transcript-viewer 0.8.1__tar.gz → 0.10.0__tar.gz

{transcript_viewer-0.8.1 → transcript_viewer-0.10.0}/PKG-INFO

@@ -1,10 +1,10 @@
 Metadata-Version: 2.5
 Name: transcript-viewer
-Version: 0.8.1
+Version: 0.10.0
 Summary: Browse agent transcripts in a local, dependency-free web viewer.
 License: MIT
 Requires-Python: >=3.12
-Requires-Dist: atif-make>=0.5.0
+Requires-Dist: atif-make>=0.7.0
 Provides-Extra: ai
 Requires-Dist: anthropic>=0.40; extra == 'ai'
 Provides-Extra: parquet
@@ -162,7 +162,13 @@ of yours, keeps its file.
 
 Some datasets ship as a single file holding many runs — ATBench publishes a
 thousand agent trajectories as one JSON array, METR's as Parquet shards
-(install with `uv tool install "transcript-viewer[parquet]"` to read those). Those are opened rather than
+(install with `uv tool install "transcript-viewer[parquet]"` to read those),
+and an Inspect `.eval` log holds one run per sample.
+
+A fetch takes at most 32 GB at once. That is a backstop, not a judgement:
+published corpora are routinely several gigabytes, and what actually guards
+against an accident is the picker, which measures what you have ticked and
+shows it before anything downloads. Those are opened rather than
 listed: the file is split into a transcript apiece and each is indexed on its
 own, so a download of one file becomes a thousand sessions you can read. The
 pieces are kept beside what they came from, so removing the folder removes them

{transcript_viewer-0.8.1 → transcript_viewer-0.10.0}/README.md

@@ -149,7 +149,13 @@ of yours, keeps its file.
 
 Some datasets ship as a single file holding many runs — ATBench publishes a
 thousand agent trajectories as one JSON array, METR's as Parquet shards
-(install with `uv tool install "transcript-viewer[parquet]"` to read those). Those are opened rather than
+(install with `uv tool install "transcript-viewer[parquet]"` to read those),
+and an Inspect `.eval` log holds one run per sample.
+
+A fetch takes at most 32 GB at once. That is a backstop, not a judgement:
+published corpora are routinely several gigabytes, and what actually guards
+against an accident is the picker, which measures what you have ticked and
+shows it before anything downloads. Those are opened rather than
 listed: the file is split into a transcript apiece and each is indexed on its
 own, so a download of one file becomes a thousand sessions you can read. The
 pieces are kept beside what they came from, so removing the folder removes them

{transcript_viewer-0.8.1 → transcript_viewer-0.10.0}/pyproject.toml

@@ -1,6 +1,6 @@
 [project]
 name = "transcript-viewer"
-version = "0.8.1"
+version = "0.10.0"
 description = "Browse agent transcripts in a local, dependency-free web viewer."
 readme = "README.md"
 requires-python = ">=3.12"
@@ -10,7 +10,7 @@ license = { text = "MIT" }
 # The floor is not politeness: this viewer imports atif_make.container, which
 # is how a benchmark published as one JSON array of a thousand runs becomes a
 # thousand sessions. An older atif-make fails to import at all.
-dependencies = ["atif-make>=0.5.0"]
+dependencies = ["atif-make>=0.7.0"]
 
 [project.optional-dependencies]
 # Claude-backed explanations are opt-in, so the default install stays

{transcript_viewer-0.8.1 → transcript_viewer-0.10.0}/src/transcript_viewer/corpus.py

@@ -312,8 +312,10 @@ def scan(roots: list[Path] | None = None, origin: str = "scanned") -> list[Entry
                 sorted(where.rglob("*.jsonl"))
                 + sorted(where.rglob("*.har"))
                 + sorted(where.rglob("*.json"))
+                + sorted(where.rglob("*.md"))
                 # Not a log, but a container of them, opened above.
                 + sorted(where.rglob("*.parquet"))
+                + sorted(where.rglob("*.eval"))
             )
         for path in candidates:
             resolved = path.resolve()

{transcript_viewer-0.8.1 → transcript_viewer-0.10.0}/src/transcript_viewer/fetch.py

@@ -58,10 +58,13 @@ HOSTS = {
 # in: a bucket of agent transcripts is far more likely to hold one zip per
 # session than loose JSONL, so excluding archives made S3 useless for exactly
 # the case it was added for.
-LOGS = (".jsonl", ".json", ".har")
+# A transcript is not always JSON: some corpora ship the readable rendering.
+LOGS = (".jsonl", ".json", ".har", ".md", ".markdown")
 # A published dataset is a container of runs rather than a log, but it is still
 # a thing worth fetching: METR's transcripts come as Parquet shards.
-DATASETS = (".parquet",)
+# An Inspect .eval is a zip of JSON, and a corpus of them is where most
+# control research keeps its transcripts.
+DATASETS = (".parquet", ".eval")
 ARCHIVES = (".zip", ".tgz", ".tar", ".tar.gz", ".tar.bz2", ".tar.xz", ".gz")
 SUFFIXES = LOGS + DATASETS + ARCHIVES
 
