benchflow 0.7.2.dev1986__py3-none-any.whl → 0.7.2.dev1994__py3-none-any.whl

benchflow/cli/main.py

@@ -1228,11 +1228,27 @@ def eval_view(
         ),
     ],
     port: Annotated[int, typer.Option(help="Server port")] = 8888,
+    confirm: Annotated[
+        bool,
+        typer.Option(
+            "--confirm",
+            help=(
+                "Show an Approve & submit / Not this one bar on the page; "
+                "print DECISION: approved|rejected and exit 0 on approve, "
+                "3 on reject."
+            ),
+        ),
+    ] = False,
 ) -> None:
     """View a trial or session trajectory in the browser."""
-    from benchflow.trajectories.viewer import serve
-
-    serve(str(rollout_dir), port)
+    from benchflow.trajectories import viewer
+
+    decision = viewer.serve(str(rollout_dir), port, confirm=confirm)
+    # Exit-code contract for --confirm: 0 approved, 3 rejected. 3 is chosen
+    # so a rejection never collides with the CLI's existing error exits
+    # (1 = errors, 2 = Typer usage errors).
+    if decision == "rejected":
+        raise typer.Exit(3)
 
 
 # ── Command-group wiring ──────────────────────────────────────────────

benchflow/cli/traj.py

@@ -54,14 +54,26 @@ SKILL_RAW_URL = (
     "/.agents/skills/benchflow-traj-upload/SKILL.md"
 )
 
+UPGRADE_COMMAND = "uv tool install --python 3.12 --upgrade --force benchflow"
+
+# Three unwrapped lines separated by blank lines: each logical unit is one
+# full physical line (no hard wraps mid-sentence), so the URL and the upgrade
+# command stay selectable in agents and terminals.
 CONTRIBUTOR_PROMPT = (
-    "Submit my best local Claude Code or Codex session to the BenchFlow eval "
-    "prize. Read "
-    f"{SKILL_RAW_URL} "
-    "and follow it: find a session, open the viewer, upload only after I review it."
+    "Submit my relevant local Claude Code, Codex, OpenCode, or Cursor session "
+    "from the re:Agent e2e agentic science / ai4bio hackathon (last 72 hours, "
+    "on this laptop) to the BenchFlow eval prize."
+    "\n\n"
+    f"1. First make sure the latest benchflow CLI is installed: {UPGRADE_COMMAND}"
+    "\n\n"
+    f"2. Then read {SKILL_RAW_URL} "
+    "and follow it: find a session, open the viewer, and upload only after "
+    "I (the human) review it."
 )
 
-UPGRADE_COMMAND = "uv tool install --python 3.12 --upgrade --force benchflow"
+CONTRIBUTOR_PROMPT_FRAMING = (
+    "Send this to your coding agent (it's a prompt for the agent, not steps for you):"
+)
 
 
 def _fetch_latest_version() -> str | None:
@@ -166,14 +178,14 @@ def register_traj(app: typer.Typer) -> None:
         ] = False,
         prompt_only: Annotated[
             bool,
-            typer.Option("--prompt", help="Print the copy-paste agent line and exit"),
+            typer.Option("--prompt", help="Print the copy-paste agent prompt and exit"),
         ] = False,
         list_sessions: Annotated[
             bool,
             typer.Option("--list", help="List recent local sessions and exit"),
         ] = False,
     ) -> None:
-        """Install the submit skill, or print the line to paste into an agent."""
+        """Install the submit skill, or print the prompt to send to an agent."""
         _maybe_print_update_hint()
         if prompt_only:
             _print_contributor_prompt()
@@ -194,7 +206,6 @@ def register_traj(app: typer.Typer) -> None:
                 _run_npx_skill_install()
         else:
             _install_project_skill(Path.cwd())
-        console.print("Paste this to your agent:")
         _print_contributor_prompt()
         if interactive and typer.confirm(
             "List recent sessions and open the viewer now?", default=False
@@ -625,7 +636,9 @@ def _print_dry_run(staged: StagedCapture) -> None:
 
 
 def _print_contributor_prompt() -> None:
-    # One physical line so people can copy it. Rich wrapping would break that.
+    # Plain print: Rich wrapping would break the unbroken URL line and make
+    # the block awkward to copy.
+    print(CONTRIBUTOR_PROMPT_FRAMING)
     print(CONTRIBUTOR_PROMPT)
 
