Fix GHOSTRUN_* env var casing left wrong by the gentest rename

The gentest->ghostrun rename turned GENTEST_JUDGE, GENTEST_MODE, etc.
into ghostrun_JUDGE, ghostrun_MODE (lowercase prefix) instead of
GHOSTRUN_JUDGE, GHOSTRUN_MODE, breaking the documented env var
convention across ghostrun/, tests/, examples/, CI, and docs.

Author: parthmax2

.github/workflows/ci.yml

@@ -41,19 +41,19 @@ jobs:
       # The core suite is hermetic: no network, no Ollama, no API keys.
       - name: Test (offline)
         env:
-          ghostrun_JUDGE: echo
+          GHOSTRUN_JUDGE: echo
         run: pytest tests -q
 
       # The example must skip cleanly rather than fail when no judge exists.
       - name: Example degrades gracefully without a judge
         env:
-          ghostrun_JUDGE: echo
+          GHOSTRUN_JUDGE: echo
         run: pytest examples -q
 
       # Replay must work with no network and no judge backend reachable.
       - name: Strict replay (no network, judge unreachable)
         env:
-          ghostrun_JUDGE_BASE_URL: http://localhost:59999
+          GHOSTRUN_JUDGE_BASE_URL: http://localhost:59999
         run: pytest examples -q --ghostrun-replay
 
   parallel:
@@ -72,7 +72,7 @@ jobs:
       # The interceptor patches global httpx state; this guards against races.
       - name: Test under parallel workers
         env:
-          ghostrun_JUDGE: echo
+          GHOSTRUN_JUDGE: echo
         run: pytest tests -q -n 4
 
   build:
@@ -118,6 +118,6 @@ jobs:
       - name: Record from the real API, then replay offline
         env:
           OPENAI_API_KEY: ${{ secrets.OPENAI_API_KEY }}
-          ghostrun_LIVE: "1"
-          ghostrun_JUDGE: echo   # don't require Ollama in CI
+          GHOSTRUN_LIVE: "1"
+          GHOSTRUN_JUDGE: echo   # don't require Ollama in CI
         run: pytest examples/test_live_smoke.py -v

CHANGELOG.md

@@ -73,7 +73,7 @@ All notable changes to ghostrun are documented here. The format is based on
   freezes whatever the judge said on one draw, including if it was wrong —
   published LLM-as-judge studies report ~13–14% flip rates on repeated grading
   of identical input, even at temperature 0. Setting `judge.votes` (or
-  `ghostrun_JUDGE_VOTES`) to an odd N > 1 grades N times on a cache miss and
+  `GHOSTRUN_JUDGE_VOTES`) to an odd N > 1 grades N times on a cache miss and
   caches the majority verdict plus the observed disagreement rate
   (`Grade.votes`, `Grade.disagreement_rate`). Even vote counts tie-break
   conservatively to FAIL. Cache keys now include the vote count, so changing
@@ -104,7 +104,7 @@ All notable changes to ghostrun are documented here. The format is based on
   judge could flip a passing test with no code change. Verdicts are now recorded
   to `.ghostrun_cache/judge/` under the same `auto`/`record`/`replay` semantics as
   HTTP calls, keyed on judge backend + model + text + criterion. Configurable via
-  `judge.cache` or `ghostrun_JUDGE_CACHE`. Cuts the bundled example from 23.3s to
+  `judge.cache` or `GHOSTRUN_JUDGE_CACHE`. Cuts the bundled example from 23.3s to
   2.7s and makes semantic assertions deterministic.
 - Committed tests for async clients and streaming (SSE) record/replay, including
   async streaming — previously working but uncovered.
@@ -161,7 +161,7 @@ Initial Phase 1 MVP.
   stub for CI.
 - **Pytest plugin** (auto-registered): `--ghostrun-record`, `--ghostrun-replay`,
   `--ghostrun-judge` flags and a `ghostrun_record` fixture.
-- **Configuration** via `.ghostrun.yaml`, `ghostrun_*` environment variables, or
+- **Configuration** via `.ghostrun.yaml`, `GHOSTRUN_*` environment variables, or
   `ghostrun.configure(...)`, resolved defaults → file → env.
 - Documentation (`README.md`, `doc/prd.md`), a runnable `examples/` app with a
   pre-recorded cache, and an offline test suite.

CONTRIBUTING.md

@@ -21,7 +21,7 @@ real Ollama judge when available and skips gracefully otherwise.
 
