A gate whose install downloads from the npm registry says so in its log with the bytes, and /capacity names the repository that fetched most in the last hour #8394

When a check, gate or preview runs its install inside a cell and that install receives more than 1 MiB from registry.npmjs.org, its log now ends with install fetched N MiB from the npm registry; the shared cache missed. /capacity and MCP capacity also name the repository the registry sent the most to in the last hour.

Where the repository comes from. meterEgress(label, repoId?) in src/apps/egress.ts now takes the repository id from the code that opened the cell (repo.id for checks and gates, record.repoId for apps), never from anything the install printed. While a meter is running, recordEgress counts that label's registry bytes per repository, minute by minute over the last hour, with the same 60-minute ring the per-destination counts use. At most 256 repositories are kept; past that, the one that fetched least this hour is dropped. Each meter also has npm(), which counts registry bytes only, so a large download from GitHub does not set off the line. Registry traffic with no meter, or with a meter that has no repository, goes into the hourly total but is attributed to no repository.

The log line. npmMissLine in src/checks/install.ts decides when to write it (strictly more than 1 MiB). It is written in two places:

• Checks and gates (src/checks/service.ts): straight after the existing install fetched … from the registry line.
• Previews and production (src/apps/process.ts install): contained installs are now metered under the cell's label, and the line goes to the deploy log. Installs outside a cell are not metered here either.

What capacity shows. egressHour(db) in src/actions/capacity.ts turns the top repository's id into owner/name, giving egress.npmTopRepoLastHour: { repo, bytes } | null. It is null when no metered install fetched anything, or when that repository has been deleted. The egress-npm line on /settings/capacity adds "…, most to owner/name (N MiB)". The capacity tool gained a field rather than a new tool, and its doc string grew by a few words. docs/MCP.md and docs/METRICS.md describe both additions.

Tests. bun test passed on all 8 files run, with the mcp-size limits unchanged. bun scripts/gate.ts, bun x tsc --noEmit and bun scripts/design-lint.ts were also clean. scripts/fast.ts selected 1470 files, so I picked the 8 that cover the change.

• test/egress-npm-repo.test.ts (new):

- 3 MiB writes the line; 512 KiB and exactly 1 MiB do not. - Bytes from other hosts don't count, and bytes after stop() aren't the install's. - Two repositories' installs are each attributed to their own repository, and the hour ages out. - Registry traffic with no meter or no repository is attributed to nobody.

• test/capacity.test.ts: MCP and the page name alice/npm-heavy with 5.0 MiB, and give null once the tally is reset.

Not done: no test runs a full contained install from start to finish. Cells cannot run in this sandbox, so the tests drive the meter and npmMissLine directly, and the two places that write the line are not tested. I also didn't add the top repository to the terminal's Egress screen or /metrics.

To try it: as a site admin, open /settings/capacity on the preview after a check has installed with a cold cache, or call MCP capacity and read egress.npmTopRepoLastHour.

🤖 Generated with Claude Code

Closes #8382

When a check, gate or preview runs its install inside a cell and that install receives more than 1 MiB from registry.npmjs.org, its log now ends with install fetched N MiB from the npm registry; the shared cache missed. /capacity and MCP capacity also name the repository the registry sent the most to in the last hour.

Where the repository comes from. meterEgress(label, repoId?) in src/apps/egress.ts now takes the repository id from the code that opened the cell (repo.id for checks and gates, record.repoId for apps), never from anything the install printed. While a meter is running, recordEgress counts that label's registry bytes per repository, minute by minute over the last hour, with the same 60-minute ring the per-destination counts use. At most 256 repositories are kept; past that, the one that fetched least this hour is dropped. Each meter also has npm(), which counts registry bytes only, so a large download from GitHub does not set off the line. Registry traffic with no meter, or with a meter that has no repository, goes into the hourly total but is attributed to no repository.