 
@@ -671,7 +684,7 @@ def _run_npx_skill_install() -> None:
         check=False,
     )
     if completed.returncode != 0:
-        print_error("npx skills add failed; the copy-paste line below still works.")
+        print_error("npx skills add failed; the copy-paste prompt below still works.")
 
 
 def _print_session_hits() -> None:

benchflow/trajectories/viewer.py

@@ -16,9 +16,15 @@ _RESULT_PREVIEW = 300  # max chars for result summary
 
 # Shared stylesheet for all viewer pages, matching the www.benchflow.ai design
 # language (light monochrome: near-white page, white cards, near-black ink,
-# pure-black accent, Satoshi/Google Sans Code with system-safe fallbacks, dark
-# code blocks). Kept inline so pages work fully offline with no external
-# requests; one constant so the three templates stop drifting apart.
+# Satoshi/Google Sans Code with system-safe fallbacks). Kept inline so pages
+# work fully offline with no external requests; one constant so the three
+# templates stop drifting apart.
+#
+# The site itself is deliberately achromatic, so tool-call accents below are
+# muted GitHub-label-style pastels (pale chip background + darker same-hue
+# text + a soft left border strip) that read as annotations rather than
+# fighting the monochrome base. The dark #141414 code treatment is reserved
+# for terminal output of shell commands; everything else stays light.
 _VIEWER_CSS = """\
 * { margin: 0; padding: 0; box-sizing: border-box; }
 :root {
@@ -46,8 +52,10 @@ body { font-family: var(--font-sans); background: var(--background); color: var(
 .meta span { font-family: var(--font-mono); font-size: 11px; font-weight: 500; color: var(--ink-secondary); background: var(--secondary); padding: 3px 9px; border-radius: 4px; border: 1px solid var(--border); }
 .step { margin-bottom: 8px; padding: 12px 16px; border-radius: var(--radius); background: var(--card); border: 1px solid var(--border); box-shadow: 0 1px 2px rgba(10, 10, 10, 0.04); }
 .step.prompt { background: var(--secondary); border-color: var(--rule-strong); margin-bottom: 14px; }
-.step.output { background: var(--code-bg); border-color: var(--code-bg); padding: 10px 16px; }
-.step.output pre { color: var(--code-ink); font-family: var(--font-mono); font-size: 12px; line-height: 1.7; white-space: pre-wrap; word-break: break-word; }
+.step.output { background: var(--card); border-color: var(--border); padding: 10px 16px; }
+.step.output pre { color: var(--ink-secondary); font-family: var(--font-mono); font-size: 12px; line-height: 1.7; white-space: pre-wrap; word-break: break-word; }
+.step.output.term { background: var(--code-bg); border-color: var(--code-bg); }
+.step.output.term pre { color: var(--code-ink); }
 .step.result { background: var(--ink); border-color: var(--ink); margin-top: 14px; }
 .step.result .msg { color: var(--background); }
 .step-header { display: flex; align-items: center; gap: 8px; margin-bottom: 8px; }
@@ -59,12 +67,70 @@ body { font-family: var(--font-sans); background: var(--background); color: var(
 .msg { font-size: 14px; line-height: 1.65; white-space: pre-wrap; word-break: break-word; }
 .thinking { font-size: 13px; color: var(--muted); font-style: italic; margin-bottom: 8px; padding: 8px 12px; background: var(--secondary); border-radius: 4px; border-left: 2px solid var(--rule-strong); white-space: pre-wrap; word-break: break-word; }
 .tool { margin-bottom: 6px; }
-.tool-name { display: inline-flex; align-items: center; font-family: var(--font-mono); font-size: 11px; font-weight: 600; color: var(--ink); background: var(--secondary); border: 1px solid var(--border); padding: 2px 9px; border-radius: 4px; }
-.tool-args { margin-top: 6px; font-family: var(--font-mono); font-size: 12px; line-height: 1.7; color: var(--code-ink); background: var(--code-bg); padding: 10px 12px; border-radius: 6px; white-space: pre-wrap; word-break: break-word; }
+.tool-name { display: inline-flex; align-items: center; font-family: var(--font-mono); font-size: 11px; font-weight: 600; color: var(--acc-ink, var(--ink)); background: var(--acc-bg, var(--secondary)); border: 1px solid var(--acc-line, var(--border)); padding: 2px 9px; border-radius: 4px; }
+.tool-args { margin-top: 6px; font-family: var(--font-mono); font-size: 12px; line-height: 1.7; color: var(--ink-secondary); background: var(--secondary); border: 1px solid var(--border); padding: 10px 12px; border-radius: 6px; white-space: pre-wrap; word-break: break-word; }
+.step.tool-step { border-left: 3px solid var(--acc-strip, var(--rule-strong)); }
+.acc-bash  { --acc-bg: #f7efda; --acc-line: #ecdcb2; --acc-ink: #8a5a12; --acc-strip: #dcb45e; }
+.acc-edit  { --acc-bg: #e8f0fa; --acc-line: #d0dff1; --acc-ink: #1d4e89; --acc-strip: #7fa8d8; }
+.acc-read  { --acc-bg: #e5f2ec; --acc-line: #cbe3d7; --acc-ink: #1a6b52; --acc-strip: #74bda0; }
+.acc-agent { --acc-bg: #efeaf8; --acc-line: #ded3ef; --acc-ink: #5b3e96; --acc-strip: #a78fd6; }
+.acc-web   { --acc-bg: #e3f1f6; --acc-line: #c8e2ea; --acc-ink: #176478; --acc-strip: #6fb6ca; }
+.acc-other { --acc-bg: var(--secondary); --acc-line: var(--border); --acc-ink: var(--ink-secondary); --acc-strip: var(--rule-strong); }
 .metrics { font-family: var(--font-mono); font-size: 11px; color: var(--faint); margin-top: 4px; }
 .turn-divider { border-top: 1px solid var(--border); margin: 24px 0; }
 """
 
