Skip to content

feat(runs): list every launch on a Launches page - #184

Merged
harenome merged 2 commits into
ThinkFlowLab:mainfrom
harenome:feat/launches-page
Nov 17, 2026
Merged

harenome merged 2 commits into
ThinkFlowLab:mainfrom
harenome:feat/launches-page

Conversation

@harenome

@harenome harenome commented Nov 17, 2026 •

Copy link
Copy Markdown
Collaborator

Closes #183

Related: #171, #172.

What

This PR adds a Launches page (/launches) under Manage, after Workspaces. It is a table with one row for each launch that Nullius made from the builder, in a workspace or not, newest first. The rows come from every run record in the registry (execution.EXECUTIONS), grouped by launches.group, which the Runs page also uses. The jobs of each record are read without their trials (parse.parse_job(..., trials=False)). A sharded launch is one row in every state: its records are grouped by the launch id that the plan gives each shard, so the row exists before any shard writes its job.

Column Shows
State an icon for the state of its runs that says the most: one still going, else the worst end (failed, refused, stopped, finished)
Launch its title as the Runs page shows it
Benchmark the benchmark of the recipe the launch was compiled from, by its catalogue name, linked to its page
Dataset the recipe's datasets, by their catalogue names, each linked to its tab on the benchmark's page (/benchmarks/<key>#<dataset>); a dataset of the standard by its label, linked to the page's Proofs (#standard); a launch with no recipe shows the dataset its jobs recorded
Started the earliest start of its runs
Runs its runs of agents, one for each shard
Checks its nop and oracle runs, counted the same way
Shards its run records: one per machine
Machines this machine or the hosts, names only
Size its run folders on this machine
Workspaces each workspace with a root in its run folders (workspace.holding), or "none"

A benchmark or dataset that the catalogue does not list (removed, or foreign) is shown as the recipe wrote it, without a link.

State. Each state has an icon of a different shape: a play triangle for running, an hourglass for preparing, a square in an open circle for stopping, two bars for paused, the warning triangle for failed, an octagon with a small x for refused, a square for stopped and the check for finished. The colours are those of the Runs page's state dots: green for running, preparing and finished, amber for paused and stopping, red for failed, refused and stopped. The running icon pulses, except with reduced motion. Each icon is an svg with role="img". Its aria-label and its <title>, the tooltip, are the state's word. The header cell shows nothing, to keep the column narrow. Its name, State, is visually hidden text for screen readers. The words and their order are the Runs page's (_TELLING in web/runs.py). A run record that says "done" shows as finished. A failed record whose jobs were all refused before they started (Execution.refused) shows as refused. For a sharded launch, the state is that of the shard that says the most. Six icons are new in the sidebar sprite (#i-play, #i-hourglass, #i-stopping, #i-pause, #i-refused, #i-stop). Failed and finished use #i-alert and #i-check.

Scroll. When the columns do not fit, at any width, the table scrolls sideways in its own box and the page does not. As on the Runs page, the text columns keep a minimum width (Launch, Benchmark, Dataset, Machines, Workspaces) and short values stay on one line (Started, the counts, Size), so the table scrolls instead of squeezing its cells to one word per line. Names still wrap, never cut. State and Launch stay at the left edge, with the page's background, so the other cells pass under them. The row hover paints them too. This is the rule that keeps the pick and Run columns of the Runs page's All runs table in place, now also for a sticky-col class (style.css).

Toolbar. Three buttons sit above the table, in one style, each an icon before its text: Columns (a table), Filter (a funnel) and Sort by (an up arrow beside a down arrow). The icons are new in the sidebar sprite (#i-table, #i-filter, #i-sort) and decorative (aria-hidden); the text names each button. The header cells have no filter carets and no sort marks.

  • Columns shows or hides every column except State, Launch and the actions. State is in the list, ticked and disabled. Checks, Machines and Size are hidden until a viewer shows them, so the table fits at 1280 px without its own scroll. The choice is kept per viewer in localStorage.
  • Filter is one menu with a group for each of State, Benchmark, Dataset, Machines and Workspaces ("none" gives the launches that no workspace holds). Each group lists the values, with All and None. The button shows how many columns a filter narrows. Clear filters shows every launch again.
  • Sort by lists Default (the page's order, newest first) and the columns State, Launch, Benchmark, Dataset, Started, Runs, Checks, Shards and Size, then Ascending and Descending. A column sorts in its usual direction first: numbers descending, text ascending. State sorts in the Runs page's state order, not by the word. Only the sorted column's header shows an arrow, the existing sorted-arrow style, and it carries aria-sort. With Default, no header has an arrow.
  • A click on a header's text, or Enter on it, still sorts by that column. It sets the same sort as Sort by, and the menu shows it.
  • Each button has aria-haspopup and aria-expanded. Opening a menu puts the focus in it. Escape, a click elsewhere, or Tab out of the menu and its button closes it, and Escape gives the focus back to its button. A click on a button's icon opens its menu as a click on its text does.

