +1761,9 @@ class TranscriptEventFileEdit(TypedDict):
     added: float | None
     atMs: float
     durationMs: float | None
-    kind: Literal["create", "modify", "delete"]
+    # `write` is a whole-file write whose result doesn't say whether the file existed before
+    # (Antigravity's, fx's, Goose's, Grok Build's and pi's write tools, for example).
+    kind: Literal["create", "modify", "write", "delete"]
     path: str
     removed: float | None
     # True when a shell-like tool or edit ran at least two minutes.
@@ -1744,7 +1828,7 @@ class TranscriptEventRetry(TypedDict):
     type: Literal["retry"]
 
 
-# Normalised agent events. Each tool call is one event, merged with its result; `durationMs` is
+# Normalized agent events. Each tool call is one event, merged with its result; `durationMs` is
 # wall time until the result arrived (Cursor's own `completed_at_ms` can be earlier). Null when
 # the harness never reported the end. Long shell or edit calls set `stalled: true`.
 TranscriptEvent: TypeAlias = "TranscriptEvent1 | TranscriptEventFileEdit | TranscriptEventWebSearch | TranscriptEventWebFetch | TranscriptEventTool | TranscriptEventError | TranscriptEventRetry"
@@ -1786,8 +1870,11 @@ class VercelGatewayDescriptor(TypedDict):
 class Verdict(TypedDict):
     # Captures this call sent the judge as file parts, when any.
     artifacts: NotRequired[list[JudgeArtifactRef]]
+    # What the gateway says the call cost, in US dollars (OpenRouter's `usage.cost`); unset when
+    # it names none.
+    costUsd: NotRequired[float]
     # The gateway from the config's `gateway` list the call went through; unset without one.
-    gateway: NotRequired[Literal["direct", "vercel"]]
+    gateway: NotRequired[Literal["direct", "vercel", "openrouter", "vertex"]]
     model: str
     question: str
     # The model id the provider says served the call, when it reports one.
@@ -1795,9 +1882,25 @@ class Verdict(TypedDict):
     # The judge's probability that the answer is yes, from 0 to 1.
     score: float
     threshold: float
+    # The upstream provider the gateway says served the call (OpenRouter's); unset when it names
+    # none.
+    upstream: NotRequired[str]
     verdict: Literal["pass", "fail"]
 
 
+# Google Cloud's Vertex AI, on Application Default Credentials: made by `vertex()` in
+# `reeval/gateway`.
+class VertexGatewayDescriptor(TypedDict):
+    kind: Literal["vertex"]
+    # The region the models are called in, e.g. `us-east5`; defaults to `global`.
+    location: NotRequired[str]
+    # The Google Cloud project the calls are billed to; defaults to `GOOGLE_CLOUD_PROJECT`.
+    project: NotRequired[str]
+    # A region per publisher, where it differs from `location`, e.g. `{ anthropic: 'us-east5'
+    # }`.
+    regions: NotRequired[dict[str, str]]
+
+
 class WaitMs(TypedDict):
     ms: float
 
@@ -1816,6 +1919,7 @@ CAMEL_KEYS: frozenset[str] = frozenset(
         "atMs",
         "cacheRead",
         "cacheWrite",
+        "costUsd",
         "delayMs",
         "durationMs",
         "exitCode",
@@ -1850,6 +1954,7 @@ RECORD_FIELDS: frozenset[str] = frozenset(
         "mcp",
         "params",
         "providerOptions",
+        "regions",
         "score",
         "untested",
         "value",

