diff --git a/apps/desktop/tests/acl-skill.built.e2e.ts b/apps/desktop/tests/acl-skill.built.e2e.ts index b0d84d3a41..c753f7ab8d 100644 --- a/apps/desktop/tests/acl-skill.built.e2e.ts +++ b/apps/desktop/tests/acl-skill.built.e2e.ts @@ -55,7 +55,7 @@ try { const directory = skill.resourceBase.path; assert.ok(!directory.includes('.asar')); const output = execFileSync('pwsh', ['-NoProfile', '-NonInteractive', '-File', - join(directory, 'scripts/diagnose-windows-sandbox-acl.ps1'), '-Path', root], + join(directory, 'scripts/diagnose-windows-sandbox-acl.ps1'), '-Path', root, '-AllowRoot', root, '-Out', join(root, 'out')], { encoding: 'utf8', timeout: 30000, windowsHide: true }); assert.ok(output.includes('READABLE=True')); await fiber.dispose(); diff --git a/packages/sandbox/sandbox-windows-acl/assets/diagnose-windows-sandbox-acl/SKILL.md b/packages/sandbox/sandbox-windows-acl/assets/diagnose-windows-sandbox-acl/SKILL.md index 014c0398e9..c0d9e67312 100644 --- a/packages/sandbox/sandbox-windows-acl/assets/diagnose-windows-sandbox-acl/SKILL.md +++ b/packages/sandbox/sandbox-windows-acl/assets/diagnose-windows-sandbox-acl/SKILL.md @@ -26,7 +26,7 @@ Substitute full, quoted paths. Keep `-Out` persistent and user-owned, preferably ## Run it -Repeat the failing operation once confined, then the command above confined too. When sandbox setup itself fails — the error names the workspace root or `SetNamedSecurityInfoW` — every confined call fails before the command runs, so do not retry confined: request approval for that one command and run it unconfined. Unconfined does not elevate the Windows token. +Repeat the failing operation once confined. The script writes permissions, which the confined token cannot do — run it there and it would report the sandbox's own restriction as a missing right — so request approval for that one command and run it unconfined. A confined run of the script is never useful. Unconfined does not elevate the Windows token. Diagnose unexpected denials of workspace writes, listing, or plainly readable paths; explain expected ones instead: writes outside the workspace, any write in `read-only`, piped grandchild `spawn EPERM`, ConstrainedLanguage errors. If approval is refused or unavailable, report the path as undiagnosed and stop. diff --git a/packages/sandbox/sandbox-windows-acl/assets/diagnose-windows-sandbox-acl/scripts/diagnose-windows-sandbox-acl.ps1 b/packages/sandbox/sandbox-windows-acl/assets/diagnose-windows-sandbox-acl/scripts/diagnose-windows-sandbox-acl.ps1 index fc6c90b9b3..bead02a280 100644 --- a/packages/sandbox/sandbox-windows-acl/assets/diagnose-windows-sandbox-acl/scripts/diagnose-windows-sandbox-acl.ps1 +++ b/packages/sandbox/sandbox-windows-acl/assets/diagnose-windows-sandbox-acl/scripts/diagnose-windows-sandbox-acl.ps1 @@ -281,8 +281,12 @@ function Get-SubtreePackageSources { $truncated = $false $queue = [System.Collections.Generic.Queue[string]]::new() # Unreadable directories are skipped, never guessed at: the caller reports the count. - try { foreach ($child in [System.IO.Directory]::EnumerateDirectories($Root)) { $queue.Enqueue($child); $enqueued++ } } - catch [System.UnauthorizedAccessException] { $failed++ } + try { + foreach ($child in [System.IO.Directory]::EnumerateDirectories($Root)) { + if ($enqueued -ge $Limit) { $truncated = $true; break } + $queue.Enqueue($child); $enqueued++ + } + } catch [System.UnauthorizedAccessException] { $failed++ } catch [System.IO.IOException] { $failed++ } while ($queue.Count -gt 0) { if ($visited -ge $Limit) { $truncated = $true; break } @@ -290,7 +294,7 @@ function Get-SubtreePackageSources { $visited++ if (Test-DangerousRoot -FullPath $current) { continue } $item = Get-Item -LiteralPath $current -Force -ErrorAction SilentlyContinue - if ($null -eq $item) { continue } + if ($null -eq $item) { $failed++; continue } if (($item.Attributes -band [System.IO.FileAttributes]::ReparsePoint) -ne 0) { continue } try { $acl = Get-Acl -LiteralPath $current @@ -611,8 +615,8 @@ try { if ($scan.truncated) { $script:scanTruncated = $true } } - # Nothing is changed while a package source is out of reach: a repair that cannot - # finish would leave the tree half repaired