镜像站点 · 本页由第三方 GitHub 只读镜像提供，非 GitHub 官方站点，不接受任何登录或凭据输入。前往 github.com
Skip to content

Forward Claude Code and goose inference to the agent gateway natively - #2916

Open
philazar wants to merge 3 commits into
coder3/dxa-1154-pi-workforcefrom
coder1/dxa-1156-host-forwarding
Open

Forward Claude Code and goose inference to the agent gateway natively#2916
philazar wants to merge 3 commits into
coder3/dxa-1154-pi-workforcefrom
coder1/dxa-1156-host-forwarding

Conversation

@philazar

@philazar philazar commented Nov 6, 2026

Copy link
Copy Markdown
Contributor

Why

Buzz can send a local agent's turn through an agent gateway that a build configures (#2632). Claude Code and goose make their own model calls, so the app needs a native way to send those calls to the gateway with the turn's context headers. It must not take over the harness's credentials, request bodies or streaming.

What

  • A native loopback transport forwards a local Claude Code harness's model traffic to the build's agent gateway.
  • The same transport forwards goose inference, tied to the goose session it serves.
  • Typed frontend bindings cover the host commands, including associateForward for goose.

How

  • Each turn opens a forwarding generation. The page's frozen context headers are added natively. The harness keeps its own credentials, bodies, streaming, retries and cancellation.
  • Admission follows the page's route revisions. Page reloads and shutdown stop every generation. Only compiled routes are forwarded, and redirects are never followed.
  • The gateway origin, header names, reserved prefixes and per-harness upstream bases come only from the version 2 BUZZ_AGENT_GATEWAY build descriptor. One validator in plugin-manager serves build.rs and the app. It refuses an invalid descriptor without echoing it. Unset means off.
  • A goose generation opens for discovery without a session. It is tied to the session goose reports, once, before the first prompt. Inference must then name that session in agent-session-id. Only goose's serving-endpoints discovery and invocation routes are forwarded.
  • The last commit merges the Pi adapter branch (feat(pi): run gateway-eligible Pi agents through the configured model gateway #2479). The merge had no conflicts.

Risk

Medium in code size (about 3,600 added lines, mostly Rust), low in effect today. No caller turns on forwarding in this PR, and an unset descriptor means off. The Claude and goose callers come in the PRs stacked on this one.

Testing

Not run yet at this head:

  • The full native Rust suite. CI runs it.
  • Native app acceptance with a live gateway. No live run is approved yet.

Bigger picture

Review order: #2632 (shared gateway client) → #2479 (Pi) → this PR (native forwarding) → Claude caller → goose caller. Each PR's diff shows only its own layer. The stack conflicts with current main and needs a main integration before merge. Ready for review does not mean ready to merge.

🤖 Generated with Claude Code

philazar and others added 3 commits November 5, 2026 16:37
@philazar coder2 @claude
feat(host): forward Claude Code inference to the agent gateway
f6becc4 
Add a native loopback transport that forwards a local Claude Code
harness's model traffic to the build's agent gateway, composing the
page's frozen context headers in natively. The harness keeps its own
credentials, bodies, streaming, retries and cancellation. Admission
follows the page's route revisions, page reloads and shutdown stop every
generation, only compiled routes are forwarded, and redirects are never
followed. Typed frontend bindings cover the four host commands.

The gateway's origin, context header names, reserved prefixes and
per-capability upstream bases come only from the version 2
BUZZ_AGENT_GATEWAY build descriptor. One validator in plugin-manager
serves build.rs, which refuses an invalid descriptor without echoing it
and compiles its canonical form, and the app, which parses that form
again at startup and keeps only what native forwarding uses. Unset means
off. The validator runs the frontend's canonical descriptor vectors.

