/flow shows actual throughput: per stage, items out per hour over the last 24 h with a sparkline, and a headline of landed in the last hour and median ask→live #1786

/flow now shows how much moves through the pipeline, not just what is waiting in it. The counts come from data the forge already has (runs, pulls, releases, issues), with no new table.

What a stage counts as "out" (src/flow/throughput.ts):

Counts are grouped by hour, counting back from now; hour 0 is the last hour.

The median ask→live covers merges that went live in the last 24 hours. It runs from the oldest issue the pull closes (the pull's own creation if it closes no issue) to when the first production deploy after the merge started serving.

On the page (src/web/pages/flow.ts):

• Each stage row gets a column with the last-hour count, a 24-bar SVG sparkline (oldest hour on the left) and the 24-hour count. The checking row also shows how many passed and failed.
• Under the bottleneck sentence there is a new line: "landed N in the last hour · M today · median ask→live X h (24 h)".
• Every count and every bar links to /flow?out=<what>[&hour=<0–23>]. That lists the counted items under the table, each linking to its issue, pull, run or commit. An unknown out or an invalid hour lists nothing.
• A sealed issue is counted, but its title is not shown.
• The sparkline's CSS goes through the page's existing nonce'd styleCss, so the content security policy is unchanged.

Things to look at:

• "today" means the last 24 hours, so it always equals the sum of the sparkline bars, rather than counting from midnight in some timezone.
• The check-queue row shows "checks started". The issue's list named seven flows for eight stages, and that row would otherwise be empty.
• The terminal's Flow command and the MCP flow tool are unchanged.

Tests: test/flow-throughput.test.ts seeds six landings over three hours (2/3/1 per hour), each closing an issue asked 1–6 hours before the deploy that served them. It checks those per-hour counts and a median of 3.5 h, plus:

• checks that passed, failed or were cancelled;
• a builder retried on one issue, which counts as one issue started;
• a release whose preflight failed, which is not counted;
• an old merge served by its own earlier deploy, which is left out of the median;
• the headline text, the links, the item lists, and query values that are rejected.

The new tests and the existing flow-pipeline and flow-stage tests pass, and the type check is clean. docs/FLOW.md and the README describe the new section.

To try it: sign in as a forge admin on the preview, open /flow, and click any count or bar.

🤖 Generated with Claude Code

Closes #1779

/flow now shows how much moves through the pipeline, not just what is waiting in it. The counts come from data the forge already has (runs, pulls, releases, issues), with no new table.

What a stage counts as "out" (src/flow/throughput.ts):

Counts are grouped by hour, counting back from now; hour 0 is the last hour.

The median ask→live covers merges that went live in the last 24 hours. It runs from the oldest issue the pull closes (the pull's own creation if it closes no issue) to when the first production deploy after the merge started serving.

On the page (src/web/pages/flow.ts):

• Each stage row gets a column with the last-hour count, a 24-bar SVG sparkline (oldest hour on the left) and the 24-hour count. The checking row also shows how many passed and failed.
• Under the bottleneck sentence there is a new line: "landed N in the last hour · M today · median ask→live X h (24 h)".
• Every count and every bar links to /flow?out=<what>[&hour=<0–23>]. That lists the counted items under the table, each linking to its issue, pull, run or commit. An unknown out or an invalid hour lists nothing.
• A sealed issue is counted, but its title is not shown.
• The sparkline's CSS goes through the page's existing nonce'd styleCss, so the content security policy is unchanged.

Things to look at:

• "today" means the last 24 hours, so it always equals the sum of the sparkline bars, rather than counting from midnight in some timezone.
• The check-queue row shows "checks started". The issue's list named seven flows for eight stages, and that row would otherwise be empty.
• The terminal's Flow command and the MCP flow tool are unchanged.

Tests: test/flow-throughput.test.ts seeds six landings over three hours (2/3/1 per hour), each closing an issue asked 1–6 hours before the deploy that served them. It checks those per-hour counts and a median of 3.5 h, plus:

• checks that passed, failed or were cancelled;
• a builder retried on one issue, which counts as one issue started;
• a release whose preflight failed, which is not counted;
• an old merge served by its own earlier deploy, which is left out of the median;
• the headline text, the links, the item lists, and query values that are rejected.

The test is now listed in GATE in scripts/gate.ts, which is what the last check failed on. Admission: bun scripts/flake-hunt.ts admit test/flow-throughput.test.ts gave 10/10 green, random order, under load (seeds 1355380017, 2952699592, 2602253883, 1306929882, 2571361766, 4118929990, 2622846114, 3814059754, 4073273512, 1707047753). test/check-list.test.ts passes. Before that change, the new tests and the existing flow-pipeline and flow-stage tests passed and the type check was clean.

docs/FLOW.md and the README describe the new section.

To try it: sign in as a forge admin on the preview, open /flow, and click any count or bar.

🤖 Generated with Claude Code

/flow now shows throughput per stage: how many items left each stage in the last hour and in the last 24 hours, with a 24-bar sparkline, under a headline of landings and the median ask→live. This update addresses @juniper's review. The same numbers can now be read on the page, through the MCP flow tool, and from the terminal's Flow command, and the sparkline is no longer a screen-reader or keyboard trap.

Both doors

• src/flow/throughput.ts gains stageOut (per stage: out, label, lastHour, day, hours) and throughputSays, which gives the headline as one sentence: "landed 2 in the last hour · 6 in 24 h · median ask→live 3.5 h (24 h)". The page headline reads the same, and a test checks they match.
• The MCP flow tool now returns throughput: { says, medianAskToLiveMs, stages } next to says and stages. It is still forbidden to anyone but a forge admin.
• The terminal's Flow command (still offered only to a forge admin) now opens a small dialog instead of a toast. A toast is one footer line for 4 seconds, too small for eight stages. The dialog is src/tui/screens/flow.ts: the bottleneck sentence, the headline, then one line per stage with its last-hour and 24 h counts. Enter, Esc or q closes it.
• I changed one existing test, in test/flow-pipeline.test.ts. It checked that the Flow command toasts the bottleneck sentence; that behaviour was changed on purpose, so the test now checks that the command opens the Flow dialog and that the dialog starts with the same sentence.
• README says it once; docs/FLOW.md describes the MCP fields and the dialog.

Phone and keyboard

• Each sparkline is now <svg aria-hidden="true" focusable="false"> with plain bars: no role="img", no links, no titles. Each row has two tab stops, the last-hour count and the 24 h count, which are the way in. A test checks every sparkline has no links or role and every stage has both count links.
• The headline says "in 24 h" instead of "today".
• ?hour=1…23 still lists that hour's items when typed into the URL; nothing on the page links to it any more.

Not done: capping a long ?out= list with a "show more" link, which was listed as next.

To try: as a forge admin, open /flow and Tab through a row (two stops). In the terminal, open the palette and run Flow. Over MCP, call flow and read throughput.says. The flow tests and test/mcp.test.ts pass, including three randomized runs, and the type check passes.

🤖 Generated with Claude Code

Parked by the crew: the check failed ("apps: production, previews, data, gates, crew > a pull gets a preview with a copy of production's data") and the decider saw no point in another try