Why

A contestant cell in the harness-engineering-bench study, gpt-5.6-sol × codex, failed
on terminal-bench and nowhere else. It ships real candidates on officeqa r1/r2/r3 and
browsecomp-plus r1/r2/r3, so it is not a routing problem with the contestant.

The fault is invalid_encrypted_content, and it is deterministic rather than
load-dependent. An unqualified model group is load-balanced across upstream
deployments, and the Responses API's encrypted reasoning content is decryptable only by
the deployment that produced it. Replaying it against a sibling deployment fails on
every turn after the first.

Reproduced against the upstream proxy directly, with no gateway and no codex in the
path:

Asking for the pinned deployment fixes it — but codex cannot express that request.
It reduces a model id to its last path component before putting it on the wire, so
azure_ai/gpt-5.6-sol arrives as gpt-5.6-sol and we are back where we started. There
is no client-side fix available for this harness.

What this does

Adds model_aliases to an inference-gateway scope: a declared map from the name a
caller may request to the upstream deployment that serves it.

The ordering is the whole design, and it is deliberate:

upstream_model = scope.model_aliases.get(model, model)

This runs after the allow-list check, not before. allowed_models continues to
govern what the optimizer may ask for — and therefore what a cell's label means —
while model_aliases decides only which deployment answers. Rewriting before the
check would force the aliased name into allowed_models, and the allow-list would stop
describing the contestant.

Both names are recorded in the request log via a new aliased_from field, so a
substitution the caller cannot see is still visible to whoever reads the log later.

refused_params is keyed on upstream_model for both its read and its write, since
refusing a parameter is a property of the deployment rather than of the name used for
it. Keying the write on the requested name would never hit on the read path for an
aliased model, and the cache would re-discover the same refusal on every request.

Where it is configured

In terminal-bench/baseline/build.azure.yaml, in the producer scope only. The
evaluation and finalization scopes are untouched — they target grok, which was never
affected.

That file is a copy of build.yaml rather than an overlay, and it differs from it by
exactly the two alias lines. Edits to build.yaml must be mirrored, which is not
hypothetical: the first version of this change missed the finalization budget and
needed a follow-up commit to bring it back into line.

Also included, from the same investigation: terminal-bench's finalization budget is
sized like the rest of the suite.

Verification

• Full suite green: 496 passed, 8 skipped.
• New tests cover the alias reaching the gateway through model_dump, self-aliases
  being dropped rather than rejected, empty targets being refused, and — the case that
  matters for a shared grid config — alias keys that a given launch can never request,
  since allowed_models is templated per cell while the alias map is not.
• Live: two gpt-5.6-sol × codex terminal-bench cells have been running on this config
  for 25 and 18 minutes with zero invalid_encrypted_content and zero 403s, where
  the old failures crashed within minutes.

Note on scope

This is a workaround at the right layer, not the root fix. The upstream remedy is
encrypted_content_affinity in the proxy's router_settings, exactly as the error text
prescribes. This change lets the study close its last gap without waiting on that.

🤖 Generated with Claude Code

Greptile Summary

Adds per-scope model aliases to pin caller-visible model names to specific upstream deployments.

• Applies aliases only after the requested model passes the scope allow-list.
• Preserves aliases while compiling build specifications into gateway configuration.
• Records requested and upstream model names and keys parameter-refusal caching by upstream deployment.
• Adds an Azure-specific Terminal-Bench configuration and aligns its finalization budget with the shared configuration.
• Adds validation and tests for alias serialization, routing, logging, and configuration parity.

%%{init: {'theme': 'neutral'}}%%
flowchart LR
    A[Caller requests model] --> B{Allowed by scope?}
    B -- No --> C[Return model_denied]
    B -- Yes --> D[Resolve model_aliases]
    D --> E[Apply cached refused parameters]
    E --> F[Send to pinned upstream deployment]
    F --> G[Record upstream model and aliased_from]

_{Reviews (1): Last reviewed commit: "Ignore Codex CLI local state" | Re-trigger Greptile}