The page shows 10, 25 (default), 100 or all launches, with previous and next buttons. A filter or a sort goes back to the first page, and the count says how many launches the filters let through. The choice is kept in localStorage. All of this happens in the browser.

Counts. Runs and Checks come from one helper (_tally in web/runs.py), which the Runs page's launch blocks also use, so a launch shows the same numbers on both pages. It counts the jobs written, less the shards that a merge written beside them stands for, and the jobs that the records name and that are not written yet (a job refused before it started among them), each shard one, with nop and oracle as checks.

The sort, the filters and the Columns menu come from the Runs page's "All runs" table, moved to static/tables.js and _tables.html (filter_button). The Filter and Sort by menus are new there: the tools macro and NulliusTable.toolbar, which a table opts in to. So are the pages: the pager macro and NulliusTable.pager, which work with the shared sort and filters and are tied to their table by its id, with no id of their own. The Runs page does not opt in and keeps its header carets, sort marks and single page. The Delete button, its question and its dialog moved to runs/_launch_ask.html, which both pages use. That is the first commit. Three changes on the Runs page: a header without data-type does not sort (before, the pick column sorted, but its style said that it did not), the sorted header carries aria-sort, and its Columns button has the table icon in place of the "▦" character.

Actions on each row:

  • Add to workspace is an icon button on every row, before the trash (aria-label "Add launch <title> to a workspace", tooltip "Add to workspace"). Its icon, #i-ws-add in the sidebar sprite, is the Workspaces grid with a plus in place of its fourth square. It is faint at rest and the text colour on hover and keyboard focus. It opens a dialog that asks for a workspace. The workspaces that already hold the launch are not in the list. The last choice is New workspace…, which shows a Name field. POST /launch/<key>/add adds the launch's job roots to the chosen workspace (shared.adopt_roots), which stays closed unless it is already open. For New workspace, the route creates a workspace with that name, or with a random two-word name if the field is empty (workspace.create), adds the roots to it and opens it, as a new workspace from adopt_roots opens. It does not reuse a workspace that already holds these roots. Both need the form token and a local client. The notice names the workspace.
  • Delete is a trash icon button (aria-label "Delete launch <title>", tooltip "Delete launch"): faint at rest, the danger colour on hover and keyboard focus. It deletes the whole launch, every shard. It asks the question of the Runs page's Delete launch, whose confirm button stays the red one, and uses the route of the Runs page (POST /launch/<key>/delete). The form sends back=launches. The route then finds the launch among all launches, with or without an open workspace, and returns to Launches. What it would take (folders, size, judgements, the workspaces that hold it, the hosts' copies) is worked out only when the dialog opens, by GET /launch/<key>/ask, which renders the same question as the Runs page from the shared ask macro. The page itself only checks whether the launch can be deleted. When it cannot, the trash stays in its place, grey, with aria-disabled="true": it still takes the focus, a click or Enter on it does nothing, and its reason is its tooltip and its description for a screen reader (aria-describedby, visually hidden text). The reason is "not while it runs", "no run record" (a job of the launch that no run of Nullius names), or "its run record cannot be read". The POST makes its own checks as before.
  • The title, and each name in Workspaces, opens that workspace on the Runs page at the launch. POST /workspaces/open/<name> takes an optional launch for this and goes to GET /launch/<key>/open, which finds the block that the open workspace's Runs page gives the launch: there, a sharded launch is keyed by the records launched into that workspace, which need not be the first record of the launch.

The Launches icon is a small rocket in the sidebar sprite. No other icon in the sprite is like it. The icon is provisional.

For speed, the page reads the workspaces once (workspace.holding_all). It finds the run of each job from one map, not with for_job for each job. execution.deletion takes the holdings that the caller has already read. execution.size_of counts a file that vanished after the folder was listed (a run still writing replaces its own record) as 0 instead of failing the page.

