fix: harden session query authorization
This commit is contained in:
@@ -11,7 +11,9 @@ Workspace-authorized model tools over `ctx.sessionQuery`. The package depends on
|
||||
|
||||
The caller comes exclusively from `ToolExecution.exec.agent`. Cross-session access requires exact equality between the target and caller session `cwd` values; a caller without `cwd` can inspect only itself. Search never exposes provider cursors, offsets, page sizes, or a model-controlled limit. Because one search consumes generation-bound provider cursors internally, both search tools execute exclusively with sibling tool calls; the three exact trace/read tools opt into parallel execution. Every exact executor passes its unchanged execution signal through authorization and the service trace/read, so cancellation waits for cooperative persistence cleanup and retains the signal's exact reason. Timestamps at the tool boundary require an explicit `Z` or numeric offset and become inclusive epoch-millisecond filters.
|
||||
|
||||
`session_search` always omits the caller session. A current-session `session_event_search` stops immediately before the step that invoked it, so the active assistant output and logged tool call cannot match themselves. Direct targets are authorized before trace, event, or title reads. Lineage output replaces unauthorized ancestor and descendant boundaries with markers that contain no hidden session id.
|
||||
`session_search` always omits the caller session. Requested parent ids are deduplicated and checked against caller-workspace authority before FTS; only authorized ids reach the provider, while missing and cross-workspace guesses behave identically and the root marker remains independently ORed. A current-session `session_event_search` stops immediately before the step that invoked it, so the active assistant output and logged tool call cannot match themselves. Direct targets are authorized before trace, event, or title reads. Lineage output replaces unauthorized ancestor and descendant boundaries with markers that contain no hidden session id.
|
||||
|
||||
Every trusted `ctx.sessionQuery` call crosses one model-boundary sanitizer. Caller cancellation is checked first and preserved exactly. Available corpus and provider diagnostics, including safely inspectable nested causes, are logged internally on a best-effort basis; unprintable failures use a fixed log placeholder. Diagnostic formatting and error classification are independently guarded, so an unprintable cause cannot escape or prevent a safely classified outer error, while unsafe classification or logging falls back to the fixed `SESSION_QUERY_TOOL_FAILED` code and message. Local argument-validation and authorization errors retain their precise tool-owned messages.
|
||||
|
||||
The package deliberately performs no byte or character truncation and does not import a spill backend. Deployments that need bounded inline output mount `@deepseek-ai/dsh-spill-policy`, which can replace the rendered text after execution while retaining the complete result.
|
||||
|
||||
|
||||
Reference in New Issue
Block a user