+# Sticky bottom confirmation bar injected only in --confirm mode. Styling
+# reuses the page's CSS variables (white surface, top border, ink text, pill
+# buttons) so it reads as part of the same www.benchflow.ai design language.
+# The <style> block rides along with the snippet so plain (non-confirm) pages
+# carry zero confirm markup or CSS.
+_CONFIRM_BAR_HTML = """\
+<style>
+body { padding-bottom: 104px; }
+.confirm-bar { position: fixed; bottom: 0; left: 0; right: 0; background: var(--card); border-top: 1px solid var(--border); box-shadow: 0 -1px 3px rgba(10, 10, 10, 0.05); z-index: 10; }
+.confirm-inner { max-width: 960px; margin: 0 auto; padding: 14px 20px; display: flex; align-items: center; justify-content: space-between; gap: 16px; flex-wrap: wrap; }
+.confirm-question { font-size: 14px; font-weight: 500; color: var(--ink); }
+.confirm-actions { display: flex; gap: 8px; flex: none; }
+.confirm-btn { font-family: var(--font-sans); font-size: 13px; font-weight: 600; letter-spacing: -0.01em; padding: 8px 18px; border-radius: 9999px; cursor: pointer; transition: background 0.15s ease; }
+.confirm-btn.approve { background: var(--ink); color: var(--background); border: 1px solid var(--ink); }
+.confirm-btn.approve:hover { background: #262626; }
+.confirm-btn.reject { background: var(--card); color: var(--ink); border: 1px solid var(--rule-strong); }
+.confirm-btn.reject:hover { background: var(--secondary); }
+</style>
+<div class="confirm-bar">
+<div class="confirm-inner" id="confirm-inner">
+<span class="confirm-question">Submit this trajectory to the BenchFlow eval prize?</span>
+<div class="confirm-actions">
+<button class="confirm-btn approve" onclick="__benchDecide('approve')">Approve &amp; submit</button>
+<button class="confirm-btn reject" onclick="__benchDecide('reject')">Not this one</button>
+</div>
+</div>
+</div>
+<script>
+async function __benchDecide(choice) {
+  try {
+    await fetch("/decision", { method: "POST", body: choice });
+  } catch (err) {
+    // The server exits right after answering; a dropped socket is still a
+    // delivered decision, so fall through to the confirmation note.
+  }
+  document.getElementById("confirm-inner").innerHTML =
+    choice === "approve"
+      ? '<span class="confirm-question">Approved — go back to your agent.</span>'
+      : '<span class="confirm-question">Rejected — tell your agent which session instead.</span>';
+}
+</script>
+"""
+
+
+def _inject_confirm_bar(page: str) -> str:
+    """Append the confirm bar just before </body> (or at the end as fallback)."""
+    if "</body>" in page:
+        return page.replace("</body>", f"{_CONFIRM_BAR_HTML}</body>", 1)
+    return page + _CONFIRM_BAR_HTML
+
+
 # Small BenchFlow wordmark header (inline SVG logo from the site's icon set —
 # no external requests) shown above the page title on every viewer page.
 _WORDMARK_HTML = (
@@ -87,6 +153,28 @@ _WORDMARK_HTML = (
 )
 
