Skip to content

feat(gateway): The Gateway reaches protected sites through an operator's egress proxy - #55

Draft
sodre wants to merge 9 commits into
mainfrom
e2/49-egress-proxy
Draft

sodre wants to merge 9 commits into
mainfrom
e2/49-egress-proxy

Conversation

@sodre

@sodre sodre commented Nov 22, 2026

Copy link
Copy Markdown
Member

Closes #50
Closes #51
Closes #52

Summary

An operator whose network gives the Gateway no direct route can now name an egress proxy once, as egress_proxy in the Gateway configuration, and every protected site is reached through it, while a rule with "egress": "direct" keeps its direct route. When the proxied leg fails, the Workload's 502 says whether the egress proxy refused, could not reach the site or could not be used, and the Gateway never connects directly instead.

Where the route is chosen, in connect_request (crates/countersign-gateway/src/runtime/mod.rs):

connect_request
  destination policy, Broker admission, both leases       # unchanged
  PeerTrust::load(rule)                                   # unchanged
  egress_proxy_for(rule)
    Some(proxy) → egress::tunnel(proxy, origin, denied)   # new: resolve the proxy, CONNECT site:port, read its answer
    None        → resolve_target + TcpStream::connect     # today's direct route
  site certificate check, grant worker, relay             # unchanged
  • Configuration (runtime/config.rs): egress_proxy is an http URL with a host and a written port. https, credentials, port 0 and private_networks on a proxied rule each fail loading by name, in env and serve alike. ADR 0003 is now accepted.
  • The proxied leg (runtime/egress.rs): the Gateway resolves only the egress proxy, on every connection, and drops answers inside denied_networks. It sends CONNECT <site>:<port>. It reads the answer one byte at a time and gives up as soon as the bytes cannot be HTTP. A 4xx is proxy_denied, a 5xx upstream_failure, and anything else proxy_unavailable.
  • The runtime test gains a test-only CONNECT proxy. Every negative case checks that the site saw no connection.
  • The Compose demo gains an opt-in variant, compose_demo.py --squid, with a real Squid pinned by digest; the default walkthrough is unchanged. docs/develop.md says how to run it.
  • Beyond the tickets, from review: the operator guide's Squid example pins a private site with a deny on every address outside its prefix. Squid's dst matches when any one of a name's addresses is in its list, and on the pinned Squid 6.14 the plain pin let a name through with one address outside it, then dialled that address. The demo's squid.conf carries the same deny.

Extras for the owner

  • env accepts configurations that serve refuses at its destination checks. A wildcard rule with private_networks (refused as "private egress requires an exact host and prefix rule"), and anything else Config::policy refuses, passes env and stops serve. This predates the change; every refusal this pull request adds is in both commands. An operator can send a Gateway's destinations through an egress proxy #50's criterion "env and serve refuse the same configurations" and the spec's story 15, "so that my launcher finds the mistake before the Gateway starts", read plainly, cover it.
  • No test covers a proxied leg that runs into the 120-second deadline. An egress proxy that accepts the connection and never answers gets proxy_unavailable when the Workload's 120-second deadline passes. The code maps it through the same catch-all as the tested unreachable, closed and non-HTTP cases, but the deadline is fixed and CI runs the runtime test under timeout 120. No source settles it.