@@ -87,7 +90,13 @@ AWS_TIMEOUT = 300
 BROWSE_LIMIT = 3_000
 
 MAX_FILES = 2_000
-MAX_TOTAL_BYTES = 2 * 1024**3
+# A ceiling on one fetch, not a judgement about what is worth having: published
+# corpora are routinely several gigabytes — MonitoringBench ships as a single
+# 2.5 GB archive — and two was refusing them. What still guards against an
+# accident is the picker, which measures a selection and shows it before
+# anything downloads; this is the backstop for the bucket nobody meant to take
+# whole, which in practice is the 144 GB one.
+MAX_TOTAL_BYTES = 32 * 1024**3
 TIMEOUT = 60
 
 
@@ -578,7 +587,8 @@ def _sized(plan: "Plan") -> "Plan":
     if known > MAX_TOTAL_BYTES:
         raise FetchError(
             f"That is {known / 1024**3:.1f} GB, over the "
-            f"{MAX_TOTAL_BYTES // 1024**3} GB limit."
+            f"{MAX_TOTAL_BYTES // 1024**3} GB limit for one fetch. "
+            f"Pick fewer folders, or take it in parts."
         )
     return plan
 

{transcript_viewer-0.8.1 → transcript_viewer-0.10.0}/src/transcript_viewer/page.html

@@ -1728,6 +1728,12 @@ const EXAMPLES=[
   {name:"Redwood agent transcripts",
    url:"s3://rr-agent-transcripts",
    note:"needs an AWS profile set in Settings"},
+  {name:"Redwood sabotage trajectories",
+   url:"s3://rr-strajs",
+   note:"136 transcripts, honest and attack in pairs — needs an AWS profile"},
+  {name:"MonitoringBench",
+   url:"https://huggingface.co/datasets/neur26anonsub/ctrldataset2026",
+   note:"2,644 attack runs as Inspect logs — the archive alone is 2.5 GB"},
 ];
 
 function drawExamples(){

{transcript_viewer-0.8.1 → transcript_viewer-0.10.0}/tests/test_fetch.py

@@ -215,7 +215,11 @@ def test_a_dataset_url_lists_its_convertible_files(monkeypatch):
     )
     assert service == "hf"
     assert label == "owner--name"
-    assert [f.name for f in files] == ["a/transcript.jsonl", "b/metadata.json"]
+    assert [f.name for f in files] == [
+        "a/transcript.jsonl",
+        "a/description.md",
+        "b/metadata.json",
+    ]
     assert all("/resolve/main/" in f.url for f in files)
     assert seen["auth"] == ["tok"], "the token was not sent"
 
@@ -310,11 +314,23 @@ def test_a_blob_url_is_rewritten_to_raw(monkeypatch):
 
 
 def test_nothing_convertible_is_said_plainly(monkeypatch):
-    _stub(monkeypatch, {"api/datasets": [{"type": "file", "path": "readme.md"}]})
+    # Markdown is fetchable — some corpora publish transcripts as nothing else —
+    # so the refusal is tested with something that could not be one.
+    _stub(monkeypatch, {"api/datasets": [{"type": "file", "path": "diagram.png"}]})
     with pytest.raises(fetch.FetchError, match="Nothing convertible"):
         fetch.plan("https://huggingface.co/datasets/owner/name", {})
 
 
+def test_a_transcript_published_as_markdown_is_worth_fetching(monkeypatch):
+    """A bucket of 136 of these was unreachable while .md was filtered out."""
+    _stub(
+        monkeypatch,
+        {"api/datasets": [{"type": "file", "path": "run/transcript_attack.md"}]},
+    )
+    plan = fetch.plan("https://huggingface.co/datasets/owner/name", {})
+    assert [f.name for f in plan.files] == ["run/transcript_attack.md"]
+
+
 def test_too_many_files_is_refused_before_downloading(monkeypatch):
     _stub(
         monkeypatch,
@@ -489,7 +505,7 @@ def test_an_s3_url_lists_what_is_convertible(monkeypatch):
     plan = fetch.plan("s3://rr-agent-transcripts/runs", {"aws": "rw-eng"})
     assert plan.service == "s3"
     assert plan.label == "rr-agent-transcripts"
-    assert [f.name for f in plan.files] == ["a.jsonl", "deep/b.jsonl"]
+    assert [f.name for f in plan.files] == ["a.jsonl", "notes.md", "deep/b.jsonl"]
     assert plan.web == "s3://rr-agent-transcripts/runs"
 
