Skip to content

compatibility.md

Latest commit

 

History

History
87 lines (65 loc) · 5.72 KB

compatibility.md

File metadata and controls

87 lines (65 loc) · 5.72 KB
Raw

Compatibility matrix

What has actually been run against version 1.0.0 of the plugin, and how. "Pass" means the named test passed end to end. Everything else is marked as not run, with a reason.

All integration tests use synthetic credentials and local TLS fixtures that stand in for api.anthropic.com and chatgpt.com. Networking is disabled where the platform allows. No test contacts a real provider or spends real quota.

CLIProxyAPI host

CPA release Platform Test Result
v8.0.27 (official release binary) linux/arm64 make host-test Pass
v8.0.27 (official release binary) linux/amd64 (emulated on an arm64 host) make host-test Pass
v8.0.27 (official Docker image eceasy/cli-proxy-api:v8.0.27) linux/arm64 tests/docker_persistence.py Pass
v8.0.27 (official Docker image) linux/amd64 tests/docker_persistence.py Runs in CI on ubuntu-24.04; not yet run for this branch
v8.0.27 (official release binary) darwin/arm64 make load-test Pass
v8.0.27 darwin/amd64, windows/amd64 make load-test Runs in CI; not yet run for this branch
v8.0.27 darwin/arm64, darwin/amd64 make host-test on macOS Runs in CI; not yet run for this branch
v7.3.15 (official release binary) linux/arm64 make host-test CPA_VERSION=7.3.15 Pass, Claude only (see below)
v7.3.15 linux/amd64 make host-test CPA_VERSION=7.3.15 Not run
CPA Home dispatch (home.enabled) — — Not supported: the host bypasses plugin schedulers

host-test scenario steps (official binary, plugin loaded from plugins.dir):

  • Claude and Codex accounts are discovered and their quota is polled.
  • Shadow mode, then active mode.
  • A session stays on its account when the ranking changes; a new session goes to the new first-ranked account.
  • A 429 with exhausted-window headers rebinds the session.
  • A 503 fails over for one request and keeps the binding; the account is used again after recovery.
  • A Claude Code subagent inherits its parent session's account.
  • Bindings survive a CPA restart.
  • Streaming requests work.
  • Settings are changed through CPA's PATCH …/config and the plugin reloads them.
  • Dashboard assets are served, management routes require authentication, and no secrets appear in any response.
  • Disabling the plugin returns routing to CPA.

docker_persistence.py covers routing and affinity, docker restart, and container removal and recreation with the state directory in the plugins bind mount. It also checks that the state file contains no secrets.

Codex on v7.3.15: CPA v7.3.15 talks to chatgpt.com over HTTP/2. The test fixture is a Python HTTP/1.1 server, so the Codex steps are skipped on v7 (http2: frame too large). This is a limitation of the fixture, not an observed plugin failure. The plugin's Codex code path is the same on v7 and v8 (same ABI, same schema). Treat Codex on v7 as untested.

Plugin ABI

Check Platform Result
make native-test: loads the release .so through CPA's exact C ABI with ctypes. Covers registration, capabilities, TLS quota reads, shadow/active routing, reset rollover, exhaustion, manual refresh, 500 picks, quiesce and resume, and buffer ownership. linux/arm64 Pass
same linux/amd64 Runs in CI; not yet run for this branch
Plugin ABI 1, schema version 6 v7.3.15 and v8.0.27 SDKs Identical; one binary loads in both

Builds

Target Built Notes
linux_amd64 yes (file: ELF x86-64) pinned golang:1.26-bookworm image
linux_arm64 yes (file: ELF aarch64) pinned image
darwin_arm64 yes host Go 1.26.8
darwin_amd64 CI patched Go toolchain (TSD slot 11), checked by otool
windows_amd64 CI MinGW-w64 10.0.0-3 in the pinned image

The dashboard bundle (web/ → internal/mgmt/dist) builds reproducibly: rebuilding from the lockfile produces byte-identical assets, and CI fails if the committed bundle is stale.

Unit and component tests

Suite Result
go test -race ./... Pass. Covers session extraction parity with CPA's ExtractSessionInfo, quota parsers and assessment, router policies, affinity, failover and rebind, persistence (atomic writes, corruption quarantine, lock), telemetry, management routes, a no-secrets check across every route and the state file, and concurrency.
cd web && pnpm test (Vitest) Pass: 30 tests (formatting, durations, reasons, quota meter)

Clients

Session identity follows CPA's header precedence, and a unit test compares it with CPA's own extractor.

Client Headers used Tested with
Claude Code X-Claude-Code-Session-Id, X-Claude-Code-Agent-Id, X-Claude-Code-Parent-Agent-Id synthetic requests in host-test, unit tests
Codex CLI Session-Id, Thread-Id, X-Codex-Turn-Metadata, X-Codex-Parent-Thread-Id, X-Openai-Subagent synthetic requests in host-test, unit tests
Generic clients X-Session-ID, X-Conversation-Id, X-Thread-Id, … unit tests

Real Claude Code and Codex CLI binaries were not run against the proxy in this work.

Dashboard

Browser Viewports Themes Result
Chromium (Playwright 1.64) 1440×960, 390×844 (mobile emulation) dark, light all six pages render with no console errors (web/scripts/screenshots.mjs)
Chromium (Playwright 1.64) + axe-core 1440×960, 390×844 dark, light no WCAG 2.1 A/AA violations on the login screen or any page (pnpm a11y)

All rows and controls are reachable by keyboard and show a visible focus ring. Scrollable tables are focusable regions. Screen-reader labels are set for icon-only controls.

Other browsers were not tested. The bundle targets current evergreen browsers and needs no polyfills.