Touches

backend (runs/parse, execution, workspace) · routes (web/runs, web/workspaces) · templates and static (runs/launches.html, runs/_launch_ask.html, _tables.html, runs/index.html, base.html, tables.js, style.css, runs/launches.css) · docs and site (concepts.md, cli/pages.md)

Before and after

Seeded data: 39 launches. Among them: an ended launch on one machine in a workspace, an ended launch in two shards on two hosts, a running launch over three machines whose shards 2 and 3 have not written their jobs, a preparing launch over two machines with no job written, a launch whose workspace was deleted, a running launch on one machine, a sharded and a whole launch with long titles that wrap, and a launch in each other state: preparing, paused, stopping, stopped, and refused (failed before any part was launched). Launches carry recipes of benchmarks in the catalogue, and one of a benchmark that is not.

Before After
Sidebar, Manage group before-sidebar-1280-light after-sidebar-1280-light
Launches page Light Dark
1280 px launches-1280-light launches-1280-dark
390 px launches-390-light launches-390-dark
1280 px, every column shown, the table scrolled sideways launches-scrolled-1280-light
390 px, the table scrolled sideways: State and Launch stay launches-scrolled-390-light launches-scrolled-390-dark
Action After
Columns menu, defaults launches-columns-menu
Filter menu, Machines on one host launches-filter-menu
Filtered on one host, the count on Filter launches-filter-machines
Sort by menu, Shards descending launches-sort-menu
Sorted by Shards: the arrow on that header only launches-sorted-by-shards
Delete of a sharded launch launches-delete-dialog
Add to workspace launches-add-dialog
Add to workspace, New workspace launches-add-dialog-new
After a Delete launches-after-delete
Runs page, By launch: the running launch over three machines runs-by-launch-pending-shards-1280-light
Runs page, All runs sorted by Agent with the shared script runs-all-sorted-by-agent-1280-light

Verified

At the tip, on main at 60d4fc90, in throwaway containers (python:3.12, with uv for the Harbor build tests):

  • python3 app/tests/parallel.py: 2050 tests in 119 modules, OK. New: SizeOf in test_run_delete.py, a file removed between the walk and its size.
  • uv run --no-project --with-requirements app/requirements.txt python app/tests/parallel.py app/ui_tests, with Node.js installed: 482 tests in 43 modules, OK. At the first commit alone: 463 tests, OK.
  • New: LaunchesPage in app/ui_tests/test_run_routes.py. It checks that every launch is a row, its Launch cell the title alone, including a launch that no workspace holds and a sharded launch on two hosts, and the Runs, Checks, Shards, Machines and Workspaces cells. The Runs cell of a sharded launch counts a run for each shard. A sharded launch is one row with no shard rows and no chevron, its state finished, and the question its Delete asks (/launch/<key>/ask) lists "(2 shards)". A row holds no Delete question; the ask route renders it with the form, back=launches, the runs and the size. Runs and Checks are the same on both pages for a launch whose two shards a merge written beside them stands for (1 run) and for a launch whose two jobs were refused before they started (1 run, 1 check, refused). A workspace that holds only the second shard of a sharded launch opens at the block its Runs page gives the launch, and the open workspace's title link goes the same way. Three shards with one job written are one row of 3 shards. Three shards preparing, running and stopping on three machines, none with a job written, are one row: running, 3 shards, the three machines, and a disabled Delete. It checks that State is the first column, an icon with the state's word as its name and tooltip (#i-refused for a failed record whose jobs were refused, #i-check for finished, #i-alert for failed, #i-pause for paused), that it sorts in the Runs page's order, and that the State header holds only visually hidden text. It checks the toolbar: Columns, Filter and Sort by, each with aria-haspopup; the Filter groups State, Benchmark, Dataset, Machines and Workspaces, the State group's words in the Runs page's order; the Sort by columns; no filter caret in a header and no sorted header at load; the pager tied to the table, 25 of 10, 25, 100 or All, with no id of its own. It checks the sidebar item, and that the title opens the workspace at the launch. For Add to workspace, it checks the form token, an unknown workspace (404), the notice and the root that the workspace gets. For New workspace, it checks the form token (nothing is created without it), a named workspace that opens with the launch's root, an empty name that gives a new random name though two workspaces already hold the launch, and the Add button and a select with only New workspace when every workspace holds the launch or when there is no workspace. It checks Delete with no workspace open, and that a running launch has a greyed Delete with aria-disabled, its reason in its tooltip and in visually hidden text that aria-describedby names, a 409 from the ask route, and its folder kept after a POST. A launch whose run record is not JSON says "its run record cannot be read", and a launch with no run record says "no run record". It checks the one-line empty page. It checks the Benchmark and Dataset cells: catalogue names linked to /benchmarks/gsm8k and /benchmarks/gsm8k#test, a standard dataset by its label linked to /benchmarks/longmemeval#standard, a benchmark that the catalogue does not list as written without a link, and a launch with no recipe. It checks the headers and the Columns menu's columns, State ticked and disabled, and the aria-label of the Add and trash buttons. Also new: test_parse (a job read without its trials) and test_workspace Holding (holding_all). The RunDelete and LaunchShards tests of the Runs page pass without change. test_the_flat_table_is_the_second_view now also checks that All runs is not inside the By launch view, after the move of the Delete dialog.

