EggPool's provider-wire acceptance suite is deliberately opt-in. It uses a temporary SQLite database and configuration, calls EggPool's public HTTP endpoints, and records only sanitized outbound observations (provider, account alias, model, path/surface, status, auth scheme, semantic field names, streaming, attempt ordinal, and selection source). It never records keys, header values, raw bodies, or provider response bodies.
The current official OpenCode Go endpoint table lists these representative surfaces:
| Model ID | Upstream surface |
|---|---|
muse-spark-1.2-contributor |
OpenAI Responses (/responses) |
muse-spark-1.3-contributor |
OpenAI Responses (/responses) |
gpt-5.6-luna |
OpenAI Responses (/responses) |
minimax-m3 |
Anthropic Messages (/messages) |
mimo-v2.5 |
OpenAI Chat Completions (/chat/completions) |
Re-check the official Go endpoint table before a release verification because the model list and endpoint assignments can change.
Set the test-only credential variable in the shell that runs the live
qualification runner below. (The historical pytest suites
tests/live/test_opencode_go_wire_live.py and
tests/integration/test_wire_negotiation_e2e.py were retired with the
Python application tooling; wire negotiation is now covered by the native
rust/tests/ targets such as wire_qualification, wire_runtime, and
coordinator_boundaries. Do not treat those pytest paths as runnable.)
The suite covers non-streaming path selection and learned steady state,
Responses/Chat/Messages streaming terminal evidence, a public
Messages-to-Responses cross-surface request, surface-native Muse Spark 1.2/1.3
and MiMo reasoning shapes, a MiniMax-M3 Chat binary reasoning-toggle request
adapted to Anthropic Messages thinking, an invalid MiniMax effort request that
is rejected before dispatch, and invalid-key isolation. The invalid-key fixture uses
ordinary account weights to make the bad account the first candidate; it does
not add a production-only routing hook. It uses bounded, low-token requests;
it is not a load test or billing benchmark. Without the environment variable,
pytest skips it cleanly. It is excluded from the default suite, smoke tests,
and CI.
The optional second account variable, EGGPOOL_E2E_OPENCODE_GO_API_KEY_2, is
reserved for future multi-valid-account checks. Gemini live checks use
EGGPOOL_E2E_GEMINI_API_KEY when a direct Gemini live matrix is enabled; the
deterministic Gemini codec and path tests do not require credentials.
For release closure, record the exact live test outcomes. A clean skip caused by a missing credential is not live verification evidence.
The bounded runtime live-provider qualification runner exercises the Rust binary through its public HTTP server. It prints the frozen seven-request matrix before dispatch, uses an isolated database/configuration, records only semantic evidence, and never writes credentials or raw bodies:
uv run python scripts/qualification_live_provider.py \
--binary rust/target/release/eggpool \
--enable-live \
--provider-key-env EGGPOOL_E2E_OPENCODE_GO_API_KEY \
--output artifacts/qualification/007-run.jsonUse --env-file only with a local, uncommitted dotenv file when the selected
credential is not exported. The deterministic loopback rehearsal is
--offline-fake; it must not be described as live-provider evidence.
The extended live-provider qualification corrective profile is a separate, still opt-in seven-request matrix for two authorized provider edges. It uses GeneralCompute's OpenAI-compatible Chat surface and MiniMax International's Anthropic Messages surface, including Responses/Chat adaptation, streaming terminal evidence, and one Chat-to-Messages cell:
uv run python scripts/qualification_live_provider.py \
--binary rust/target/release/eggpool \
--enable-live --profile q011-multi \
--provider-key-env Q011_GENERALCOMPUTE_KEY \
--secondary-provider-key-env Q011_MINIMAX_KEY \
--output artifacts/qualification/011-run.jsonBoth credentials are supplied only through the invoking environment. The profile retains the live-provider qualification maximum request budget and writes only bounded, secret-free evidence.
The mandatory stale-profile check (historically
tests/integration/test_wire_negotiation_e2e.py, now covered natively)
uses an in-process fake upstream. It first accepts Responses, then changes to a safe Responses rejection with
Chat acceptance. The same account succeeds after an in-process alternate-wire
retry, the Chat preference is learned, and the next request uses Chat without
a restart or database reset. The test also verifies that the outbound hook
sees the actual paths and attempt ordinals without exposing credentials.
The same integration module covers unhinted known-model negotiation, strong
model-absence control, and Messages↔Responses/Chat cross-surface adaptation.