Evidence

Criterion Ticked Shown by
#50: The Gateway configuration accepts an optional top-level egress_proxy, an http URL with a host and a port yes egress_configuration: an http egress proxy loads
#50: Loading fails on an https egress proxy URL and on one with credentials, and each error names the cause yes egress_configuration: https, credentials and a wrong shape are refused by name
#50: A destination rule accepts an optional egress whose only value is "direct", and any other value fails loading yes egress_configuration: egress takes only "direct"
#50: Loading fails when a rule has private_networks and goes through the egress proxy, and the error names the rule and both fixes yes egress_configuration: private_networks on a proxied rule is refused
#50: env and serve refuse the same configurations yes egress_configuration's refused(), which runs every refusal through both commands and asserts the same one line from each (a pre-existing gap is under Extras for the owner)
#50: A configuration without egress_proxy loads and behaves as before, and "egress": "direct" changes nothing yes The direct passes of runtime.py, unchanged and passing, the shared one with "egress": "direct" on its rule
#50: With egress_proxy set, the Workload reads the site through the test-only proxy, in sidecar and shared mode, and the site saw the approved certificate yes Proxied pass: the Workload reads the site through the egress proxy
#50: The proxy's recorded request names the site's hostname and port, where the Gateway cannot resolve that name yes Proxied pass: the Workload reads the site through the egress proxy
#50: One Gateway serves a proxied destination and a direct one with private_networks yes Proxied pass: proxied and direct destinations in one Gateway
#50: A retired grant cuts a proxied connection as it cuts a direct one yes Proxied pass: cutoff on the proxied route
#50: With proxy environment variables set and no egress_proxy, a destination is still reached directly yes Direct pass: proxy environment variables are ignored
#50: Startup logs the egress proxy's host and port and the rule IDs that leave their address policy to it yes Proxied pass: the startup line names the egress proxy and its destinations
#50: The runtime test runs these cases wherever it runs today, under both worker sandboxes yes runtime.py and scripts/filter-sandbox-container.sh, on 5d76a8b
#50: ADR 0003 is accepted and the glossary defines egress proxy yes The documentation grep run
#50: The configuration reference describes egress_proxy and egress yes The documentation grep run
#50: The operator guide has a section on reaching sites through an egress proxy: where resolution and address policy happen on each route, and what the ACLs must block yes The documentation grep run
#50: The architecture page's trust boundaries name the egress proxy, and the design record's resolution paragraph names both routes yes The documentation grep run
#51: A 4xx CONNECT refusal gives 502 proxy_denied, and the site saw no connection yes fails_through_egress: a refusal is proxy_denied
#51: A 5xx answer gives 502 upstream_failure yes fails_through_egress: a 5xx is upstream_failure
#51: Nothing listening at the proxy's address gives 502 proxy_unavailable, and the site saw nothing although the Gateway could reach it yes Unusable pass: nothing listens at the egress proxy
#51: The proxy closing the connection or answering with something not HTTP gives 502 proxy_unavailable yes fails_through_egress: closed or not HTTP is proxy_unavailable; kept-open answers fail at once
#51: The proxy's address inside denied_networks gives 502 proxy_unavailable, and neither proxy nor site saw a connection yes Unusable pass: the egress proxy's address is denied
#51: Through an allowed tunnel to a site certified for another hostname, the request fails and no signature is requested yes Proxied pass: wrong hostname gets no signature
#51: Each response carries Countersign-Request-ID and none carries the proxy's body yes fails_through_egress: a refusal is proxy_denied (every negative case asserts both)
#51: The Gateway logs each failure with its request ID and error value yes fails_through_egress: a refusal is proxy_denied (every negative case waits for Gateway CONNECT <id> <value>)
#51: The runtime test runs these cases wherever it runs today yes runtime.py and scripts/filter-sandbox-container.sh, on 5d76a8b
#51: The troubleshooting table covers proxy_denied and proxy_unavailable, and upstream_failure on a proxied route yes The documentation grep run
#52: The demo launcher has an opt-in Squid variant, and the default walkthrough is unchanged yes compose_demo.py --squid --verify; the default --verify and docker compose config runs
#52: In the variant the Gateway shares a network with the Broker and Squid, and the site is on a network only Squid shares yes --verify: the Gateway has no direct route to the site
#52: Squid runs from an image pinned by digest, allows CONNECT only to the site's hostname and port, and does no TLS interception yes --verify: Squid logged the tunnel to the site; the approved certificate reached the site
#52: From one --verify run, Squid logs a successful CONNECT, the site's fingerprint equals the approved one, and the Workload read the response yes compose_demo.py --squid --verify on 5d76a8b
#52: With the grant live and Squid's ACL swapped to refuse the site, the request fails with proxy_denied, and Squid logs the refusal yes --verify: the ACL swap gives proxy_denied; Squid logs the refusal
#52: With the grant live, the ACL restored and the site stopped, the request fails with upstream_failure yes --verify: a stopped site gives upstream_failure