 
+# Tool kind → accent class. Substring matching so it covers Claude Code tool
+# names (Bash, Write, Edit, Read, Agent, WebSearch, ...), ACP kinds (execute,
+# edit, read, search, fetch, ...), and Codex function names (shell, ...).
+# Order matters: earlier entries win (e.g. "websearch" hits web before read).
+_TOOL_ACCENTS: tuple[tuple[str, tuple[str, ...]], ...] = (
+    ("acc-web", ("web", "search", "fetch", "grep", "glob", "browser")),
+    ("acc-bash", ("bash", "shell", "exec", "terminal", "command")),
+    ("acc-edit", ("write", "edit", "patch", "delete", "move", "notebook")),
+    ("acc-read", ("read", "cat", "view", "ls", "list")),
+    ("acc-agent", ("agent", "task", "skill", "oracle")),
+)
+
+
+def _tool_accent_class(name: str) -> str:
+    """Map a tool name/kind to its accent CSS class (presentation only)."""
+    lowered = name.lower()
+    for css_class, needles in _TOOL_ACCENTS:
+        if any(needle in lowered for needle in needles):
+            return css_class
+    return "acc-other"
+
+
 def render_turn(events: list[dict], turn_number: int, prompt: str = "") -> str:
     """Render one turn's events as HTML blocks."""
     blocks = []
@@ -103,6 +191,9 @@ def render_turn(events: list[dict], turn_number: int, prompt: str = "") -> str:
     # Group: thinking → text → tool_use → tool_result → thinking → ...
     pending_thinking = ""
     pending_text = ""
+    # tool_use id → accent class, so each tool_result can match its call's
+    # accent and shell results alone keep the dark terminal treatment.
+    accent_by_tool_id: dict[str, str] = {}
 
     for event in events:
         etype = event.get("type", "")
@@ -160,7 +251,13 @@ def render_turn(events: list[dict], turn_number: int, prompt: str = "") -> str:
                         f"</div>"
                     )
 
-                    blocks.append(f'<div class="step agent">{"".join(parts)}</div>')
+                    accent = _tool_accent_class(block.get("name", ""))
+                    if block.get("id"):
+                        accent_by_tool_id[str(block["id"])] = accent
+                    blocks.append(
+                        f'<div class="step agent tool-step {accent}">'
+                        f"{''.join(parts)}</div>"
+                    )
 
         elif etype == "user":
             content = event.get("message", {}).get("content", "")
@@ -181,8 +278,13 @@ def render_turn(events: list[dict], turn_number: int, prompt: str = "") -> str:
                             display = "[binary content]"
                         else:
                             display = html.escape(raw[:400])
+                        accent = accent_by_tool_id.get(
+                            str(block.get("tool_use_id", "")), "acc-other"
+                        )
+                        term = " term" if accent == "acc-bash" else ""
                         blocks.append(
-                            f'<div class="step output"><pre>{display}</pre></div>'
+                            f'<div class="step output tool-step {accent}{term}">'
+                            f"<pre>{display}</pre></div>"
                         )
                     elif block.get("type") == "text" and block.get("text"):
                         texts.append(str(block["text"]))
@@ -408,8 +510,13 @@ def _render_acp_events(
             kind = html.escape(event.get("kind", ""))
             event_title = html.escape(event.get("title", ""))
             status = event.get("status", "")
+            # ACP kinds are coarse ("other" covers Skill/Task/...); when the
+            # kind doesn't map, fall back to the human title for the accent.
+            accent = _tool_accent_class(event.get("kind", ""))
+            if accent == "acc-other":
+                accent = _tool_accent_class(event.get("title", ""))
             blocks.append(
-                f'<div class="step agent">'
+                f'<div class="step agent tool-step {accent}">'
                 f'<div class="tool"><span class="tool-name">{kind}</span> {event_title}</div>'
                 f'<div class="metrics">{status}</div>'
                 f"</div>"
@@ -578,9 +685,21 @@ def _stream_json_page(title: str, events: list[dict], turn_blocks: list[str]) ->
 