The log line. npmMissLine in src/checks/install.ts decides when to write it (strictly more than 1 MiB). It is written in two places:

• Checks and gates (src/checks/service.ts): straight after the existing install fetched … from the registry line.
• Previews and production (src/apps/process.ts install): contained installs are now metered under the cell's label, and the line goes to the deploy log. Installs outside a cell are not metered here either.

What capacity shows. egressHour(db) in src/actions/capacity.ts turns the top repository's id into owner/name, giving egress.npmTopRepoLastHour: { repo, bytes } | null. It is null when no metered install fetched anything, or when that repository has been deleted. The egress-npm line on /settings/capacity adds "…, most to owner/name (N MiB)". The capacity tool gained a field rather than a new tool, and its doc string grew by a few words. docs/MCP.md and docs/METRICS.md describe both additions.

Tests. bun test passed on all 8 files run, with the mcp-size limits unchanged. bun scripts/gate.ts, bun x tsc --noEmit and bun scripts/design-lint.ts were also clean. scripts/fast.ts selected 1470 files, so I picked the 8 that cover the change.

• test/egress-npm-repo.test.ts (new):

- 3 MiB writes the line; 512 KiB and exactly 1 MiB do not. - Bytes from other hosts don't count, and bytes after stop() aren't the install's. - Two repositories' installs are each attributed to their own repository, and the hour ages out. - Registry traffic with no meter or no repository is attributed to nobody.

• test/capacity.test.ts: MCP and the page name alice/npm-heavy with 5.0 MiB, and give null once the tally is reset.

Not done: no test runs a full contained install from start to finish. Cells cannot run in this sandbox, so the tests drive the meter and npmMissLine directly, and the two places that write the line are not tested. I also didn't add the top repository to the terminal's Egress screen or /metrics.

To try it: as a site admin, open /settings/capacity on the preview after a check has installed with a cold cache, or call MCP capacity and read egress.npmTopRepoLastHour.

🤖 Generated with Claude Code

When a contained check, gate or app/preview install gets more than 1 MiB from registry.npmjs.org, its log now ends with install fetched N MiB from the npm registry; the shared cache missed. The forge also counts registry bytes for each repository over the last hour. The one with the most shows up in three places, all read from the same egressHour(db):

• the Egress connections card on /settings/capacity;
• MCP capacity, as egress.npmTopRepoLastHour;
• the terminal's Egress connections screen.

The repository comes from the code that opened the cell, never from anything the install printed (#8382).

Changes since review:

• Terminal: src/tui/screens/egress.ts now prints The npm registry sent X over the last hour, most to owner/name (Y).. test/egress-admin.test.ts asserts that line. docs/METRICS.md names egressHour as the source for all three places.
• Both log sites tested: the meter / install / log steps are now one function, meteredInstall in src/checks/install.ts. Both src/checks/service.ts and install() in src/apps/process.ts call it. test/egress-npm-repo.test.ts covers:

- the check case: the 7546 fetched line, then the miss line; - the app case: only the miss line; - an install that throws: the miss line is still logged; - an uncontained install: nothing is metered or logged.

• Repository kept: meterEgress now sets repoId when the first holder of a label had none. A test covers it.
• Behaviour change: the lines are written in finally, so failed, aborted or throwing installs log them too. That includes the check's install fetched … from the registry line, which before was skipped when the install threw. The doc comment and docs/METRICS.md say so.

Not done:

• Neither real call site is run through an actual cell, because cells need unix sockets, which this sandbox can't bind. Each call site is one call to the tested function.
• Repositories without a contained install are not counted for the top-repository figure.

To try it: as a site admin, open /settings/capacity, or in the terminal go to About, press c. After a cold contained install, the repository is named next to the registry's bytes for the hour.

Checks run here and passing: bun scripts/gate.ts, bun x tsc --noEmit, bun scripts/design-lint.ts, and bun test on these files:

• egress-npm-repo
• egress-admin
• capacity
• egress-totals
• egress-metrics
• check-install
• app-install-cache
• checks

🤖 Generated with Claude Code