Co-authored-by: coder2 <4acc49cb93576448ba82e631688c51a40bf03d6a889d78fddf57917d198de10f@buzz.block.builderlab.xyz>
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Signed-off-by: Philip Azar <pazar@squareup.com>
@philazar coder2 @claude
feat(host): forward goose inference to the agent gateway
59b5400 
Extend the native transport to goose. A goose generation opens for
discovery without a session and is associated with the session goose
reports, once, before its first prompt; inference must then name that
session in `agent-session-id`. Only goose's serving-endpoints discovery
and invocation routes are forwarded, with plain endpoint names. Its
upstream base comes from the descriptor's `goose` capability, and a build
that configures none leaves goose unforwarded while Claude Code may still
forward. Typed bindings add `associateForward`.

Co-authored-by: coder2 <4acc49cb93576448ba82e631688c51a40bf03d6a889d78fddf57917d198de10f@buzz.block.builderlab.xyz>
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Signed-off-by: Philip Azar <pazar@squareup.com>
@philazar Coder1 @claude
Merge the Pi gateway adapter into the native gateway forwarding
ea2e2f0 
No conflicts.

Co-authored-by: Coder1 <db50ef1f1688dc7f3e536c17351e8e994a7466bcbe009794f61d64355c0ffc2e@buzz.block.builderlab.xyz>
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Signed-off-by: Philip Azar <pazar@squareup.com>
@philazar
philazar requested a review from a team as a code owner November 6, 2026 23:16
@philazar philazar mentioned this pull request Nov 6, 2026
Open
2 tasks
@baxen baxen mentioned this pull request Nov 7, 2026
Merged
baxen added a commit that referenced this pull request Nov 7, 2026
@baxen mbot
test(browser): fix three flaky journeys seen across open PRs (#2957)
bdc015f 
## Summary

These are fixes for three browser-journey flakes that are failing open
PRs which don't touch the code under test. The change is test-only. No
browser cases are added or removed, and no assertion is relaxed.

The red `Browser journeys (chromium, 3/9)` / `(webkit, 4/14)` shards on
most open PRs come from a different problem:
`sidebar-geometry.spec.mjs:224` is broken on main. #2945 and #2946
already fix it, so this PR leaves it alone.

### 1. `channel-completion.spec.mjs`: cold dev-server load timed out

**Evidence:** `Flake check` on #2935 failed
`channel-completion.spec.mjs:7` on both engines with `page.goto: Timeout
15000ms exceeded` on
`/tests/fixtures/mentions.html?channels&stream&test-controls`
([chromium](https://github.com/block/buzz-app/actions/runs/42317084302/job/128232679087),
[webkit](https://github.com/block/buzz-app/actions/runs/42317084302/job/128232678918)).
#2935 doesn't touch that page.

**Cause:** this is the same cold load #2776 fixed for
`mentions.spec.mjs`. Each test here started its own Vite dev server for
the same ~1,600-module fixture.

**Fix:**
- `beforeAll` builds `mentions.html` once and serves it with
`build`/`preview`, using the same config as `mentions.spec.mjs`.
- The build keeps the dev server's CSS. Tailwind's optimizer and Vite's
CSS minifier each round the 20/14 line-height token to `1.42857`, so
line boxes become `19.99998px`. In WebKit that put the popup `4.03125px`
above the composer, and the test asserts exactly `4`. I confirmed this
locally: it fails 5/5 with either setting left on and passes with both
off. Turning both off keeps the geometry the test checked before.

### 2. `profiles.spec.mjs:940`: reopening Members right after closing
the profile

**Evidence:** `Flake check (webkit, 8/9)` on #2909 failed in repeat 8 of
10
([job](https://github.com/block/buzz-app/actions/runs/42313241314/job/128219877861)).
The trace shows this sequence:
1. `Escape` closes the profile panel.
2. Focus returns to "Manage members of One" within 40ms.
3. `Enter` is pressed 10ms later.
4. The DOM after that `Enter` shows the Profile panel again and no
Members dialog, so the next `row.click` times out.