 
@@ -604,7 +620,10 @@ def test_an_archive_counts_as_something_worth_fetching(monkeypatch):
         ]
     }))
     plan = fetch.plan("s3://rr-agent-transcripts/chippy/abc", {"aws": "rw-eng"})
-    assert [f.name for f in plan.files] == ["transcripts.zip"]
+    # The Markdown comes too: a corpus that publishes transcripts as Markdown
+    # and nothing else was unreachable while it was filtered out. One that is
+    # not a transcript simply does not index.
+    assert [f.name for f in plan.files] == ["transcripts.zip", "notes.md"]
 
 
 def test_the_listing_is_capped_rather_than_exhaustive(monkeypatch):
@@ -650,7 +669,8 @@ def test_s3_browsing_returns_folders_and_files(monkeypatch):
     nodes = fetch.browse("s3://bucket", "chippy", {"aws": "rw-eng"})
     assert "--delimiter" in seen["command"]
     assert [(n.kind, n.name) for n in nodes] == [
-        ("folder", "abc"), ("folder", "def"), ("file", "loose.zip"),
+        ("folder", "abc"), ("folder", "def"),
+        ("file", "loose.zip"), ("file", "readme.md"),
     ]
     assert [n.path for n in nodes][0] == "chippy/abc"
 
@@ -667,7 +687,9 @@ def test_github_browsing_reads_one_level(monkeypatch):
         {"type": "file", "path": "logs/notes.md", "size": 3},
     ]})
     nodes = fetch.browse("https://github.com/owner/name/tree/main/logs", "", {})
-    assert [(n.kind, n.name) for n in nodes] == [("folder", "inner"), ("file", "a.jsonl")]
+    assert [(n.kind, n.name) for n in nodes] == [
+        ("folder", "inner"), ("file", "a.jsonl"), ("file", "notes.md"),
+    ]
 
 
 def test_hf_browsing_reads_one_level(monkeypatch):

{transcript_viewer-0.8.1 → transcript_viewer-0.10.0}/uv.lock

@@ -44,11 +44,11 @@ wheels = [
 
 [[package]]
 name = "atif-make"
-version = "0.5.0"
+version = "0.7.0"
 source = { registry = "https://pypi.org/simple" }
-sdist = { url = "https://files.pythonhosted.org/packages/60/2c/9e5d1b1ba278fde5135b8066487e029dd0620161b6a220f5cb2a118a0333/atif_make-0.5.0.tar.gz", hash = "sha256:d4c3122a63b553099f0843d5d1c666136b0bccb1c11e8ebd2b4f9c4a9200cb5e", size = 215842, upload-time = "2026-08-23T05:15:04.439Z" }
+sdist = { url = "https://files.pythonhosted.org/packages/6a/ed/b1fbadcc65d59186442517059beffe7c4a590be85d17308d9143278a2104/atif_make-0.7.0.tar.gz", hash = "sha256:858966e102e772f35ffc3fdcd25a11cb52759190233b41cead3878dc607ff4ce", size = 350496, upload-time = "2026-08-24T21:49:42.323Z" }
 wheels = [
-    { url = "https://files.pythonhosted.org/packages/00/47/f9a09d10b3d53600c43fa1b22c1d01cb731e628b3572c2e40109de195cfb/atif_make-0.5.0-py3-none-any.whl", hash = "sha256:b52fac5489f5cbc34f947f5d8b21b51a6a04276420d908a96f938e1bf7443791", size = 49769, upload-time = "2026-08-23T05:15:03.265Z" },
+    { url = "https://files.pythonhosted.org/packages/14/7b/d135b4b2bec98599cb65166f278487f87d9842492c2f13e8a1385c6a3e85/atif_make-0.7.0-py3-none-any.whl", hash = "sha256:ef1601b5d2ac3e07b9e9e97ee0a2e0856b953b4ba317633e6f590747b450659e", size = 63987, upload-time = "2026-08-24T21:49:41.087Z" },
 ]
 
 [package.optional-dependencies]
@@ -388,7 +388,7 @@ wheels = [
 
 [[package]]
 name = "transcript-viewer"
-version = "0.8.1"
+version = "0.10.0"
 source = { editable = "." }
 dependencies = [
     { name = "atif-make" },
@@ -410,7 +410,7 @@ dev = [
 [package.metadata]
 requires-dist = [
     { name = "anthropic", marker = "extra == 'ai'", specifier = ">=0.40" },
-    { name = "atif-make", specifier = ">=0.5.0" },
+    { name = "atif-make", specifier = ">=0.7.0" },
     { name = "atif-make", extras = ["parquet"], marker = "extra == 'parquet'", specifier = ">=0.5.0" },
 ]
 provides-extras = ["parquet", "ai"]