skycap-viewer

A browser viewer for skycap trajectory records: every model call of an RL rollout, as the trainer saw it, token by token.

Point it at a directory of records and you can:

• Browse a run as a list of rollouts, or grouped the way GRPO trains them (all samples of one prompt at one step), filtered by step, task, status or any annotation.
• Spot the odd ones out. A health band flags rollouts that are worth a look: truncated, failed calls, histories that forked, groups with no learning signal.
• Read a rollout token by token, with what was trained, what was context and what the model sampled marked on the text, and logprobs on hover.
• Follow forks (summarization, retries, edited histories) in a tree, and compare two branches from the point where they diverge.

It is read-only and local: it reads record files from disk, makes no network calls, and needs no tokenizer, no Python and no credentials.

Quick start

You need Node.js 22.15 or newer. There are no dependencies to install and no build step.

git clone https://github.com/NovaSky-AI/skycap-viewer && cd skycap-viewer
node bin/skycap-viewer.js /path/to/records

Then open http://127.0.0.1:8765. (npm link makes the same command available as skycap-viewer.)

The path can be a single record directory, or a folder that contains several; each record directory shows up as its own run.

Where records come from

skycap writes one record per trajectory into its --record-dir. Point the viewer at that directory, even while the run is going; new records appear as they are written.

What you see

Runs are listed on the left. Each run opens as a table:

• Groups (the default for training runs): one row per prompt per step, with a dot per rollout coloured by reward on a red-to-green scale. Click a group to expand its rollouts in place.
• Rollouts: one row per trajectory, with a strip showing where its tokens were trained.

Columns sort on click, and the view, filters and sort live in the URL, so a link shows exactly what you see.

A rollout opens in a side drawer:

• Path: the conversation as the model saw it, cut into blocks: what the model sampled and trained on, the chat template's scaffolding, the inputs it was given, and replayed text it didn't produce. Hover a token for its id and logprob.
• Tree: every message as a node. Forks are marked, and opening a branch point shows its branches side by side.
• Forks: pick two paths and compare them from the last message they share.
• Calls: each model call behind the rollout, with timing, token counts and how it ended.

Health flags sit in a band above the table; click one to filter to it. Among them:

Flag	What it means
truncated	a model call stopped on its token limit
failed-calls	calls that never produced a reply
forked	the history diverged, so the rollout trains as more than one row
unbridged	a prompt had to be re-rendered instead of extending the previous call token for token
no-train	nothing in a path is trainable
tokens-missing	the record's tokens aren't available (e.g. a mirror that left them out); only message text is shown
no-signal	every rollout in a group got the same reward, so the group teaches nothing
masked	a rollout in the group timed out or errored, so the whole group was masked

The viewer shows records, not training curves: rewards appear as they were recorded, and there are no averages, advantages or pass rates.

Command line

skycap-viewer <dir>... [--port 8765] [--host 127.0.0.1]   # serve the viewer
skycap-viewer summary <dir>...                            # a quick text summary of each run
skycap-viewer check <dir>...                              # validate records against skycap's format

Record format

The viewer reads skycap records, format_version 1, as specified in skycap's docs/format.md, and refuses other versions.

Development

node --test

The tests run the real views against real and generated records in test/fixtures*.

License

Apache License 2.0, the same license as SkyRL. See LICENSE.