MCP 服务 2026-10-09 18:43 UTC
rocketride-server
由提交人 dsapandora 提交,主要更新类型为:新增功能与能力扩展。提交说明:feat(log): list every run log of a source in one call (#2583)。带来了新的能力与可配置项,扩展了使用场景。
Commit提交信息原文
feat(log): list every run log of a source in one call (#2583) * feat(log): rrext_log streams names every continuum of a source the caller may read Where a run's events land depends on how the run was started: a dev run logs into the owner's tree, a team deploy (a scheduled run) into the team's tree, an @me deploy into the owner's tree again. rrext_log reads exactly one of those, and nothing in a stream says another exists, so an app reading chapters() without a teamId sees its dev stream and every scheduled run is silently absent (#2108). The new `streams` subcommand takes projectId + source and returns every continuum the caller may read, own scopes first, each tagged with teamId, teamName, runKind and owner and carrying the chapters body. Team scopes come from the session's org record filtered through resolve_task_permissions, the same math as the team gate, so nothing a direct teamId request would be denied on is listed. Never-logged scopes are absent, not empty. Both SDKs expose it as client.log.streams(); the docs name the keying. Nothing existing changes: the other subcommands keep their one-continuum contract and seq stays per stream. Part of #2108. * test(client-python): streams lists a dev run and a team run against a live server A dev run and a team deploy run of one source, then client.log.streams() names both while chapters() alone sees one. Runs with the other client-python integration tests against ROCKETRIDE_URI. Part of #2108. * fix(log): streams reads scopes with bounded concurrency; stream fields are required in the Python types An org admin can monitor every team, so the per-scope control-file reads run a few at a time instead of one after another. Order and omission of never-logged scopes are unchanged. LogStreamEntry and LogStreamsResult declare the fields the command always sends as required, as the TypeScript contract does. Part of #2108. * chore(shell): freeze the shell-api contract as v5 The shell re-exports every SDK type, so LogStreamEntry, LogStreamIdentity and LogStreamsResult are new shell-api exports. Generated by ./builder shell:freeze; v5 adds exactly those three names over v4. Part of #2108. * test(log): streams refuses without task.monitor; SDK docstrings carry Args and Returns The task.monitor gate in front of the caller's own streams had no test: the connection stub replaces verify_permission, so dropping the line kept every test green. The new test fails without it. The streams() docstrings in both SDKs now document parameters and return value, since the reference is generated from them. Part of #2108. * fix(store): a missing file raises StorageNotFoundError, so a failed read no longer reads as "never logged" Every store provider already told the two apart in the message; now it is a type. _load_control translates only StorageNotFoundError into FileNotFoundError, so chapters() and streams() surface a failed read instead of answering an empty timeline or dropping the scope. Still a StorageError, so nothing that catches that changes. Part of #2108. * test(store): pin StorageNotFoundError per provider; Azure move_file keeps other errors; writer never starts fresh on a failed read The provider tests asserted the base StorageError, which the subclass satisfies, so any of the 24 not-found sites could silently go back to the base type and streams() would then fail for nearly every caller. Every not-found test now asserts StorageNotFoundError, BaseStoreTest covers open_read (the path FileStore.read takes), and the mocked Azure store covers open_read and move_file. Reverting s3.py's open_read site to StorageError turns the suite red. Azure move_file only reports "not found" for BlobNotFound or ResourceNotFound and re-raises the rest, like delete_file and S3. RunLogWriter._load_control had the reader's old except Exception: a transient store error at run start began from an empty control and wrote it over the stored chapters. It now starts fresh only on StorageNotFoundError (or an unreadable control) and lets a failed read propagate, with a test. The cmd_log module header lists streams. Part of #2108.