The Open Framework for Building Agents
Version Packages (#4589)
github-actions[bot]730048d18cd1
diff --git a/.changeset/approval-denial-note.md b/.changeset/approval-denial-note.mddeleted file mode 100644index 5f662d9be..000000000--- a/.changeset/approval-denial-note.md+++ /dev/null@@ -1,5 +0,0 @@-----"eve": patch------A denied tool approval that carries `text` now passes that note to the model in the denial reason instead of dropping it.diff --git a/.changeset/calm-functions-start.md b/.changeset/calm-functions-start.mddeleted file mode 100644index e49e609cb..000000000--- a/.changeset/calm-functions-start.md+++ /dev/null@@ -1,5 +0,0 @@-----"eve": patch------Use build-stamped package metadata in bundled runtimes so serverless functions can start without resolving `eve/package.json` at runtime.diff --git a/.changeset/chatgpt-current-models.md b/.changeset/chatgpt-current-models.mddeleted file mode 100644index d81bd3d19..000000000--- a/.changeset/chatgpt-current-models.md+++ /dev/null@@ -1,5 +0,0 @@-----"eve": patch------`/login chatgpt` now lists current subscription models such as `gpt-6-luna`, and the default `chatgpt()` model (`gpt-6-luna-fast`) works: eve sends it to Codex as `gpt-6-luna` on the priority service tier, since Codex rejects `-fast` model IDs.diff --git a/.changeset/openai-prompt-cache-key.md b/.changeset/openai-prompt-cache-key.mddeleted file mode 100644index cd6bb236d..000000000--- a/.changeset/openai-prompt-cache-key.md+++ /dev/null@@ -1,5 +0,0 @@-----"eve": patch------Direct OpenAI and `chatgpt()` model calls now send a per-session `promptCacheKey`, which helps OpenAI route a session's calls to the cache holding its prompt prefix. `chatgpt()` calls also send it as the `session-id` header the Codex backend routes on. An authored `providerOptions.openai.promptCacheKey` still takes precedence.diff --git a/.changeset/retire-incompatible-child-runs.md b/.changeset/retire-incompatible-child-runs.mddeleted file mode 100644index 939385efd..000000000--- a/.changeset/retire-incompatible-child-runs.md+++ /dev/null@@ -1,5 +0,0 @@-----"eve": patch------On worlds without per-deployment routing, incompatible subagent sessions and workflow tool runs (including tasks) are now cancelled when their queue deliveries arrive, instead of attempting replay on the new eve version. Top-level sessions remain available for replacement or reset, and session timeout workflows continue to run.diff --git a/.changeset/stranded-sessions.md b/.changeset/stranded-sessions.mddeleted file mode 100644index 34d8f7778..000000000--- a/.changeset/stranded-sessions.md+++ /dev/null@@ -1,5 +0,0 @@-----"eve": patch------Sessions whose workflow run cannot execute after an eve upgrade (a stranded session) are no longer accepted and left unanswered, and the local World and Postgres stop replaying and failing them with `CORRUPTED_EVENT_LOG`. eve treats every World other than Vercel as running one eve version at a time, so the next message through a channel address (Slack, `from(address).send()`) ends a stranded session and starts a fresh one. Sends by session ID throw `SessionStrandedError` (`ClientSessionStrandedError` in `eve/client`), and the eve HTTP channel returns `409 session_stranded`; `reset()` ends a stranded session while `clear()` refuses one, and recorded history stays readable with `stream({ follow: false })` or `follow=false` on the stream route. On Worlds other than Vercel, a retired session's own session timeout now ends it at its deadline, so a retired session no longer stays active forever when no message arrives.diff --git a/.changeset/web-search-compat-recovery.md b/.changeset/web-search-compat-recovery.mddeleted file mode 100644index ee349820f..000000000--- a/.changeset/web-search-compat-recovery.md+++ /dev/null@@ -1,5 +0,0 @@-----"eve": patch------Agents whose OpenAI model is served by an endpoint without OpenAI web search, such as Amazon Bedrock through `createOpenAI({ baseURL })` or AI Gateway, now retry the step without `web_search` instead of failing every turn.diff --git a/packages/eve/CHANGELOG.md b/packages/eve/CHANGELOG.mdindex 59e97a454..027b024f4 100644--- a/packages/eve/CHANGELOG.md+++ b/packages/eve/CHANGELOG.md@@ -1,5 +1,17 @@# eve+## 0.76.1++### Patch Changes++- fcbd479: A denied tool approval that carries `text` now passes that note to the model in the denial reason instead of dropping it.+- 9626748: Use build-stamped package metadata in bundled runtimes so serverless functions can start without resolving `eve/package.json` at runtime.+- 25b3f35: `/login chatgpt` now lists current subscription models such as `gpt-6-luna`, and the default `chatgpt()` model (`gpt-6-luna-fast`) works: eve sends it to Codex as `gpt-6-luna` on the priority service tier, since Codex rejects `-fast` model IDs.+- 416e63e: Direct OpenAI and `chatgpt()` model calls now send a per-session `promptCacheKey`, which helps OpenAI route a session's calls to the cache holding its prompt prefix. `chatgpt()` calls also send it as the `session-id` header the Codex backend routes on. An authored `providerOptions.openai.promptCacheKey` still takes precedence.+- b04acbb: On worlds without per-deployment routing, incompatible subagent sessions and workflow tool runs (including tasks) are now cancelled when their queue deliveries arrive, instead of attempting replay on the new eve version. Top-level sessions remain available for replacement or reset, and session timeout workflows continue to run.+- b04acbb: Sessions whose workflow run cannot execute after an eve upgrade (a stranded session) are no longer accepted and left unanswered, and the local World and Postgres stop replaying and failing them with `CORRUPTED_EVENT_LOG`. eve treats every World other than Vercel as running one eve version at a time, so the next message through a channel address (Slack, `from(address).send()`) ends a stranded session and starts a fresh one. Sends by session ID throw `SessionStrandedError` (`ClientSessionStrandedError` in `eve/client`), and the eve HTTP channel returns `409 session_stranded`; `reset()` ends a stranded session while `clear()` refuses one, and recorded history stays readable with `stream({ follow: false })` or `follow=false` on the stream route. On Worlds other than Vercel, a retired session's own session timeout now ends it at its deadline, so a retired session no longer stays active forever when no message arrives.+- cf72486: Agents whose OpenAI model is served by an endpoint without OpenAI web search, such as Amazon Bedrock through `createOpenAI({ baseURL })` or AI Gateway, now retry the step without `web_search` instead of failing every turn.+## 0.76.0### Minor Changesdiff --git a/packages/eve/package.json b/packages/eve/package.jsonindex 6ba67227f..27cc1fac1 100644--- a/packages/eve/package.json+++ b/packages/eve/package.json@@ -1,6 +1,6 @@{"name": "eve",- "version": "0.76.0",+ "version": "0.76.1","private": false,"description": "Filesystem-first framework for durable backend AI agents that run anywhere.","keywords": [