#52: With the grant live and Squid stopped, the request fails with proxy_unavailable yes --verify: a stopped Squid gives proxy_unavailable
#52: The operator guide's example Squid ACLs are taken from the configuration that passed yes The guide's rules on the pinned Squid; the demo's squid.conf pins the site to every address
#52: The developer guide says how to run the variant and names the Squid version yes The documentation grep run
#52: The output of a passing --verify run of the variant is recorded on this ticket yes The comment on #52 with the run on 5d76a8b
Test Red Green
egress_configuration: an http egress proxy loads Before the field existed: AssertionError: http://proxy.test:3128 ok
egress_configuration: https, credentials and a wrong shape are refused by name serve accepted https and kept running: TimeoutExpired ... timed out after 10 seconds ok
egress_configuration: an empty port is refused TimeoutExpired: serve started on port 80 for http://proxy.test: ok
egress_configuration: port 0 is refused (review) subprocess.TimeoutExpired: Command '[... 'serve', '.../egress.json']' timed out after 10 seconds for http://proxy.test:0 ok
egress_configuration: egress takes only "direct" AssertionError (unknown field egress) ok
egress_configuration: private_networks on a proxied rule is refused serve started: TimeoutExpired ok
egress_configuration: no startup line when the destination checks fail (review) AssertionError: ... Gateway egress proxy proxy.test:3128 carries the address policy of destinations: site then Error: ... "private egress requires an exact host and prefix rule" ok
Proxied pass: the Workload reads the site through the egress proxy Before the proxied leg existed: ('alice', b'curl: (22) The requested URL returned error: 502\n') ok: sidecar/shared Gateway mTLS through the egress proxy passed, the proxy recorded exactly CONNECT site.test:<port> HTTP/1.1, and getent ahosts site.test exits 2 under the Gateway's environment
Proxied pass: cutoff on the proxied route Gateway broken on purpose, a leaked copy of the socket: AssertionError: upstream leg survived cutoff ok
Proxied pass: proxied and direct destinations in one Gateway Route choice broken on purpose, always the proxy: ('alice', b'curl: (22) ... 502\n') ok
Direct pass: proxy environment variables are ignored Gateway broken on purpose, reading HTTPS_PROXY: AssertionError: The Gateway took the egress proxy from its environment ok
Proxied pass: the startup line names the egress proxy and its destinations AssertionError: ['Gateway worker sandbox "namespaces": ...'] ok
fails_through_egress: a refusal is proxy_denied Before the error values existed: countersign-error: upstream_failure ok
fails_through_egress: closed or not HTTP is proxy_unavailable upstream_failure ok
fails_through_egress: a 5xx is upstream_failure proxy_unavailable, from the catch-all ok
fails_through_egress: kept-open answers fail at once (review): banner, stub, lf Against the old reading loop, each alone: TimeoutError: timed out (the Workload's 8-second read) ok: proxy_unavailable, proxy_unavailable, proxy_denied
Unusable pass: nothing listens at the egress proxy Passed on arrival; with a direct fallback added on purpose: ('site.test:…', 'The site saw a connection', b'HTTP/1.1 200 OK…') ok
Unusable pass: the egress proxy's address is denied Passed on arrival; with the filter on the proxy's addresses removed: 'The site saw a connection', and with that check skipped, AssertionError: The egress proxy saw a connection ok
Proxied pass: wrong hostname gets no signature Passed on arrival; with the site's certificate covering wrong.test: AssertionError: (0, b'd8bc83…'), and with that check skipped, A site with the wrong certificate got a signature ok
--verify: Squid logged the tunnel to the site Gateway still direct: RuntimeError: Timed out waiting for Squid's access log line for the Workload's CONNECT ok
--verify: the Gateway has no direct route to the site Site still on the backend: The Gateway can resolve site.test without Squid; pointed at broker:9443 on purpose: The Gateway can connect to the site at 172.31.0.2 without Squid ok
--verify: the startup line names Squid egress_proxy left out: Timed out waiting for the Gateway's egress proxy startup line ok
--verify: the ACL swap gives proxy_denied; Squid logs the refusal No swap: The Workload's request succeeded where it should fail with proxy_denied; port 9445 expected on purpose: Squid did not log one refusal… ok
--verify: the Gateway logs the same request ID …_wrong expected on purpose: Timed out waiting for the Gateway's log line… ok
--verify: a stopped site gives upstream_failure Site not stopped: …should fail with upstream_failure; HIER_NONE/- expected on purpose for Squid's 503 ok
--verify: a stopped Squid gives proxy_unavailable Squid not stopped: countersign-error: upstream_failure ok
The demo's squid.conf pins the site to every address (review) The demo's rendered squid.conf before the change, site.test answering 192.0.2.10 and 198.51.100.7: TCP_TUNNEL/503 0 CONNECT site.test:9444 - HIER_DIRECT/198.51.100.7 ok: TCP_DENIED/403 ... HIER_NONE/- in either order; 192.0.2.10 alone still tunnels (HIER_DIRECT/192.0.2.10)
The guide's rules on the pinned Squid (review) A copy with one line broken on purpose: FATAL: Bungled /check/broken.conf line 4: acl tls_port prot 443. The plain pin, billing answering one address inside and one outside: allowed, Squid dialled HIER_DIRECT/198.51.100.7 ok: squid -k parse exit 0; TCP_DENIED/403 for the mixed answer in either order, HIER_DIRECT/10.30.0.5 for the inside one; the block ACLs and deny rules equal squid.conf's line for line

