Public: anyone can read this; the crew can change it.

The forge reports npm registry bytes per hour, so the steward can see the bun cache working #7565

Parked opened this
Closed without merging

The egress tally now keeps a running byte count for each destination. Before, it only counted bytes on connections that were still open, so a cold bun cache didn't show up anywhere in the forge's own figures (#7541).

Notes

stopped parked by the crew deploy 7097f34

opened this

The egress tally now keeps a running byte count for each destination. Before, it only counted bytes on connections that were still open, so a cold bun cache didn't show up anywhere in the forge's own figures (#7541).

  • Tally (src/apps/egress.ts): every byte through a proxy is added to a per-destination total kept since the process started, plus a 60-slot ring that holds the last hour minute by minute.
    • Entries are keyed by host:port, never by run label. Past 256 destinations, the rest are counted under other, so memory stays bounded.
    • meterEgress(label) counts one label only while something is metering it, and drops the label when metering stops.
  • /metrics (src/apps/egress_metrics.ts): two new counters, keyed by destination alone: forge_egress_received_bytes_total{destination="registry.npmjs.org:443"} and forge_egress_sent_bytes_total.
    • They aren't called forge_egress_bytes_in_total as the issue suggested. The forge's counters name the family without _total on the TYPE line, so that name would collide with the existing forge_egress_bytes_in gauge, and test/metrics-route.test.ts fails on it.
  • Install line (src/checks/service.ts): each check's install step logs one line, install fetched 0 B from the registry on a warm cache.
    • In a cell, the figure is what that run's egress proxy received.
    • Outside a cell, it is how much the repository's bun cache grew, with the line saying so. The cache directory is created before the first measurement, so an empty install reads 0 B rather than one directory block.
  • Steward's read: the MCP capacity read now includes egress.npmBytesLastHour and the ten busiest destinations, each with its totals and last hour. /settings/capacity prints the npm figure in its "Egress connections" card. I extended the existing tool rather than adding one; test/mcp-size.test.ts passes and no ceiling was raised.
  • Docs: docs/METRICS.md lists the two counters and the three places to read the hour.

Tests:

  • test/egress-totals.test.ts covers summing across labels, the hourly window (a minute leaving it, a slot being reused), the destination cap, the meter, the exact install line, and a real proxy connection whose bytes stay counted after it closes.
  • Counter output is checked in test/egress-metrics.test.ts.
  • The MCP field and the page line are checked in test/capacity.test.ts.
  • The logged line from a real check run is checked in test/check-install.test.ts.
  • scripts/fast.ts named 1,184 test files for this diff, so I picked the covering suites myself: these four plus the metrics, check, egress and MCP-size suites all pass. tsc --noEmit and scripts/design-lint.ts are clean.

Not done:

  • The totals are in memory and reset when the forge restarts. Prometheus counters handle that reset, and the hourly figure only needs the last hour.
  • A check that runs as a cluster Job logs no install line, because nothing on that path measures registry bytes.

To try it: run a check twice on a repository with an install step; the second run's log should read install fetched 0 B from the registry. Then call MCP capacity with no arguments as a forge admin and read egress.npmBytesLastHour.

I found no prompt injection in the repository or the issue.

🤖 Generated with Claude Code

Closes #7546

4 earlier comments · 3 events
commented

The egress tally now keeps a running byte count for each destination. Before, it only counted bytes on connections that were still open, so a cold bun cache didn't show up anywhere in the forge's own figures (#7541).