 
 def serve(
-    rollout_path: str, port: int = 8888, prompts: list[str] | None = None
-) -> None:
-    """Serve a trial directory or a session JSONL file as a web page."""
+    rollout_path: str,
+    port: int = 8888,
+    prompts: list[str] | None = None,
+    confirm: bool = False,
+) -> str | None:
+    """Serve a trial directory or a session JSONL file as a web page.
+
+    With ``confirm=True`` the page carries an Approve/Reject bar posting to
+    ``/decision``; the first valid decision shuts the server down and is
+    returned as ``"approved"`` or ``"rejected"`` after printing a
+    machine-readable ``DECISION: <value>`` line to stdout. Without it the
+    server runs until Ctrl+C and the return value is ``None`` — exactly the
+    pre-confirm behavior (no bar, no POST endpoint).
+    """
+    import threading
     from http.server import HTTPServer, SimpleHTTPRequestHandler
 
     path = Path(rollout_path)
@@ -600,11 +719,20 @@ def serve(
         print(f"No trajectories found in {path}")
         sys.exit(1)
     if write_sidecar:
+        # The sidecar stays the plain page: the confirm bar is a one-shot
+        # interaction against this live server, not part of the artifact.
         (path / "trajectory.html").write_text(html_content)
+    if confirm:
+        html_content = _inject_confirm_bar(html_content)
 
     print(f"Trajectory viewer: http://localhost:{port}")
     print(f"Trial: {path}")
-    print("Press Ctrl+C to stop\n")
+    if confirm:
+        print("Waiting for Approve / Not this one in the browser (Ctrl+C to stop)\n")
+    else:
+        print("Press Ctrl+C to stop\n")
+
+    decision: str | None = None
 
     class Handler(SimpleHTTPRequestHandler):
         def do_GET(self):
@@ -616,11 +744,44 @@ def serve(
         def log_message(self, format, *args):
             pass
 
-    server = HTTPServer(("localhost", port), Handler)
+    handler_cls: type[SimpleHTTPRequestHandler] = Handler
+
+    if confirm:
+
+        class ConfirmHandler(Handler):
+            def do_POST(self):
+                nonlocal decision
+                if self.path != "/decision":
+                    self.send_error(404)
+                    return
+                length = int(self.headers.get("Content-Length") or 0)
+                body = self.rfile.read(length).decode("utf-8", "replace").strip()
+                if body not in ("approve", "reject"):
+                    self.send_error(400, "Body must be 'approve' or 'reject'")
+                    return
+                decision = "approved" if body == "approve" else "rejected"
+                self.send_response(200)
+                self.send_header("Content-Type", "text/plain; charset=utf-8")
+                self.end_headers()
+                self.wfile.write(decision.encode())
+                # shutdown() blocks until serve_forever returns, and this
+                # handler runs inside that loop — stop from a helper thread.
+                threading.Thread(target=server.shutdown, daemon=True).start()
+
+        handler_cls = ConfirmHandler
+
+    server = HTTPServer(("localhost", port), handler_cls)
     try:
         server.serve_forever()
     except KeyboardInterrupt:
         print("\nStopped.")
+        return None
+    finally:
+        server.server_close()
+
+    if decision is not None:
+        print(f"DECISION: {decision}", flush=True)
+    return decision
 
 
 if __name__ == "__main__":

{benchflow-0.7.2.dev1986.dist-info → benchflow-0.7.2.dev1994.dist-info}/METADATA

@@ -1,6 +1,6 @@
 Metadata-Version: 2.5
 Name: benchflow
-Version: 0.7.2.dev1986
+Version: 0.7.2.dev1994
 Summary: Multi-turn agent benchmarking with ACP — run any agent, any model, any provider.
 Project-URL: Homepage, https://github.com/benchflow-ai/benchflow
 Project-URL: Repository, https://github.com/benchflow-ai/benchflow
@@ -89,20 +89,23 @@ BenchFlow is a universal environment framework: it runs AI agents against task e
 