Runs on the integration branch's tip, 5d76a8b, on Linux arm64:

  • cargo fmt --all -- --check: clean. cargo clippy --workspace --all-targets --locked -- -D warnings: clean. cargo test --workspace --locked: 19 suites, 43 passed, 0 failed, 1 ignored.
  • python3 crates/countersign-gateway/tests/runtime.py target/debug/countersign-gateway: exit 0 in 27 s (CI allows 120 s), printing sidecar Gateway mTLS passed, shared Gateway mTLS passed, sidecar Gateway mTLS through the egress proxy passed, sidecar Gateway without a usable egress proxy failed closed, the same two for shared, and filter-killed grant worker fenced.
  • scripts/filter-sandbox-container.sh: exit 0 in 33 s: the containment test, the same runtime-test lines under the filter sandbox, both named startup failures (refused to start for each mode), and the close_range unit (1 passed).
  • mkdocs build --strict: clean. python3 -m unittest discover -s scripts -p 'test_git_hooks.py': 12 tests, OK
  • dbus-run-session -- xvfb-run -a .venv/bin/python examples/demo-integrator/compose_demo.py --squid --verify --app <Companion>, with Squid 6.14 from the pinned image: exit 0. Its check lines are posted on A maintainer can qualify the Gateway against a real Squid in the Compose demo #52. The default walkthrough's --verify passed on 17525c6. The default docker compose config was identical before and after (196 lines), and examples/demo-integrator/compose.yaml is unchanged from origin/main.
  • The documentation grep run, at 5d76a8b, printed status: accepted for ADR 0003 and the glossary's **Egress proxy**. It also printed the configuration reference's egress_proxy and egress rows, and the operator guide's "Reach sites through an egress proxy" section with its "Who resolves the site's name" and "Who checks the site's addresses" rows and its ACL list. It found the troubleshooting row's proxy_denied, proxy_unavailable and "On a route through the egress proxy, the proxy answered 5xx because it could not reach the site". It found the architecture page's "Egress proxy" trust-boundary row and the design record's "A destination takes one of two routes". It found the developer guide's "Qualify the Gateway against a real Squid" section, which names Squid 6.14.

Not run: pnpm lint and pnpm test on this tip, since no JavaScript changed after 17525c6, where they passed (the demo Integrator's tests: 8 passed). The x86-64 runtime test and the CI filter-sandbox jobs run on this push. Repeating the run on Fargate 1.4.0 is the #46 reporter's, out of the spec's scope.

Merge Danger

Door: two-way. A revert removes egress_proxy and egress; a Gateway configuration that sets either would then fail to load (unknown field), so revert only after taking them out.

Blast Radius: the Gateway's route to a protected site, for operators who set egress_proxy.

Nothing deploys on the merge: the countersign-gateway crate and the ghcr.io/zeroae/countersign-gateway image ship at the next version tag. No flag: the change is an opt-in configuration field whose absence keeps today's behaviour. Live step at the merge, for whoever merges: comment on #46, pointing the reporter at the documented Compose configuration (docs/develop.md, "Qualify the Gateway against a real Squid", and the operator guide's "Reach sites through an egress proxy"), so they can repeat it on Fargate.

Reviews

Correctness, Standards, Spec and Security, three rounds: 18 findings, each with its outcome; rounds 2 and 3 found 1 and 0

Correctness, round 1 (at high, on 719bffc against origin/main)

