figmar/BrowserSkill
1
0
Fork 0
mirror of https://github.com/Tencent/BrowserSkill.git synced 2026-09-10 13:01:23 +08:00
Code Issues Packages Projects Releases Wiki Activity

feat(bsk): harden browser file transfer transactions

Browse Source
This commit is contained in:
Ljy-0827
2026-08-24 15:02:52 +08:00
parent a86cbf724f
commit bf59753aa3
24 changed files with 1397 additions and 407 deletions
Download Patch File Download Diff File Expand all files Collapse all files
docs/architecture.md
+5 -4
View File
@@ -146,10 +146,11 @@ flowchart LR
### File-transfer boundary
- The invoking agent/harness decides whether a transfer is authorized and supplies the task-local source or destination path.
- The CLI is the only component that reads an upload source or writes the final download destination. The extension never receives either agent-facing path.
- The daemon issues opaque, session-scoped transfer IDs and stages bounded chunks in a private runtime directory. For upload it injects only its private staged paths. For download it mints a relative Chrome directory capability, validates the reported completed path against that capability, rejects symlinks and oversized/non-regular files, then imports the bytes into private staging.
- The extension only performs the browser-side mechanism, scoped to the requested session and exact Agent Window target. Upload installs a transaction-scoped DOM guard that records the one `<input type="file">` actually activated by the requested click while cancelling its native default action, then uses `DOM.setFileInputFiles`. File System Access picker entry points are intercepted only for that transaction and return a structured fallback. Download requires an exact-target `Page.downloadWillBegin` intent before a matching `chrome.downloads` item is routed into the daemon-minted relative directory. The extension does not classify content or make policy decisions.
- Download staging is released after the CLI commits the file. Upload staging remains until session teardown because the page may read an attached file only on a later form submission. All staging is released on session stop/browser disconnect and on daemon startup after a crash. Existing download destinations are not overwritten unless the CLI caller explicitly opts in.
- The CLI is the only component that reads an upload source or writes the final download destination. Before browser dispatch it owns rollback of partially staged uploads; after dispatch, ownership moves to the session because a transport timeout cannot prove that Chrome did not attach the file. Download output becomes visible through one atomic commit, and replacement is opt-in without a pre-delete window. The extension never receives either agent-facing path.
- The daemon is the authority for storage capabilities and limits. It issues opaque session-scoped transfer IDs, stages bounded chunks in a private runtime directory, and injects only private staged upload paths. For download it mints one relative Chrome directory capability. Only after validating the reported path, file type, symlink boundary, and authoritative byte limit does it take ownership of browser-file cleanup and import the bytes.
- The extension owns only the browser transaction. Both tools resolve one `ResolvedActionTarget` and use that same target for protocol setup and the triggering click. Upload arms Chrome's chooser interception before clicking, then accepts either an exact `Page.fileChooserOpened` input node or an independent probe anchored in the trigger node's own document; one verified input is committed with `DOM.setFileInputFiles`. A non-input picker is rejected immediately as unsupported. Download correlates exact-target CDP intent and `chrome.downloads` filename candidates in either arrival order, claims only one unique match, and never cancels an unclaimed candidate.
- Browser-side operations report `effect_state` (`none`, `committed`, or `unknown`), `phase`, and `cleanup_state`. Confirmed success wins over a late cancel; an unknown effect is preserved across timeout or transport loss and must not be retried blindly. A transfer deadline sends cancellation to the extension and keeps the session queue occupied for bounded compensation rather than abandoning an in-flight browser effect.
- Download staging is released after CLI commit. Upload staging remains until session teardown because the page may read an attached file only on a later form submission. Remaining staging is released on session stop/browser disconnect and on daemon startup after a crash. BrowserSkill does not inspect content or decide whether a transfer is appropriate.
## Repository layout