Second review pass on the outbound proxy work.
The installed dispatcher was undici's EnvHttpProxyAgent, which reuses the HTTP
proxy for `https:` whenever no HTTPS proxy is present. That is exactly the state
this package resolves after refusing a SOCKS or malformed URL the user named for
`https:`, so the scheme the diagnostic reported as direct was tunnelled anyway.
The dispatcher is now an Agent whose per-origin factory calls `proxyForUrl`, so
routing and `proxyForUrl` cannot disagree by parsing the same list twice.
`childProxyEnv` returned only the names the user exported, which left a child
Node direct whenever the proxy came from `ALL_PROXY` or from cordis.yml — Node's
`NODE_USE_ENV_PROXY` reads neither — and stripped the merged loopback bypass so
the child sent its own localhost traffic to the proxy. A scheme the user named
in either casing still reaches the child exactly as written; one they named in
neither now carries the resolved value, and the bypass list is always the merged
one.
A nested install (the plugin mounted over the launcher's policy) recorded the
outer policy's published values as the user's, then cleared the record on
disposal, so every later child inherited the normalization instead. The record
now belongs to the outermost install and is restored, not dropped.
`web_fetch` read the active policy twice — once to skip address pinning, again
inside the transport — so a disposal landing between the two reads produced an
unpinned direct connection to a host nothing validated. One snapshot now decides
both.
Also: the node:http proxy test asserted a route the engines range does not always
have, and the gate could not see undici bound through `await import('undici')`,
the form this repository actually uses.
The workflow worker no longer receives proxy configuration: it executes the
model-authored script body, and a proxy URL may carry credentials. A child
process now inherits the values the user exported rather than this process's
normalization, so a SOCKS proxy set for curl survives and no HTTPS_PROXY is
invented. `mode: 'off'` installs a direct dispatcher instead of recording a
policy the global dispatcher ignores, and the environment snapshot is taken
before any write so Windows restores the user's values.
E2B picks its proxy from the control-plane URL the SDK will really call, the
OTLP agent honors `exporter.keepAlive`, and a scheme whose own value was
refused stays direct instead of borrowing another scheme's proxy.
verify-no-bare-dispatcher parses the TypeScript AST as scripts/AGENTS.md
requires; it immediately found the `{ dispatcher }` shorthand the regex missed.
NODE_USE_ENV_PROXY reaches Node 24.0+ and 22.21+, while engines admits 22.19.
Assert the direct connection on an older runtime instead of only the proxied
one, so the seam is executable rather than prose.
Node's built-in fetch ignores HTTP_PROXY, so every harness request connected
directly regardless of what the user exported. Resolve one policy from the
launch environment and install it as undici's global dispatcher, then wire the
four surfaces a global dispatcher cannot reach: web_fetch's pinned transport,
the OTLP exporter's node:http agent, the E2B SDK's own proxy option, and the
environment a child process or worker thread is given.
Each outbound call site carries an egress test that drives its real code path
through a fake proxy; that measurement is what found the OTLP and E2B gaps.
Review found the shared observation defeated the very fence it fed:
it carries the original PID-to-start-time pairing forward, so a
recycled PID still matches it and takes a signal meant for the process
that exited. Capturing it outside the per-member try also let one
failed read abort a whole teardown round, breaking the synchronous
host-exit contract, and an empty round paid a read for no members.
signalProcess now reads ProcessInspector.isAlive immediately before
delivering, from the narrowest per-identity source each platform
offers; signalMembers and waitForMembers return before capturing when
a round has no members. snapshot() keeps serving the readiness poll,
whose per-poll table read stays at one.
Windows enumerates Toolhelp32 lazily on the first tree question, so a
snapshot asked only for liveness — the 25 ms teardown poll — performs
no table walk at all.
MacProcessInspector answered the descendant tree and every member's
liveness with its own `/bin/ps` fork, so one readiness poll cost N+1
full table reads for N tracked descendants. With execFileSync on that
path and a 50 ms poll interval, any command spawning two or more
children saturated the host event loop until it exited.
ProcessInspector.snapshot() now returns one ProcessSnapshot that
answers tree, session, and alive from a single observation, and
signalProcess takes the caller's observation so its PID-reuse fence
does not re-read the table per member.