Browser check, Chromium (Playwright 1.49), on the seeded data. 143 checks passed, with no page JavaScript errors:

  • Launches comes right after Workspaces in the sidebar.
  • Light and dark, 1280 px and 390 px: the columns shown by default. The State header shows nothing (its text is 1 px, visually hidden) and is named State. There is no shard row and no chevron. The running launch over three machines is one row, running, 3 shards. The preparing launch over two machines with no job written is one row, preparing, 2 shards. No header has a sort arrow at load. Columns, Filter and Sort by sit on one line, inside the page. Each has its own icon (#i-table, #i-filter, #i-sort), 18 px, aria-hidden, before its text and inside the button, and its text is its accessible name. The page does not scroll sideways.
  • With every column shown, at 1280 px, 1024 px and 390 px, the table is wider than its box (by 232 px at 1280 px) and scrolls in it, Launch keeps its minimum width, and Started and Size stay on one line. At 1024 px, scrolled to the end, State and Launch stay in place. With the default columns at 1280 px, the table fits its box. At 390 px the table scrolls in its own box. Scrolled sideways, in light and dark, at 390 px and at 1280 px with every column shown: State and Launch keep their place, Launch starts where State ends, and Benchmark moves by the scroll. On every row in view, the point near Launch's right edge is the Launch cell, not a cell under it. The sticky cells and the Launch header have the page's background. On hover, the sticky cells take the row's hover colour.
  • Columns, Filter and Sort by: each closed at load with aria-haspopup; Enter opens it with aria-expanded true and the focus inside; Escape closes it and gives the focus back to its button; a click on its icon opens it and a click elsewhere closes it. Opening Filter closes Columns. Tab out of Filter closes it; Shift+Tab from Sort by's menu to its button keeps it open, and one more closes it. In Columns, State is listed, ticked and disabled. In Filter, the State group is named State and lists running, preparing, stopping, paused, failed, refused, stopped and finished, in that order. In Sort by, State is listed, Default is chosen and the directions are disabled.
  • Filter: State on failed shows only failed launches and the button counts 1. With Machines on gpu-box too, it counts 2, nothing matches, and the empty note shows. Clear filters shows every launch and hides the count. Machines on gpu-box alone shows the one launch. Workspaces on "none" shows only launches that no workspace holds. Benchmark on GSM8K shows that one launch.
  • Sort by Shards sorts descending first, with the arrow on Shards only and aria-sort "descending" there and on no other header. Ascending sorts ascending, with an up arrow. A click on the Runs header sets Sort by to Runs, descending, with the arrow on Runs only. Sort by State gives the state order. Default puts the page's order back, removes every arrow and disables the directions. A click on the blank State header sorts by State. Enter on the Launch header sorts by it, with aria-sort "ascending".
  • With every column shown, the page does not scroll sideways at 1280 px or 390 px.
  • 25 of 39 launches by default, the next page has the other 14, and Show 10 gives 10; its second page is 11–20, a sort goes back to 1–10, and a filter on finished goes back to the first page and counts only the finished launches. The pager has no id of its own; its select is labelled Show and its buttons are named Previous page and Next page. Runs and checks are counted from the written jobs, as on the Runs page: the launch over three machines has 3 runs (2 not written yet) and 0 checks. Machines lists only the names. Every Launch cell is the title alone, with no second line.
  • The running launch has a grey Delete at 0.4 opacity that takes the focus, with the tooltip "Delete launch: not while it runs" and the same reason as its description; Enter or a click on it opens nothing. No Delete question is asked with the page. Delete of the sharded launch asks /launch/<key>/ask once, then opens the dialog with the focus on Keep it, and lists "(2 shards)". Keep it closes the dialog and deletes nothing. Add opens a dialog with a select of workspaces and New workspace last. Add with New workspace and a name gives the notice, and the row lists the workspace. Delete removes a row, with the notice.
  • On the Runs page: a participant's shards fold and unfold. All runs' Columns button has the table icon, aria-hidden, and its text is its name; a click on the icon opens its menu. All runs keeps its header carets and the ↕ mark on sortable headers, sorts by Agent, opens and closes its header filter menu, and its Columns menu shows, hides and remembers columns.

