Add native OpenAI spine CLI

This commit is contained in:
Meta Vens 2026-05-13 05:28:13 +01:00
commit 71eca3889e
7 changed files with 198 additions and 2 deletions

View file

@ -141,5 +141,20 @@ contracts in its hello/schema catalog. This is the contract bridge the native
runtime should consume next. Do not add another JSON-RPC model endpoint and call
it progress.
The first native operator/debug edge now exists as the
`epiphany-openai-spine` binary in `epiphany-openai-codex-spine`. It can report
typed adapter status and consume a serialized `EpiphanyOpenAiModelRequest`
document for a model turn without going through Codex app-server JSON-RPC. This
is not the final CultNet daemon. It is a buildable extracted edge that proves
the spine can be called from Epiphany-owned code and gives the next cut a place
to route through while the old `thread/epiphany/*` model path is starved.
The bad news is useful: the current spine still drags in `codex-api` and its
large transport dependency stack. That is tolerable as a sealed compatibility
reliquary, but not as the final shape. The next purification should shrink the
surviving Codex dependency surface to credential loading/refresh, provider
auth, model selection, and the smallest OpenAI Responses HTTP call that keeps
Codex subscription compatibility.
The point is ownership: Epiphany calls a model adapter; it does not live inside
the Codex host brain.

View file

@ -567,6 +567,16 @@ plus the coordinator-owned model request contract. The native runtime should
consume that contract next; do not make a new JSON-RPC model endpoint and
pretend the whale got lighter.
The `epiphany-openai-spine` binary is now the first native edge for that
wrapper. It can print typed adapter status and consume a serialized
`EpiphanyOpenAiModelRequest` document for a model turn outside Codex
app-server. Treat its JSON as CLI/file-edge serialization of typed documents,
not internal data cargo. The current wrapper still pulls `codex-api` and a
large dependency stack, so the next real cut is not celebration; it is routing
native runtime/CultNet model calls through this spine and then shrinking the
surviving Codex dependency surface to auth, provider/model routing, and the
smallest OpenAI Responses call needed for subscription compatibility.
Also: MCP itself is allowed to be JSON. The target is not "replace MCP JSON";
the target is an Epiphany-owned boundary that speaks typed Epiphany
intent/result/receipt documents internally and normal MCP JSON-RPC externally.