Own-plan turns get 165k tokens, 200 s and 16 rounds, operator turns 55k,
120 s and 8; the final-answer reserve, the concurrent-turn slot and the
step cap move with them.
Codex requests carry a per-conversation prompt_cache_key and a stable
session id, api.openai.com gets the same key, Anthropic usage counts
cache reads and writes in the gross, and local servers report cached
prompt tokens, so the turn budget discounts cached input on every
provider.
A gateway can answer a failed generation with HTTP 200 and put the reason
in the body: a top-level `error` object and no `choices`, or
`choices[0].error` beside `finish_reason: "error"` when a generation dies
mid-flight. Both fell through to the empty-content throw, so a dead credit
balance, a data-policy block, a rate limit and a model that is down all
recorded the same hop: status 0, reason "returned empty content", no
excerpt. Nothing on the provider health card told them apart.
Raise the shape a non-2xx reply raises instead, with the gateway's own
message leading and its `error.code` as the status when that code is an
HTTP error status. The chain classifier, the health ledger and the card
then see 402 / 403 / 429 / 503 and treat each of them the way they treat
the same status on the wire. Message and excerpt are bounded and
key-redacted by the same helper the non-2xx path uses, so remote text is
no more trusted here than it is there.
Gated on the gateway tag: api.openai.com signals every failure with a
status, so the pinned tags parse exactly as before, and a genuine empty
reply stays the empty-content throw it was. The local-endpoint provider
had the same hole on its buffered path, its buffered-reply fallback and
its stream, and gets the same treatment. The body classifier learned the
two account-side refusals a gateway words in its own vocabulary rather
than the vendor's.
Replace the CompletionParams { systemPrompt, userPrompt } string pair with a
structured { system, messages[] } request across all five AI clients. One
refactor unlocks three things: provider prompt-caching (Anthropic cache_control
on the stable system prefix, OpenAI/Codex automatic prefix cache, with
cachedInputTokens surfaced for observability), the tool-calling wire surface,
and clean multi-turn.
Each client maps the new shape to its own wire and degrades unsupported
features silently: the local OpenAI-compatible server never receives a tools
field (most reject unknown keys), and prompt-caching is best-effort everywhere.
CompletionResult gains toolCalls, finishReason and cachedInputTokens; the tool
types and per-provider mapping ship now, the tool-use loop does not.
Vision inputs fold into the user turn's content parts via singleUserTurn, with
each provider's pre-existing block ordering preserved. The corrective-retry
paths append to the last user message rather than a flat string. The Coach
keeps assembling one user turn so its snapshot-once and grounding behaviour is
unchanged. No wire schema visible to API consumers changed.
Behaviour is unchanged across every surface; the full suite stays green.
Marc-André Bombeck