菜单
Prove what the Rerun viewer rendered — pixel evidence over logs. Use when a .rrd, blueprint, or Rerun rendering must be visually verified, when a timeline sweep or video of a recording is wanted, when an .rrd must be embedded in an HTML page, or when a gradio/WebViewer surface needs browser validation.
用 Codex 或 Claude 帮你安装 复制这段 Prompt，粘贴到 Codex、Claude 或其他助手里，让它检查 Skill 页面并帮你完成安装。
基于 SOC 职业分类
Model routing for delegated work — use when spawning subagents or workflows, choosing a model for a task, or handing work to the Codex runtime. (Claude Code side; other agents rarely delegate.)
Fleet operations — use when a task involves other machines (discovering, reaching, or reasoning about them), installing/updating/removing any user-level tool, converging a machine, or working with paseo or the agent-fleet repo.
Python conventions for every repo — typing, dataclasses, return types, docstrings. Use whenever touching Python code.
Create self-contained single-file HTML artifacts when an explanation, plan, report, review, comparison, diagram, deck, prototype, or lightweight editor would be more readable, useful, or reusable in a browser than in markdown. Use for making dense information scannable and beautiful with semantic HTML, inline CSS, inline JavaScript, responsive layout, typography, charts, tables, annotations, tabs, collapsibles, timelines, diagrams, and copy/export affordances. Especially use when the user wants an artifact they are likely to open, read, share, or iterate on directly as a .html file.
Create a new single-purpose CV node in the monorepo with API layer (class-based node with verbose Rerun logging, result logging helper, config + result dataclasses), Gradio UI with embedded Rerun viewer (streaming binary stream), CLI entry point (tyro + RerunTyroConfig), and bundled example data. Use when adding a new model, predictor, or pipeline step as a standalone reusable app.
Performs conda-forge operations. Fixes failing builds by analyzing CI logs, creates new packages via staged-recipes, adds cross-compilation and ARM support to feedstocks, and migrates recipes from v0 to v1 format. Use when working with conda-forge feedstocks, staged-recipes, build failures, recipe migrations, or when the user mentions conda-forge.
| name | rerun-viewer-validation |
| description | Prove what the Rerun viewer rendered — pixel evidence over logs. Use when a .rrd, blueprint, or Rerun rendering must be visually verified, when a timeline sweep or video of a recording is wanted, when an .rrd must be embedded in an HTML page, or when a gradio/WebViewer surface needs browser validation. |
| compatibility | Requires the rerun binary (with the viewer-mcp subcommand) and ffmpeg on PATH, plus a GPU or software rasterizer for headless rendering. The web branch additionally needs a browser automation tool (playwright or chrome-devtools) and network access. |
Prove what rendered. Logs, metadata, and rrd stats say what was sent; only pixels say what the viewer shows.
scripts/rrd_to_video.py (this skill's helper). Never loop MCP screenshots for video.references/web.md — the iframe embed recipe (CORS, Chrome Local Network Access, tailscale) and the playwright validation recipes live there. The MCP structurally cannot reach a WASM viewer in a browser — no gRPC server to dial.Video vs embed (branches 3 vs 4, when both fit): the embedded rrd is the richer artifact — fully inspectable, orbitable, scrubbable — so prefer it when the recording is browser-sized. The WASM viewer holds the whole recording in memory, so check size first (ls -lh, rerun rrd stats) and gate embeds at a few hundred MB (hard ceiling ~1.5 GiB — see references/web.md). Choose video when the recording is huge, when the audience only needs to watch (Slack, PR description), or when the data needs a visualizer the web viewer doesn't have (custom visualizers). Best of both, often: a trimmed/downsampled preview rrd for the embed plus a full-fidelity video.
Version rule for every branch: the viewer that validates must be ≥ the SDK that wrote the data — Rerun has no forward compat, so an .rrd written by a newer SDK will not load in an older viewer (including the WASM viewer inside a gradio-rerun app pinned to an older release). Don't hardcode version numbers in prose or scripts; the environment's package pins (this skill's run_constraints) guarantee a capable viewer.
Default headless for both ViewerClient.spawn(headless=True) and rerun --headless:
screenshot times out if a headed window can't render, notably on macOS).DISPLAY. A headed spawn without DISPLAY wedges silently (channel fills, rr.log blocks forever).Go headed only when a human co-views: the user wants to watch you scrub, or wants the viewer left open afterwards. All tools work identically against either — headed vs headless is user preference, not capability.
Lifecycle gotchas (both modes):
connect → work.ViewerClient.spawn resolves rerun from PATH — stale global installs win. Always pass executable_path= pointing at the project env's rerun binary.detach_process defaults: headless → attached (dies with your script / close()); headed → detached (survives; only explicit close() kills it). Clean up detached viewers when done.The server is rerun viewer-mcp (stdio); it dials a running viewer's gRPC ViewerControlService. In order of preference:
mcp__rerun__* tools already in your surface → use them.initialize → notifications/initialized → tools/call); reuse McpStdioClient from scripts/rrd_to_video.py.claude mcp add rerun -- <env>/bin/rerun viewer-mcp (or codex mcp add …). (The viewer-mcp subcommand exists since 0.34.)claude -p --mcp-config … from a non-Claude harness) crosses a provider boundary — confirm with the user first.17 tools: connect, disconnect, viewer_state, set_time, open_url (rerun-specific) + query_tree, get_node, screenshot, click, drag, hover, scroll, press_key, type_text, resize, wait_for, batch (egui UI, accessibility-tree based). Work observe → act → verify.