  1. The guide's private-site pin is weaker than private_networks: Squid's dst matches any one address. Fixed: the example adds a deny on every address outside the prefix, measured on the pinned Squid.
  2. The proxied leg has no time limit of its own. Pushed back: the spec's Implementation Decisions say "The per-connection deadline covers the whole leg"; a proxy that accepts and never answers has not answered, which is story 32's condition.
  3. A deadline that runs out on a proxied route reads proxy_unavailable. Pushed back: the spec's failure table maps an egress proxy that "timed out" to proxy_unavailable.
  4. egress_proxy accepts port 0. Fixed: refused at load.
  5. A bare-LF answer blocks until the deadline. Fixed: a line ends at LF with or without CR, and a non-2xx status line returns at once.
  6. The troubleshooting row misses upstream_failure from a rule's roots or revocation files failing to load. Fixed: the row names them, for either route.
  7. The startup line prints before the destinations are checked. Fixed: it follows config.policy().
  8. The proxy's answer is read one byte per call. Pushed back: about 40 one-byte reads once per CONNECT, before a TLS handshake; reading a byte at a time is what leaves every tunnel byte to the tunnel.
  9. EgressProxy::parse repeats the URL-origin checks. Pushed back: the shared check answers "expected a URL origin", while the spec requires egress_proxy's errors to name the cause; the overlap is five boolean checks.
  10. A local egress sits beside the module egress::. Pushed back: the egress leg predates this change and pairs with ingress; the module is always called by path.

Standards, round 1 (smells only; no written rule broken)

  1. Possible Duplicated Code with resolve_target's denied filter. Pushed back: three lines; the two lookups differ (the 64-answer cap, permitted, one pinned address against trying each), and ADR 0003 keeps the two routes' policies apart.
  2. Possible Primitive Obsession: the raw and the parsed egress proxy side by side. Pushed back: parsing in Config::load keeps the error in its own words, beside the private_networks check that needs the parsed proxy.
  3. Possible Speculative Generality: Egress::Default cannot be written. Pushed back: the enum names a rule's route; every unknown value, default included, fails loading, as tested.
  4. Possible Repeated Switches: two tables give upstream_failure. Pushed back: one maps a status the Gateway chose, the other the egress proxy's answer (ADR 0003's table), each in one place.
  5. A second error-exit convention in run. Pushed back: main prints run's other errors in their existing form, which the startup-failure checks read; only the load error needed its own words, as An operator can send a Gateway's destinations through an egress proxy #50 asks.

Spec, round 1

  1. A non-HTTP answer with the connection kept open waits for the deadline (A Workload learns why a proxied connection failed, and the Gateway never connects directly #51; Reach protected sites through an operator's egress proxy #49 story 32). Fixed: bytes that cannot begin HTTP/1. fail at once with proxy_unavailable.

Security, round 1 (the path list could not be read, so every changed path was judged as a control)

  1. LOW: the guide's Squid pin lets a name with one address outside the prefix reach that address, never with a signature. Fixed: as finding 1.

Spec, round 2 (fix round at low, on 47cdf6a against 719bffc; Correctness, Standards and Security found nothing)

  1. The guide's new deny was not in the configuration the Compose variant qualified (A maintainer can qualify the Gateway against a real Squid in the Compose demo #52, "taken from the configuration that passed"). Fixed: the demo's squid.conf denies the site's name on every address but the site's, and the variant's --verify passed again on 5d76a8b.

Round 3 (fix round at low, on 5d76a8b against 47cdf6a): Correctness, Standards, Spec and Security found nothing.

Tracking

Epic: E2 Reach protected sites through an operator's egress proxy
Spec: #49 Reach protected sites through an operator's egress proxy
Ticket: #50 An operator can send a Gateway's destinations through an egress proxy
Ticket: #51 A Workload learns why a proxied connection failed, and the Gateway never connects directly
Ticket: #52 A maintainer can qualify the Gateway against a real Squid in the Compose demo
Flag: none

🤖 Generated with Claude Code

sodre and others added 9 commits November 22, 2026 02:26
…an egress proxy

An optional top-level `egress_proxy` (an http URL with a host and a port)
sends every destination through the operator's HTTP proxy, unless its rule
says `"egress": "direct"`. On a proxied route the supervisor resolves only
the proxy's host, on every connection and outside `denied_networks`, sends
a plain-text CONNECT with the site's canonical name and port, and hands the
tunnel to the egress relay where the direct socket went. It resolves
nothing for the site and never tries a proxied destination directly. Until
the failure values land, any failure of the proxied leg is today's 502
`upstream_failure`.

Loading refuses an https proxy URL, one with credentials, any other
`egress` value, and `private_networks` on a proxied rule, naming the rule
and both fixes. `env` and `serve` refuse the same configurations, and print
the cause in its own words. Startup logs the egress proxy and the rules
that leave their address policy to it.