 ## Quick start: 1. Submit a trajectory
 
-Proud of a Claude Code or Codex session? Copy this line into Claude Code,
-Codex, Cursor, or any coding agent. That is the $2,000 eval prize path. No
-BenchFlow account, API key, or Azure login.
+**Send these to your coding agent.** Copy
+the block below and paste it as a message to Claude Code, Codex, OpenCode, or
+Cursor. That is the $2,000 eval prize path. No BenchFlow account, API key, or
+Azure login.
 
-```
-Submit my best local Claude Code or Codex session to the BenchFlow eval prize. Read https://raw.githubusercontent.com/benchflow-ai/benchflow/main/.agents/skills/benchflow-traj-upload/SKILL.md and follow it: find a session, open the viewer, upload only after I review it.
-```
+> Submit my relevant local Claude Code, Codex, OpenCode, or Cursor session from the re:Agent e2e agentic science / ai4bio hackathon (last 72 hours, on this laptop) to the BenchFlow eval prize.
+>
+> 1. First make sure the latest benchflow CLI is installed: uv tool install --python 3.12 --upgrade --force benchflow
+>
+> 2. Then read https://raw.githubusercontent.com/benchflow-ai/benchflow/main/.agents/skills/benchflow-traj-upload/SKILL.md and follow it: find a session, open the viewer, and upload only after I (the human) review it.
 
 The agent finds sessions on this machine, opens the viewer, and uploads after
 you like what you see.
 
-Be on the latest BenchFlow first — `uv tool install --python 3.12 --upgrade
---force benchflow`. The `bench traj` commands print a one-line upgrade hint
-when a newer release is available.
+The prompt has the agent install the latest BenchFlow first (`uv tool
+install --python 3.12 --upgrade --force benchflow`). The `bench traj` commands
+also print a one-line upgrade hint when a newer release is available.
 