reeval_python-0.15.1/src/reeval_python/binaries.json

@@ -0,0 +1,55 @@
+{
+  "version": "0.15.1",
+  "binaries": {
+    "linux-x64": {
+      "name": "reeval-0.15.1-linux-x64",
+      "sha256": "1514aa35048abc1bb90174f4bae90cbad829008c62cc7c61b7a061d616f190aa",
+      "size": 90777056
+    },
+    "linux-arm64": {
+      "name": "reeval-0.15.1-linux-arm64",
+      "sha256": "9765c6f7b5c83cea1f4040143ee15403540404598897b6cef90eaa695b4c7f90",
+      "size": 90753320
+    },
+    "linux-x64-musl": {
+      "name": "reeval-0.15.1-linux-x64-musl",
+      "sha256": "98524643d0be87ba8f8a4320ec2e20f6c6c1e7a77cd5418005e6602a0cc54f0b",
+      "size": 84747824
+    },
+    "linux-arm64-musl": {
+      "name": "reeval-0.15.1-linux-arm64-musl",
+      "sha256": "822041705bdaaff553fa34286abf29435c2e9866afd9ee6fa76b06a1c55730f4",
+      "size": 83889360
+    },
+    "darwin-x64": {
+      "name": "reeval-0.15.1-darwin-x64",
+      "sha256": "cea280d80792e701b9b1d559415d24a02f3cdf6ba1040f9d428e5bf5bb97e533",
+      "size": 78786832
+    },
+    "darwin-arm64": {
+      "name": "reeval-0.15.1-darwin-arm64",
+      "sha256": "e7f1fdeb31f41e79ec137d22ff036a25201f0e960f996b669602aef246462c16",
+      "size": 71754354
+    },
+    "windows-x64": {
+      "name": "reeval-0.15.1-windows-x64.exe",
+      "sha256": "efe14b3ee751bbf4132c476eb138e8a32036122d1f4ec6f927bb31f302d0e32d",
+      "size": 95552512
+    },
+    "linux-x64-baseline": {
+      "name": "reeval-0.15.1-linux-x64-baseline",
+      "sha256": "0d3dca1f8ed50316aff10edf990611394da259d280bbeb52704548080508967d",
+      "size": 90777056
+    },
+    "linux-x64-musl-baseline": {
+      "name": "reeval-0.15.1-linux-x64-musl-baseline",
+      "sha256": "a5091a4a02eed53d676c5f670a661c18657507b425bf28705b4f7afeb8417d80",
+      "size": 84747824
+    },
+    "windows-x64-baseline": {
+      "name": "reeval-0.15.1-windows-x64-baseline.exe",
+      "sha256": "8bc8772981f5352f3a08824193a1a59c324e105508b20c41814ec4e41f500c58",
+      "size": 95552512
+    }
+  }
+}

{reeval_python-0.14.0 → reeval_python-0.15.1}/src/reeval_python/cli.py

@@ -40,13 +40,26 @@ class ShimError(Exception):
         self.code = code
 
 
-PRERELEASES = {"a": "alpha", "b": "beta", "rc": "rc"}
+PRERELEASES = {
+    "a": "alpha",
+    "alpha": "alpha",
+    "b": "beta",
+    "beta": "beta",
+    "c": "rc",
+    "pre": "rc",
+    "preview": "rc",
+    "rc": "rc",
+}
 
 
 def semver(version: str) -> str:
-    """reeval's version for a reeval-python version: 1.0.0b1 is 1.0.0-beta.1."""
-    if m := re.fullmatch(r"(\d+\.\d+\.\d+)(a|b|rc)(\d+)", version):
-        return f"{m[1]}-{PRERELEASES[m[2]]}.{m[3]}"
+    """reeval's version for a reeval-python version: 1.0.0b1 is 1.0.0-beta.1.
+
+    Takes any PEP 440 spelling of a prerelease (1.0.0-rc1, 1.0.0RC1 and 1.0.0c1 are 1.0.0rc1).
+    """
+    pattern = r"(\d+\.\d+\.\d+)[-_.]?(alpha|a|beta|b|preview|pre|rc|c)[-_.]?(\d+)"
+    if m := re.fullmatch(pattern, version, re.IGNORECASE):
+        return f"{m[1]}-{PRERELEASES[m[2].lower()]}.{m[3]}"
     return version
 
 
@@ -58,9 +71,13 @@ def own_version() -> str | None:
         return None
 
 
+def pinned_version() -> str:
+    """REEVAL_VERSION as reeval spells its version, so 1.0.0rc1 is 1.0.0-rc.1; empty if unset."""
+    return semver(os.environ.get("REEVAL_VERSION", "").strip().removeprefix("v"))
+
+
 def binary_version() -> str:
-    pinned = os.environ.get("REEVAL_VERSION", "").strip().removeprefix("v")
-    if pinned:
+    if pinned := pinned_version():
         return pinned
     if own := own_version():
         return own
@@ -415,7 +432,7 @@ def resolve() -> Path:
                 f"REEVAL_BINARY is {override}, which is this `reeval` command, not the binary.", 2
             )
         return path
-    pinned = os.environ.get("REEVAL_VERSION", "").strip().removeprefix("v")
+    pinned = pinned_version()
     own = bundled()
     if own.is_file() and pinned in ("", own_version()):
         if not WINDOWS and not os.access(own, os.X_OK):