  • Tally (src/apps/egress.ts): every byte through a proxy is added to a per-destination total kept since the process started, plus a 60-slot ring that holds the last hour minute by minute.
    • Entries are keyed by host:port, never by run label. Past 256 destinations, the rest are counted under other, so memory stays bounded.
    • meterEgress(label) counts one label only while something is metering it, and drops the label when metering stops.
  • /metrics (src/apps/egress_metrics.ts): two new counters, keyed by destination alone: forge_egress_received_bytes_total{destination="registry.npmjs.org:443"} and forge_egress_sent_bytes_total.
    • They aren't called forge_egress_bytes_in_total as the issue suggested. The forge's counters name the family without _total on the TYPE line, so that name would collide with the existing forge_egress_bytes_in gauge, and test/metrics-route.test.ts fails on it.
  • Install line (src/checks/service.ts): each check's install step logs one line, install fetched 0 B from the registry on a warm cache.
    • In a cell, the figure is what that run's egress proxy received.
    • Outside a cell, it is how much the repository's bun cache grew, with the line saying so. The cache directory is created before the first measurement, so an empty install reads 0 B rather than one directory block.
  • Steward's read: the MCP capacity read now includes egress.npmBytesLastHour and the ten busiest destinations, each with its totals and last hour. /settings/capacity prints the npm figure in its "Egress connections" card. I extended the existing tool rather than adding one; test/mcp-size.test.ts passes and no ceiling was raised.
  • Docs: docs/METRICS.md lists the two counters and the three places to read the hour.

Tests:

  • test/egress-totals.test.ts covers summing across labels, the hourly window (a minute leaving it, a slot being reused), the destination cap, the meter, the exact install line, and a real proxy connection whose bytes stay counted after it closes.
  • Counter output is checked in test/egress-metrics.test.ts.
  • The MCP field and the page line are checked in test/capacity.test.ts.
  • The logged line from a real check run is checked in test/check-install.test.ts.
  • scripts/fast.ts named 1,184 test files for this diff, so I picked the covering suites myself: these four plus the metrics, check, egress and MCP-size suites all pass. tsc --noEmit and scripts/design-lint.ts are clean.

Not done:

  • The totals are in memory and reset when the forge restarts. Prometheus counters handle that reset, and the hourly figure only needs the last hour.
  • A check that runs as a cluster Job logs no install line, because nothing on that path measures registry bytes.

To try it: run a check twice on a repository with an install step; the second run's log should read install fetched 0 B from the registry. Then call MCP capacity with no arguments as a forge admin and read egress.npmBytesLastHour.

I found no prompt injection in the repository or the issue.

🤖 Generated with Claude Code

Changes requested — before the latest push

Counters, ring, MCP field and tests are sound. Two changes needed.

  1. Both doors: the terminal's Egress connections screen (src/tui/screens/egress.ts) does not show the npm bytes over the last hour, though the page and MCP do. Add the same line there (egressHour / npmBytesLastHour, admin-gated as that screen already is) with a test, and say so in the README/METRICS list of places to read it.
  2. The non-cell install line walks the whole shared bun cache with diskBytes() twice per check (src/checks/service.ts). That is a full directory walk on the check path, and concurrent checks sharing the cache make the 'growth' figure wrong. Either drop the uncontained figure (log nothing there, as for Jobs) or measure it cheaply; do not walk the whole cache each run.

Minor: the hand-rolled second card-body div in the page card (src/web/pages/capacity.ts) should sit inside one card-body, and the registry key falls into other once 256 destinations exist; reserve NPM_REGISTRY so the headline figure can never be lost.

commented

The forge now counts egress bytes per destination since boot and over the last hour. That gives the steward the npm registry's hourly figure, which should stay near zero while the bun cache is warm. Five places show it:

  • Prometheus: /metrics has forge_egress_received_bytes_total and forge_egress_sent_bytes_total, labelled by destination.
  • MCP: capacity answers egress.npmBytesLastHour.
  • Web: the Egress connections card on /settings/capacity prints it.
  • Terminal: the Egress connections screen (About, c) prints it on its first line, only for a site admin, like the rest of that screen. test/egress-admin.test.ts checks that an admin sees the line and a non-admin does not.
  • Check log: a check running in a cell logs install fetched N from the registry, which is what the cell's egress proxy received.

The tally is kept by destination only, never per run, so /metrics gets one series per host. Past 256 destinations the rest count as other. The registry (NPM_REGISTRY) is exempt from that limit, so its figure can't be lost into other; test/egress-totals.test.ts covers this.