connect takes endpoint: "http://127.0.0.1:<port>" — plain http, not the SDK's rerun+http://…/proxy URL.open_url loads recordings: absolute file path (no file:// prefix), rerun:// dataset URI, or https URL.viewer_state first, always: recordings + per-timeline {timeline, type, min, max} + current time. Choose the timeline from this data, never by assumption.set_time: time is a sequence index for sequence timelines, nanoseconds for duration/timestamp timelines. play: true to run from there; default stays paused.screenshot always returns the PNG inline into context; save_path writes to disk in addition. Budget ≤ ~10 MCP screenshots per validation — seeing evidence frames is the point; sweeping is the helper's job.id from query_tree, role/label_contains) over raw pos; everything is in logical points (screenshot pixels at pixels_per_point: 1.0 align 1:1 with click coordinates).batch chains act+observe (e.g. set_time → screenshot) in one round trip.import rerun as rr
from rerun.experimental import ViewerClient
with ViewerClient.spawn(
headless=True, port=9877, hide_welcome_screen=True,
executable_path="<env>/bin/rerun", # NEVER rely on PATH
) as viewer:
rr.init("rrd_check", default_enabled=True, strict=True)
rr.connect_grpc(url=viewer.url)
rr.log_file_from_path("recording.rrd") # preserves saved blueprint/layout
rr.get_global_data_recording().flush(timeout_sec=30.0)
import time; time.sleep(3.0) # let ingestion + first render settle
viewer.save_screenshot("native-full.png")
Prefer save .rrd → reload → screenshot: it validates serialization, blueprint, and viewer loading in one pass. ViewerClient has no time-cursor setter — the playhead lives in the MCP (set_time) only.
Per-view capture works only for views the viewer is currently rendering. The safe pattern is authoring the blueprint in-process: view = rrb.Spatial3DView(…); rr.send_blueprint(view); viewer.save_screenshot(p, view_id=view.id) — returns in milliseconds. The trap: a view_id the viewer can't resolve to a rendered view (an unknown uuid, or a saved-blueprint view right after replaying an .rrd) gets no reply and the blocking call hangs forever, with no diagnostic on 0.34.0. So always run view_id calls in a killable child process with a timeout, and for replayed recordings prefer cropping the full screenshot (view rectangles are deterministic for a fixed viewport). To enumerate a recording's saved views (their /view/<uuid> ids are the same namespace as view.id, but resolve only while rendered):
import rerun.experimental as rrx
r = rrx.RrdReader("recording.rrd")
for chunk in r.stream(store=r.blueprints()[0]).to_chunks():
if str(chunk.entity_path).startswith("/view/"):
print(chunk.entity_path, chunk.to_record_batch())
python scripts/rrd_to_video.py --rrd recording.rrd --out sweep.mp4 \
--rerun-bin <env>/bin/rerun [--timeline frame] [--frames 150] [--fps 15] [--collapse-panels]
Spawns a headless viewer, drives rerun viewer-mcp over stdio (set_time → screenshot save_path per frame — zero agent context), ffmpeg-encodes. 120 frames at 1080p ≈ 10 s: the per-frame cost is the settle wait plus a ~32 ms screenshot RPC, so --settle-ms is the speed/fidelity dial. Auto-picks the first non-log_time timeline; handles sequence and temporal timelines (--frames samples evenly across the range); stdlib-only — needs just ffmpeg on PATH and the project env's rerun binary. The default --settle-ms 30 is enough for decoded video frames; raise to 100–400 when overlay-heavy views (detections, segmentation) must fully stabilize per frame — a mostly-duplicate sweep fails loudly with that advice (--allow-static overrides for genuinely static scenes). Verify 2–3 sampled frames visually (Read start/middle/end PNGs with --keep-frames) before trusting the mp4.
Collapse the blueprint/selection/time panels whenever the frame should be all content — videos, embeds, clean screenshots:
click with label_contains = "Blueprint panel toggle", "Time panel toggle", "Selection panel toggle". A fresh viewer starts with panels expanded, so one click each collapses; confirm via query_tree (the _streams_tree / _selection_panel panes disappear). The video helper does this for you: --collapse-panels.rrb.Blueprint(<views>, collapse_panels=True), or per-panel rrb.BlueprintPanel(state="collapsed") / rrb.SelectionPanel(…) / rrb.TimePanel(…) with "collapsed" | "hidden" | "expanded". An .rrd re-saved this way opens chrome-free everywhere, including the WASM viewer iframe./tmp/rerun-viewer-validation/<timestamp>/: screenshots, notes.md recording Rerun version, command, .rrd path/size, chosen timeline + range, wait times, renderer string, pass/fail.rerun rrd verify|stats|print <file> (use the project env's binary).tailscale serve --https <port> --bg <report-dir>, pick an unused port, curl -k -I the URL to confirm reachability. Path mode is fine for plain HTML + screenshots; a report that embeds an .rrd needs the CORS proxy setup in references/web.md.ViewerClient: https://ref.rerun.io/docs/python/main/experimental/