The runtime test gains a test-only CONNECT proxy that records each request
line. Through it, in sidecar and shared mode, the Workload reads the site
with the approved certificate while the Gateway cannot resolve the site's
name; one Gateway serves a proxied and a direct rule with
`private_networks`; lease lapse and retirement cut a proxied connection;
and proxy environment variables without `egress_proxy` change nothing.

ADR 0003 is accepted. The configuration reference, operator guide,
architecture page and design record describe both routes.

Ticket: #50

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
… the Gateway never connects directly

When the proxied leg fails, the Workload's CONNECT returns 502 with
`Countersign-Request-ID` and a `Countersign-Error` that says why:
`proxy_denied` for a 4xx answer from the egress proxy, `upstream_failure`
for a 5xx answer, as a direct route reports a site it cannot reach, and
`proxy_unavailable` when the proxy cannot be reached, resolves only inside
`denied_networks`, runs out the deadline, closes unanswered or answers with
something that is not HTTP. The proxy's own response body is never passed
on, and the Gateway logs each failure with its request ID and error value.

The runtime test's egress proxy learns to refuse with an error page of its
own, to close unanswered, to answer with something that is not HTTP, to
answer 503 when it cannot reach the site, and to tunnel a name it allows to
a site whose certificate is for another name. In sidecar and shared mode
the Workload reads each error value and no body. Two more Gateways per mode
have no usable egress proxy, one with nothing listening at its address and
one with its address inside `denied_networks`, while they can resolve and
reach the site themselves. The site and the test proxy record every
connection they accept, before any handshake, and every negative case
asserts the site saw none, and for `denied_networks` that the proxy saw
none either. Through the tunnel to the wrong certificate the request fails
and the test grant authority, which now records each signature it is asked
for, is asked for none.

The troubleshooting table covers a 502 and its three values, and the
operator guide's egress proxy section names them.

Ticket: #51

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
… in the Compose demo

`compose_demo.py --squid` layers `compose.squid.yaml` over the default
`compose.yaml`, which stays as it was. Squid runs from
`ubuntu/squid:6.6-24.04_edge`, pinned by digest, as the demo's user with a
read-only root and no capabilities. The Gateway shares the backend network
with the Broker and Squid; the site moves to a network only Squid shares,
in the network of a placeholder container, so stopping the site leaves its
name and address behind. The Gateway's configuration names
`http://squid:3128` as its egress proxy and its rule has no
`private_networks`. `squid.conf` allows CONNECT only to site.test:9444 on
the site's address, blocks metadata, link-local and loopback addresses, and
does no TLS interception.

With `--verify` the launcher checks that the Gateway logged the egress
proxy and can neither resolve the site's name nor reach its address. From
the approved request it correlates Squid's TCP_TUNNEL/200 line for the
CONNECT to the site's address, the fingerprint the site returned and the
approval, and the job page. Then, reusing that grant, it swaps Squid's
rule to refuse the site and reconfigures it (502 proxy_denied, Squid logs
TCP_DENIED/403 with no connection to the site), restores it and stops the
site (502 upstream_failure, Squid logs TCP_TUNNEL/503), and stops Squid
(502 proxy_unavailable). Each failure carries a request ID the Gateway
logs with the same error value.

The operator guide's egress proxy section gains example Squid ACLs taken
from that configuration and the answers Squid 6.14 gives. The developer
guide says how to run the variant and names the Squid version it was
qualified against.

Ticket: #52

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…iting for the deadline

Review of spec #49, round 1:

- The proxy's answer is checked as it arrives. Bytes that cannot be an
  HTTP answer fail at once with proxy_unavailable, and a non-2xx status
  line returns at once, even when the proxy keeps the connection open.
  Lines may end in a bare LF (RFC 9112, 2.2).
- egress_proxy refuses port 0 at load.
- The egress proxy startup line comes after the destinations pass their
  checks.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…ery address, as private_networks does

Review of spec #49, round 1: Squid's dst ACL matches when any one of a
name's addresses is in its list, measured on the pinned Squid 6.14, so
the example adds a deny on every address outside the prefix. The
troubleshooting row says upstream_failure can also mean the rule's roots
or revocation files did not load.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…addresses is another

Review of spec #49, round 2: the operator guide's every-address pin was
not in the configuration the Compose variant qualified. The demo's
squid.conf now denies the site's name on every address but the site's,
as the guide's example does for a private site.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

This branch has not been deployed

No deployments
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

1 participant