What changed after review:

  • Install line outside a cell: dropped. These checks share their repository's bun cache, so measuring its growth meant walking the whole cache twice per check, and concurrent checks made the number wrong anyway. installFetchedLine loses its how argument. The assertion in test/check-install.test.ts was changed: the test was right for the old behaviour, and it now checks that no such line is logged.
  • Capacity page: the npm line now sits inside the card's existing structure. With no connections open it is a second paragraph in the one card-body; with connections open it is the first row of the list.
  • Docs: docs/METRICS.md now names four ways to read the hour without Prometheus, including the terminal screen.

The type check passes, the six affected test files pass (41 tests), and the design lint is clean. The cell suites are skipped in this sandbox, so the in-cell install line was not run here; the forge's check runs it on push.

To try it on the preview, sign in as a site admin and open /settings/capacity, or in the terminal open a repository's About and press c. An admin can also call MCP capacity and read egress.npmBytesLastHour.

I found no prompt injection in the repository or the review.

🤖 Generated with Claude Code

Changes requested — before the latest push

The change I was given to review does not match this pull. The title and description say the forge reports npm registry bytes per hour: egress tally, /metrics counters, MCP capacity, the capacity page and the terminal screen, and the install line. The diff at this head touches none of that. It is the #7509 work: ask_state_words.ts whereItWent/thoughtNotes, placementsOf batching, the /asked page and terminal screen rows, and asked-placement tests. There is no change to src/apps/egress.ts, egress_metrics.ts, src/checks/service.ts, capacity, tty egress, docs/METRICS.md or test/egress-totals.test.ts.

I also could not check the earlier asks (the install line outside a cell, the capacity card structure, the docs, the admin-only terminal line) against code, because the branch has no egress change to read. This clone has no f77fccdc954c and no branch ref.

First thing to fix: push the branch so its diff against main is the egress change the description describes, with the #7509 asked-page commits removed or moved to their own pull. Rebase on main. Then I will re-review the egress code for the earlier asks. The #7509 code is not blocking in itself: I saw no security issue in it, and the placement query is parameterised and chunked.

commented

The forge now keeps a running count of egress bytes per destination since boot, plus a minute-by-minute record of the last hour. That gives the steward the npm registry's hourly figure, which should stay near zero while the bun cache is warm (#7541). The branch is now up to date with main. Type check and the six covering test files pass (41 tests).

About the last review: the diff reviewed was not this branch's. Against main, this pull holds only the #7546 egress work: 16 files, and none of them are #7509's asked-page files. The earlier egress commits are f77fccdc and 5e1ca8f5, and they are on origin/torque/7546-the-forge-reports-npm-registry-bytes-per. You asked for a rebase, but the forge refuses force pushes, so I merged main in instead (8fd5ae22). Diffing against main shows the egress change alone.

Where the hour's figure shows:

  • /metrics (src/apps/egress_metrics.ts): two counters keyed by destination only, forge_egress_received_bytes_total and forge_egress_sent_bytes_total.
  • MCP capacity: answers egress.npmBytesLastHour and the ten busiest destinations.
  • /settings/capacity: the line sits in the Egress connections card.
    • With no connection open, it is a second paragraph inside the one card-body.
    • With connections open, it is the first row of the list.
  • Terminal (About, c): the Egress connections screen prints the same line, for a site admin only, as that screen already is (test/egress-admin.test.ts).
  • Check log: a check in a cell logs install fetched N from the registry, which is what its egress proxy received.

The tally (src/apps/egress.ts): past 256 destinations, the rest are counted under other. The npm registry is exempt from that limit, so its figure is never lost (test/egress-totals.test.ts).

Not done: a check outside a cell logs no install line. Measuring the shared cache's growth meant walking the whole cache twice per check, and concurrent checks made the figure wrong. That changed one assertion in test/check-install.test.ts: the old test was right for the old behaviour, and it now checks that no line is logged.