 - `pytest tests -q` must pass.
 - If you touch the interceptor, cache, or judge-caching logic, also run under
-  parallel workers: `ghostrun_JUDGE=echo pytest tests -q -n 4` — several real
+  parallel workers: `GHOSTRUN_JUDGE=echo pytest tests -q -n 4` — several real
   bugs in this project (cross-thread cache contamination, torn writes) were
   only found this way, not by the sequential suite.
 - New behavior needs a test that fails without the fix. This project has

doc/guide/configuration.md

@@ -16,8 +16,8 @@ judge:
   cache: true
 ```
 
-Everything is overridable by environment variable (`ghostrun_MODE`,
-`ghostrun_JUDGE`, `ghostrun_JUDGE_MODEL`, `ghostrun_CACHE_DIR`, `ghostrun_JUDGE_VOTES`,
+Everything is overridable by environment variable (`GHOSTRUN_MODE`,
+`GHOSTRUN_JUDGE`, `GHOSTRUN_JUDGE_MODEL`, `GHOSTRUN_CACHE_DIR`, `GHOSTRUN_JUDGE_VOTES`,
 …) or at runtime via `ghostrun.configure(judge="echo")`. Resolution order:
 defaults → `.ghostrun.yaml` → environment variables → `ghostrun.configure(...)`.
 

doc/guide/recording.md

@@ -5,7 +5,7 @@ and Anthropic SDKs sit on top of. It does *not* monkey-patch the SDKs, so it
 survives SDK upgrades. Only requests to known provider hosts are cached; all
 other traffic passes through untouched.
 
-Modes (via `ghostrun_MODE` or `.ghostrun.yaml`):
+Modes (via `GHOSTRUN_MODE` or `.ghostrun.yaml`):
 
 | Mode | Behavior |
 | :--- | :--- |
@@ -39,7 +39,7 @@ test passes:
 
 In the bundled example this takes a run from **23.3s to 2.7s**, and in
 `--ghostrun-replay` no model is invoked at all. Disable with `judge.cache: false`
-or `ghostrun_JUDGE_CACHE=false` if you want a live grade every run.
+or `GHOSTRUN_JUDGE_CACHE=false` if you want a live grade every run.
 
 ## Providers
 

examples/README.md

@@ -37,7 +37,7 @@ When no real judge is available (echo, or Ollama down / model not pulled), the
 example **skips** rather than fails:
 
 ```bash
-ghostrun_JUDGE=echo pytest examples/test_support_reply.py   # -> skipped
+GHOSTRUN_JUDGE=echo pytest examples/test_support_reply.py   # -> skipped
 ```
 
 For real semantic grading, use Ollama (above) or point the judge at a cloud
@@ -47,10 +47,10 @@ model in `.ghostrun.yaml`.
 
 [`test_live_smoke.py`](test_live_smoke.py) is the one check the mock-based suite
 can't cover: recording from the **real** OpenAI API and replaying it offline.
-It's opt-in — skipped unless you provide a key and set `ghostrun_LIVE=1`:
+It's opt-in — skipped unless you provide a key and set `GHOSTRUN_LIVE=1`:
 