Render time of GET /launches (Flask test client, in a python:3.12 container limited to 2 CPUs, the median of three requests after the first): about 0.10 s at 50 launches and 1.05 s at 500; the first request after start takes 0.18 s and 1.25 s. With 60 runs of about 300 files each, held by 10 workspaces, it is 0.24 s (0.33 s for the first request). Working out the Delete question for every row on that seed took 0.79 s a request, and 10.6 s for the first: it reads the jobs of every workspace that holds a launch. The dialog now asks for it when it opens. The page stays rendered on the server, and sorting, filtering and pages stay in the browser.

Risk

The Launches page and the Runs page's launch blocks count runs and checks with one helper, so a change to it changes both. execution.deletion can now take holdings that the caller has read, and workspace.holding is now holding_all for one folder. Their default behaviour does not change, and the RunDelete tests cover it. If the Launches page lists a launch in the wrong group, Delete removes the wrong runs. Delete still finds the runs again on the server from the key, and only runs that Nullius made can be deleted. The Runs page uses the shared script for sort, filters and the Columns menu, and the sticky-column rule now also matches sticky-col. If that script or rule breaks, those features break on both pages. To undo, revert the two commits. Nothing is migrated.

… the launch Delete dialog

The sortable headers and the per-column value filters move to
static/tables.js and the filter_button macro to _tables.html. The All runs
Columns menu moves to NulliusTable.columns. The launch Delete button, its
question and the one dialog move to runs/_launch_ask.html. The Runs page
uses them as before. A header without data-type no longer sorts: the pick
column did, though its style said it did not. The sorted header carries
aria-sort. A click on anything inside the Columns button opens its menu.

A table can opt in to a Filter and a Sort by menu beside its Columns
button (the tools macro, NulliusTable.toolbar) in place of header carets:
the sort arrow shows only on the sorted column, and a header click still
sorts. Each menu closes on Escape, a click elsewhere, or Tab out of it. It
can opt in to pages too (the pager macro, NulliusTable.pager): 10, 25,
100 or all rows a page, back to the first page after a filter or a sort.
A .sticky-col cell keeps its place while .table-wrap scrolls sideways, by
the rule the pick and Run columns use.

The Columns, Filter and Sort by buttons carry an icon before their text,
new in the sprite: a table, a funnel, and an up arrow beside a down
arrow. The Runs page's Columns button takes the table in place of the
"▦" character.
Deleting a workspace leaves the launches Nullius made on disk, and no page
listed them. The Launches page, under Manage after Workspaces, has one row
per launch: every run record, grouped as the Runs page groups them (a
sharded launch by the launch id its plan gives each shard, one row before
any shard writes its job), each record's jobs read without their trials.

A row gives the launch's state, its title, its benchmark and datasets by
their catalogue names, linked to the benchmark's page, start, runs,
checks, shards, machines, size on this machine and the workspaces that
hold it. The state is an icon of its own shape per state, the word its
name and tooltip, in the Runs page's words and order, the most telling of
its shards. A record that says done is finished, and a failed one whose
jobs were all refused before they started is refused. Runs and checks
are counted by one helper with the Runs page's launch blocks: jobs
written, less those a merge stands for, and those still to come, each
shard one, nop and oracle as checks.

When the columns do not fit, the table scrolls sideways in its box, its
text columns kept to a minimum width and short values to one line, and
State and Launch stay in place. Columns, Filter and Sort by sit above the
table: Columns hides Checks, Machines and Size until a viewer shows them,
so the table fits 1280 px; Filter narrows by state, benchmark, dataset,
machine and workspace; Sort by sorts by any column but Machines and
Workspaces. The table pages 10, 25, 100 or all launches at a time. The
title opens a workspace that holds the launch at the block its Runs page
gives it.

Each row has two icon buttons. Add to workspace asks in a dialog for a
workspace, or New workspace with an optional name, and adds the launch's
job roots to it; a new workspace opens. The trash asks as the Runs page's
Delete launch, for every shard, and returns here. What it would take is
worked out only when it opens (GET /launch/<key>/ask), not with the page.
A launch it cannot delete keeps a greyed trash that still takes the focus
and says why: not while it runs, no run record, or its run record cannot
be read. The size of a run folder no longer fails on a file the run
replaces while it is read.
@harenome
harenome merged commit fe260f5 into ThinkFlowLab:main Nov 17, 2026
2 checks passed
@harenome
harenome deleted the feat/launches-page branch November 17, 2026 13:30
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

Launches: list every launch Nullius made, in a workspace or not

1 participant