@vincemakes/kiso-provider-openai-responses 0.43.0 → 0.45.0

package/README.md

@@ -6,7 +6,7 @@ new dependency (plain `fetch` plus a hand-rolled SSE reader).
 | target | auth | endpoint | what is different |
 |---|---|---|---|
 | OpenAI first-party | API key (`kiso login openai`, or `OPENAI_API_KEY`) | `https://api.openai.com/v1/responses` | the public Responses API |
-| ChatGPT subscription | the stored OAuth credential (`kiso login chatgpt`) | `https://chatgpt.com/backend-api/codex/responses` | `chatgpt-account-id` / `originator` / `OpenAI-Beta` headers; `store: false`, `include: ["reasoning.encrypted_content"]`, `prompt_cache_key` |
+| ChatGPT subscription | the stored OAuth credential (`kiso login chatgpt`) | `https://chatgpt.com/backend-api/codex/responses` | `chatgpt-account-id` / `originator` / `OpenAI-Beta` headers, and `session-id` (the session's cache lane, the same value as `prompt_cache_key`); `store: false`, `include: ["reasoning.encrypted_content"]`, `prompt_cache_key` |
 
 The target is inferred from the options the factory is handed — an
 `apiKey` builds the first-party adapter, an `oauth` thunk builds the
@@ -27,7 +27,8 @@ LOCAL double (`tests/helpers/rig.ts`); no test reaches a vendor.
 | streaming | ✓ | ✓ | `or1-stream` — the nine adapter events in order, `usage` before `stop`, both from the same terminal frame |
 | tool call + next turn | ✓ | ✓ | `or1-tool-turn` — the emitted `callId` is the `call_id`, and the next request replays it as `function_call_output` |
 | reasoning effort | ✓ | ✓ | `or1-request-rig` shape 3 — `reasoning.effort` carries the level it was handed |
-| reasoning replay (`store: false`) | n/a | ✓ | `or1-continuation` — one `stop.continuation` entry per encrypted reasoning item, replayed on scope match only |
+| reasoning summary | ✓ | ✓ | `or1-request-rig` shapes 3 and 4b — with an effort, `reasoning.summary: "auto"`; the summary streams as `thinking` events (`or1-stream`); no effort, no `reasoning` key |
+| reasoning replay (`store: false`) | n/a | ✓ | `or1-continuation` — one `stop.continuation` entry per encrypted reasoning item, replayed verbatim (its summary included) on scope match only |
 | cancel | ✓ | ✓ | `or1-errors` — an abort ends the turn with an `AbortError`, no `stop`, nothing retryable |
 | error mapping | ✓ | ✓ | `or1-errors` — 429/401/400/503 by status, both `Retry-After` forms in milliseconds |
 | retry authority | ✓ | ✓ | `or1-retry-authority` — exactly one request per stream, on every failure class |

package/dist/index.js

@@ -71,6 +71,12 @@ function resolveTarget(config) {
                     "chatgpt-account-id": token.accountId,
                     originator: "kiso",
                     "OpenAI-Beta": "responses=experimental",
+                    // The backend routes a session's requests to one cache by
+                    // this header. With the body's prompt_cache_key alone most
+                    // requests landed elsewhere and missed the cache (measured:
+                    // about 14% of input cached, 54% with the header). Same
+                    // value as the key, so a conversation is one lane.
+                    ...(config.promptCacheKey !== undefined ? { "session-id": config.promptCacheKey } : {}),
                 };
             },
             // The backend rejects `store: true`; with nothing stored, the
@@ -122,7 +128,12 @@ function buildBody(options, target, scopeProviderId) {
         // dialect parameter for a human reader and never reaches an
         // adapter — the dialect path is this adapter's own knowledge, and
         // the level is transported as handed. Absent adds NO key.
-        ...(options.reasoning?.effort !== undefined ? { reasoning: { effort: options.reasoning.effort } } : {}),
+        // With an effort, the reasoning summary is asked for too: the API
+        // streams summaries only on request (they arrive as `thinking`
+        // events below), and "auto" takes the most detailed one the model
+        // offers. No effort still means no key: whether a model reasons is
+        // not this adapter's call.
+        ...(options.reasoning?.effort !== undefined ? { reasoning: { effort: options.reasoning.effort, summary: "auto" } } : {}),
         ...(options.maxTokens !== undefined ? { max_output_tokens: options.maxTokens } : {}),
         ...(options.temperature !== undefined ? { temperature: options.temperature } : {}),
         ...target.extraBody,

package/package.json

@@ -1,6 +1,6 @@
 {
 	"name": "@vincemakes/kiso-provider-openai-responses",
-	"version": "0.43.0",
+	"version": "0.45.0",
 	"description": "kiso OpenAI Responses adapter \u2014 the first-party Responses API and the ChatGPT subscription backend, one adapter, plain fetch + SSE.",
 	"type": "module",
 	"license": "MIT",
@@ -21,7 +21,7 @@
 		"test": "vitest run"
 	},
 	"dependencies": {
-		"@vincemakes/kiso-core": "0.43.0"
+		"@vincemakes/kiso-core": "0.45.0"
 	},
 	"devDependencies": {
 		"@types/node": "^26.1.2",