 ```bash
-OPENAI_API_KEY=sk-... ghostrun_LIVE=1 pytest examples/test_live_smoke.py -v
+OPENAI_API_KEY=sk-... GHOSTRUN_LIVE=1 pytest examples/test_live_smoke.py -v
 ```
 
 It costs one cheap `gpt-4o-mini` call the first time, writes the response to

examples/test_live_smoke.py

@@ -3,9 +3,9 @@
 
 This is opt-in. It is skipped unless BOTH are true:
   * OPENAI_API_KEY is set
-  * ghostrun_LIVE=1
+  * GHOSTRUN_LIVE=1
 
-    OPENAI_API_KEY=sk-... ghostrun_LIVE=1 pytest examples/test_live_smoke.py -v
+    OPENAI_API_KEY=sk-... GHOSTRUN_LIVE=1 pytest examples/test_live_smoke.py -v
 
 What it proves, end to end:
   1. record  -> one real call to api.openai.com is captured to the cache
@@ -29,8 +29,8 @@
 from ghostrun.interceptor import CacheMiss, Interceptor
 
 pytestmark = pytest.mark.skipif(
-    not (os.environ.get("OPENAI_API_KEY") and os.environ.get("ghostrun_LIVE") == "1"),
-    reason="live test: set OPENAI_API_KEY and ghostrun_LIVE=1 to run",
+    not (os.environ.get("OPENAI_API_KEY") and os.environ.get("GHOSTRUN_LIVE") == "1"),
+    reason="live test: set OPENAI_API_KEY and GHOSTRUN_LIVE=1 to run",
 )
 
 CACHE_DIR = str(Path(__file__).with_name(".ghostrun_live_cache"))

examples/test_support_reply.py

@@ -6,7 +6,7 @@
     pytest examples/test_support_reply.py
 
 The semantic assertions are graded by a local Ollama model (default). If a real
-judge isn't available (no Ollama, or ghostrun_JUDGE=echo), the test skips rather
+judge isn't available (no Ollama, or GHOSTRUN_JUDGE=echo), the test skips rather
 than fails — see examples/conftest.py.
 
 To re-record against the real OpenAI API instead:

ghostrun/config.py

@@ -3,7 +3,7 @@
 Resolution order (lowest to highest precedence):
   1. Built-in defaults
   2. ``.ghostrun.yaml`` found by walking up from the current working directory
-  3. Environment variables (``ghostrun_*``)
+  3. Environment variables (``GHOSTRUN_*``)
 
 Nothing here talks to the network; a missing config file is fine.
 """
@@ -93,22 +93,22 @@ def _from_file(path: Path) -> dict:
 def _from_env() -> dict:
     env = os.environ
     out: dict = {}
-    if "ghostrun_MODE" in env:
-        out["mode"] = env["ghostrun_MODE"]
-    if "ghostrun_CACHE_DIR" in env:
-        out["cache_dir"] = env["ghostrun_CACHE_DIR"]
-    if "ghostrun_JUDGE" in env:
-        out["judge"] = env["ghostrun_JUDGE"]
-    if "ghostrun_JUDGE_MODEL" in env:
-        out["judge_model"] = env["ghostrun_JUDGE_MODEL"]
-    if "ghostrun_JUDGE_BASE_URL" in env:
-        out["judge_base_url"] = env["ghostrun_JUDGE_BASE_URL"]
-    if "ghostrun_JUDGE_TIMEOUT" in env:
-        out["judge_timeout"] = float(env["ghostrun_JUDGE_TIMEOUT"])
-    if "ghostrun_JUDGE_CACHE" in env:
-        out["judge_cache"] = env["ghostrun_JUDGE_CACHE"].strip().lower() not in ("0", "false", "no")
-    if "ghostrun_JUDGE_VOTES" in env:
-        out["judge_votes"] = int(env["ghostrun_JUDGE_VOTES"])
+    if "GHOSTRUN_MODE" in env:
+        out["mode"] = env["GHOSTRUN_MODE"]
+    if "GHOSTRUN_CACHE_DIR" in env:
+        out["cache_dir"] = env["GHOSTRUN_CACHE_DIR"]
+    if "GHOSTRUN_JUDGE" in env:
+        out["judge"] = env["GHOSTRUN_JUDGE"]
+    if "GHOSTRUN_JUDGE_MODEL" in env:
+        out["judge_model"] = env["GHOSTRUN_JUDGE_MODEL"]
+    if "GHOSTRUN_JUDGE_BASE_URL" in env:
+        out["judge_base_url"] = env["GHOSTRUN_JUDGE_BASE_URL"]
+    if "GHOSTRUN_JUDGE_TIMEOUT" in env:
+        out["judge_timeout"] = float(env["GHOSTRUN_JUDGE_TIMEOUT"])
+    if "GHOSTRUN_JUDGE_CACHE" in env:
+        out["judge_cache"] = env["GHOSTRUN_JUDGE_CACHE"].strip().lower() not in ("0", "false", "no")
+    if "GHOSTRUN_JUDGE_VOTES" in env:
+        out["judge_votes"] = int(env["GHOSTRUN_JUDGE_VOTES"])
     return out
 
 

ghostrun/interceptor.py

@@ -72,7 +72,7 @@ def handle_request(self, request: httpx.Request) -> httpx.Response:
         if self._mode == "replay":
             raise CacheMiss(
                 f"No cached response for {request.method} {request.url} "
-                f"(key {key}). Re-run with ghostrun_MODE=record to capture it."
+                f"(key {key}). Re-run with GHOSTRUN_MODE=record to capture it."
             )
 
         response = self._inner.handle_request(request)
@@ -102,7 +102,7 @@ async def handle_async_request(self, request: httpx.Request) -> httpx.Response:
         if self._mode == "replay":
             raise CacheMiss(
                 f"No cached response for {request.method} {request.url} "
-                f"(key {key}). Re-run with ghostrun_MODE=record to capture it."
+                f"(key {key}). Re-run with GHOSTRUN_MODE=record to capture it."
             )
 
         response = await self._inner.handle_async_request(request)