 Prefer the terminal instead? The guided upload inspects before anything leaves
 your machine — it renders a redacted trajectory report (step counts, masked
@@ -128,7 +131,7 @@ bench traj setup
 ```
 
 `npx skills add` asks which agents to install for. `bench traj setup` copies
-the skill into this project and prints the same paste line. See the
+the skill into this project and prints the same agent prompt. See the
 [upload skill](./.agents/skills/benchflow-traj-upload/SKILL.md).
 
 ## Quick start: 2. Run with a ChatGPT or Claude subscription

{benchflow-0.7.2.dev1986.dist-info → benchflow-0.7.2.dev1994.dist-info}/RECORD

@@ -108,7 +108,7 @@ benchflow/cli/environment.py,sha256=q3G8G2i0rs_wqGIWzULnnc6Nt8TlJLO12zuYlro-5-A,
 benchflow/cli/eval_artifacts.py,sha256=4CbS8aIaynCk2PUeWA4Qo31F8SH7_rxtwB7po8_O3v4,11275
 benchflow/cli/eval_lift.py,sha256=0Tyn9ZVIzAIeZ9vm0T7onBLre48gliKf0K8NrD7m1j0,3072
 benchflow/cli/hub.py,sha256=uX5O-q_9CMRb9VlW06HkA5EbKUa6xORnNNTNmMM2H_0,7029
-benchflow/cli/main.py,sha256=nSVz2r1Iq0_i7Mu4Mqx1-ok3jy_7C2k74nblROKe1UA,47711
+benchflow/cli/main.py,sha256=gjsxz90nwHaYq-SMpfzFG-zSbAzvxW6GZAy6JCbiWe4,48319
 benchflow/cli/monitor.py,sha256=-0C4OzZqgUj7Lvi5obJVsK-M81OZwaZNEpis71Kqyj4,3427
 benchflow/cli/review.py,sha256=ysR6No9oFu0TgcbucNG9YmXY5FkMUHobdr6F4rHQmQw,7892
 benchflow/cli/sandbox.py,sha256=QrLASbsfzPGVosW_s5XxE3_mcXqTzbU2ZnnC9MADpXM,9849
@@ -116,7 +116,7 @@ benchflow/cli/skills.py,sha256=2Apv1iqvmstWEhY6O-PRdxKc8RYwR6f7MvbYfl2dWm0,7497
 benchflow/cli/tasks.py,sha256=b1XP1EZ79ZtK_1shF65qZyv27EcKKFZYV35M4DzxYKI,18534
 benchflow/cli/trace_import.py,sha256=NHDieFdiKQ-2Sr1-GF8Z5EDcbGsz2qq_kFE0rXacC2o,14705
 benchflow/cli/train.py,sha256=MIP_JxptitH348bWWURpcWt-TmDAw1s8d-RUU7jBaIg,23635
-benchflow/cli/traj.py,sha256=2QlXCLNywIKsm8uQVsQ1TAEnaafV6b0dfLx37O2wEsU,23691
+benchflow/cli/traj.py,sha256=tX4mtAXbEwEhbwegUTHl_IBIFoxKtVlrwrlv8MPf764,24290
 benchflow/continue_run/__init__.py,sha256=dWidng7kJphVUtbP1a1Ma0Lyjxj5yLgQZIApaKCJUew,1479
 benchflow/continue_run/batch.py,sha256=8BNSInyp_Ol6Fcrg4ezBJdi3cBS12nC8mf963_x4fIg,4146
 benchflow/continue_run/orchestrator.py,sha256=wYoseBNE_rZ0EjNwWDyqfave54xNRe43sCd_O4VJjyk,27676
@@ -291,9 +291,9 @@ benchflow/trajectories/sessions.py,sha256=u_6x1WyT8YtKUVOb6b3myTCHhqn58a8NHCLAvQ
 benchflow/trajectories/tree.py,sha256=c0jyoP9OurDQq-zRtPc66edzWUvLGwQQz0jgtyqSw2U,5576
 benchflow/trajectories/trl_sft_tokenization.py,sha256=-Ef0bP9cuzfzOz5nIv2mEF3taQIEw68h6SM1YO-Mwn4,8643
 benchflow/trajectories/types.py,sha256=aSJclec3dOCUlfFNDL8GfzvkAcBo5r-dvlHE0MnGdMo,27554
-benchflow/trajectories/viewer.py,sha256=59wr8pcWWsCYzuI_sFzTSFJFHeXzH6Yklrh8UsF1ksg,26561
-benchflow-0.7.2.dev1986.dist-info/METADATA,sha256=D7gO4mBmAwaKA5_bFHpuy2Hu2_vfMEBQLdPfVoCyFEk,15261
-benchflow-0.7.2.dev1986.dist-info/WHEEL,sha256=zOwg4jB6zX2kU910N-cMawjivD6tO8NEWvE12je1bVk,87
-benchflow-0.7.2.dev1986.dist-info/entry_points.txt,sha256=qPMBVtH4qz3P4nMALEAdTbvpxvsVch9RhqshZKQn1dE,84
-benchflow-0.7.2.dev1986.dist-info/licenses/LICENSE,sha256=ohnbHkuBylrTtOvaZJk9Hz7dRwmaEEGAb5ZcBV4Xuhw,10779
-benchflow-0.7.2.dev1986.dist-info/RECORD,,
+benchflow/trajectories/viewer.py,sha256=4V7HMjJtUt8_HE1dStZwzF9W_3yefHNas7R3Y1iZmvY,34760
+benchflow-0.7.2.dev1994.dist-info/METADATA,sha256=BZC_LqHIgXKHtV4fMAqIDeaodQRbu2iDK2awvDuJ3fI,15565
+benchflow-0.7.2.dev1994.dist-info/WHEEL,sha256=zOwg4jB6zX2kU910N-cMawjivD6tO8NEWvE12je1bVk,87
+benchflow-0.7.2.dev1994.dist-info/entry_points.txt,sha256=qPMBVtH4qz3P4nMALEAdTbvpxvsVch9RhqshZKQn1dE,84
+benchflow-0.7.2.dev1994.dist-info/licenses/LICENSE,sha256=ohnbHkuBylrTtOvaZJk9Hz7dRwmaEEGAb5ZcBV4Xuhw,10779
+benchflow-0.7.2.dev1994.dist-info/RECORD,,