**Fix:** after each `Escape`, wait for the address to leave the profile
(no `"detail"` in the `#buzz=` hash) before reopening Members. This is a
lifecycle barrier, not a delay. One step in this test already waited for
the panel to unmount before reopening; the other steps didn't wait at
all.

**Product follow-up (not fixed here):** the order of a dismissal's
history return and a quick reopen matters.
- If I delay `history.go` by 300ms, the unpatched test takes a different
path: the dismissal's `go(-1)` never runs, and the URL stays on the
profile visit under the reopened dialog.
- A `browser-history` unit sketch (deferred traversal: `push(channel)`,
`push(panel)`, `dismiss([panel])`, then `push(settings)` before the
traversal commits) lands on `channel`. The new `settings` visit is
undone.

Someone who owns navigation should decide whether a push during a
pending dismissal return should be deferred or replayed. That is a
navigation-history change, not a test fix.

### 3. `timeline.mjs` `edge()`: one wheel gesture stopped short

**Evidence:** `Browser journeys (webkit, 6/14)` on #2937 failed
`history-prefix.spec.mjs:37` with `wheel reaches timeline edge`,
`Received: 199`
([job](https://github.com/block/buzz-app/actions/runs/42320781857/job/128244819852)).
The trace shows the one `-3251px` wheel ended with `scrollend` and the
geometry then held still, so the poll could never pass.

**Fix:** `edge()` repeats the gesture, at most four times, until the
edge is reached. Every extra gesture must reduce the distance, so an
immobile timeline still fails. `upper()` already follows this rule. The
case of already being at the edge is unchanged: it still sends one small
wheel so production can start paging.

## Not in this PR, from the same survey of the 40 most recently updated
open PRs

- `sidebar-geometry.spec.mjs:224` (Drafts & sent row) is broken on main
since #2733 + #2728 crossed. #2945 is green and #2946 is in CI; either
one fixes it.
- `solo-chats.spec.mjs:322` (Share, webkit, on #2927) is being handled
in draft #2933.
- `layout.spec.mjs:803` on #2916 is already fixed on main by #2750.
#2916's merge ref predates it.
- #2911 (thread virtualization) and #2918 (per-channel notifications)
fail consistently in tests their own changes reach, so those look like
regressions for their authors, not flakes.

## Validation

All runs were local on macOS with
`tests/browser/playwright.ci.config.mjs` and `--no-deps`. These are not
hosted CI measurements.

- `channel-completion.spec.mjs`, `--repeat-each=5`, both engines: 20/20
passed with the change. Before the CSS settings were added, the built
fixture failed WebKit 5/5 (`4.03125`).
- `profiles.spec.mjs:940`, `--repeat-each=24 --workers=6`, WebKit:
passed. It also passed with `history.go/back/forward` delayed by 300ms,
in both engines.
- `history-prefix`, `history-loading`, `scroll`, `timeline-setup` and
`new-message-boundary` (all callers of `edge()`/`end()`), both engines:
70/70 passed.
- Full `history-prefix`, `channel-completion` and `profiles` files, both
engines: 43/44 passed. The one failure was `profiles.spec.mjs:6` in
WebKit, a test this PR doesn't touch. It then passed 4/4 both with and
without this change.

**Deferred and remaining gaps:**
- I couldn't reproduce the cold-load timeout or the WebKit stopping
short locally, so I have no local fail-before for those two. The
evidence is the CI traces linked above.
- The other `edge()` callers (24 spec files) are left to CI.
- A green run doesn't prove these flakes are gone.

Signed-off-by: Bradley Axen <baxen@squareup.com>
Co-authored-by: mbot <3754f8729004d95654c46dbab3129e4ab9ef05cc2534e2a3fbfc155983bd637b@buzz.block.builderlab.xyz>

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

Reviewers

No reviews

Assignees

No one assigned

Labels

None yet

Projects

None yet

Milestone

No milestone

Development

Successfully merging this pull request may close these issues.

1 participant

@philazar