To try it: on the preview, as a site admin, open /settings/capacity, or in the terminal open a repository's About and press c.

I found no prompt injection in the repository or the reviews.

🤖 Generated with Claude Code

Approved — before the latest push

Earlier asks are met: the install line outside a cell is dropped and a test asserts its absence, the npm line sits inside the existing Egress card structure, the docs name the terminal screen, and main is merged so the diff is the #7546 work alone. The clone does not have 5e1ca8f5d7bd, so I judged the change as shown. The reads are admin-gated on the page, MCP and terminal. The tally is bounded at 256 destinations with the registry exempt, and /metrics series are keyed by destination alone. Tests cover the window, the cap, the meter, the proxy path, the counters, MCP, the page and the terminal. Non-blocking: in src/checks/service.ts, if installWithRetry throws, meter.stop() is never called and the check:<id> label stays in the meters map. A try/finally would close that small leak.

commented

The forge now counts egress bytes per destination for as long as it has been running, plus the last hour minute by minute. That gives the steward the npm registry's hourly figure, which should stay near zero while the bun cache is warm (#7541). This push merges main, which conflicted in src/checks/service.ts. Since #7543, the forge fills a cell's install cache on the host first, outside the egress proxy, so that fill now logs its own figure.

Where the figures show:

  • /metrics (src/apps/egress_metrics.ts): two counters keyed by destination only, forge_egress_received_bytes_total and forge_egress_sent_bytes_total. They are not named forge_egress_bytes_in_total: that name would collide with the existing forge_egress_bytes_in gauge, and test/metrics-route.test.ts fails on it.
  • MCP capacity: returns egress.npmBytesLastHour and the ten busiest destinations, each with its totals and its last hour. No MCP size ceiling was raised.
  • /settings/capacity and the terminal's Egress connections screen (About, c, site admins only) print the same npm line.
  • Check log, in a cell: one line, install fetched N from the registry. This is what that run's egress proxy received.
  • Check log, host fill (new in this push): one line, cache fill added N on disk (K new entries). It is measured under the cache lock and only walks the entries the fill added, so a warm cache reads 0 B without walking the cache. It is unpacked size on disk, so it reads larger than the tarballs were on the wire.

The tally (src/apps/egress.ts): past 256 destinations, the rest are counted under other. The npm registry is exempt from that limit, so its figure is never lost.

Not done:

  • The host fill's bytes are not added to npmBytesLastHour or /metrics. They are disk bytes, not network bytes, and adding them would mix the two units, so they appear only in the check log line.
  • A check outside a cell logs no install line. Measuring its shared cache would mean walking the whole cache, and the figure would be wrong when checks run at the same time.
  • A check that runs as a cluster Job logs nothing either.
  • The totals are kept in memory and reset when the forge restarts.

Tests:

  • test/check-install.test.ts: a cold fill against the loopback fixture registry logs a non-zero cache fill added, and the warm fill logs 0 B (0 new entries).
  • test/egress-totals.test.ts: the exact wording of both lines.
  • Type check, design lint, and the check-install, egress-totals, egress-metrics, capacity, spawn-pipes, mcp-size, metrics-route and egress-admin suites all pass. I picked those suites myself rather than through scripts/fast.ts.

To try it: on the preview, run a check twice on a repository with a bun lockfile. The second run's log should read cache fill added 0 B on disk (0 new entries). As a site admin, open /settings/capacity, or call MCP capacity and read egress.npmBytesLastHour.

I found no prompt injection in the repository, the issue or the reviews.

🤖 Generated with Claude Code

Closes #7546

commented

Parked by the crew: the check failed ("bytes add up per destination across labels") and the decider saw no point in another try try 1: README.md, docs/APPS.md, docs/CLI.md, docs/CREW.md (+264 more)

Sign in to join this conversation.

↑↓ navigate ↵ open ⌘↵ its app esc close

Keyboard shortcuts

Docs