and only then report the refusal. + # Nothing is changed while a package source or a grant target is out of reach: a + # repair that cannot finish would leave the tree half repaired before refusing. foreach ($packagePath in $packageTargets) { $refusal = Get-RepairRefusal -FullPath $packagePath -Root $AllowRoot -AllowEqual if ($refusal) { Write-Line ('REPAIR_REFUSED {0}' -f $refusal); Write-Decision $mode $packagePath refused $refusal; $refused++; break paths } @@ -687,7 +691,7 @@ try { } } if ($sources.Count -gt 0) { - Write-Report decision remove_package_sources $full selected 'Remove explicit package allow ACEs from their sources, ancestor first. Inherited entries are verified after the source changes; inheritance stays enabled.' @{ + Write-Report decision remove_package_sources $full selected 'Remove explicit package allow ACEs from their sources, ancestor first, leaving inheritance enabled so inherited copies follow their repaired source.' @{ sources = @($sources); affectedPaths = @($packageTargets) } } diff --git a/packages/sandbox/sandbox-windows-acl/tests/diagnose-script.spec.ts b/packages/sandbox/sandbox-windows-acl/tests/diagnose-script.spec.ts index 4ccf770fe0..1240d7b078 100644 --- a/packages/sandbox/sandbox-windows-acl/tests/diagnose-script.spec.ts +++ b/packages/sandbox/sandbox-windows-acl/tests/diagnose-script.spec.ts @@ -368,6 +368,26 @@ describe.skipIf(!isWin32 || !pwshAvailable())('diagnose-windows-sandbox-acl scri } }, timeout) + it('keeps an inherit-only ACE when it grants the missing right', () => { + const scratch = newScratch() + try { + const target = makeDir(scratch, 'inherit-only') + icacls(target, '/inheritance:r', '/grant:r', '*S-1-5-11:(M)') + icacls(target, '/grant', '*S-1-5-32-545:(OI)(CI)(IO)(F)') + const inheritOnlyBefore = aclLines(target).filter(line => line.includes('(IO)')) + expect(inheritOnlyBefore).toHaveLength(1) + + const run = runScript(['-Path', target, '-AllowRoot', scratch, '-Out', join(scratch, 'out')]) + expect(run.code, run.output).toBe(0) + expect(run.output).toContain(`GRANTED ${target} SID=${meSid}`) + expect(reports(run)).toContainEqual(containingObject({ kind: 'verification', operation: 'grant', path: target, status: 'verified' })) + expect(aclLines(target).filter(line => line.includes('(IO)'))).toEqual(inheritOnlyBefore) + expect(aclLines(target).join('\n')).toMatch(/\(F\)/u) + } finally { + dispose(scratch) + } + }, timeout) + it('preserves a deny that shares the removed package allow SID', () => { const scratch = newScratch() try { diff --git a/snapshots/session/windows-acl-skill/session.v4.jsonl b/snapshots/session/windows-acl-skill/session.v4.jsonl index 53947f3a37..a7f2e5740a 100644 --- a/snapshots/session/windows-acl-skill/session.v4.jsonl +++ b/snapshots/session/windows-acl-skill/session.v4.jsonl @@ -15,7 +15,7 @@ {"type":"session/title","data":{"title":"Load the diagnose-windows-sandbox-acl sk","messageSeqs":[8],"source":{"kind":"fallback"}}} {"type":"assistant/message","data":{"turn":1,"step":1,"message":{"role":"assistant","content":[{"type":"reasoning","text":"Load the requested skill."},{"type":"tool-call","id":"call_skill_load","name":"skill","arguments":"{\"name\":\"diagnose-windows-sandbox-acl\"}"}],"source":{"kind":"model","provider":"deepseek-official","model":"deepseek-v4-flash"},"id":"{{message:5}}"},"usage":{"inputTokens":100,"outputTokens":20,"cacheReadTokens":0,"reasoningTokens":5},"stream":[{"type":"chunk","time":1788699430688,"chunk":{"type":"block-start","index":0,"blockType":"reasoning"}},{"type":"reasoning-chunks","time0":1788699430688,"index":0,"dt":[],"texts":["Load the requested skill."]},{"type":"chunk","time":1788699430688,"chunk":{"type":"block-start","index":1,"blockType":"tool-call"}},{"type":"tool-call-chunks","time0":1788699430688,"index":1,"dt":[],"id":"call_skill_load","name":"skill","args":["{\"name\":\"diagnose-windows-sandbox-acl\"}"]},{"type":"chunk","time":1788699430688,"chunk":{"type":"block-end","index":0,"block":{"type":"reasoning","text":"Load the requested skill."}}},{"type":"chunk","time":1788699430688,"chunk":{"type":"block-end","index":1,"block":{"type":"tool-call","id":"call_skill_load","name":"skill","arguments":"{\"name\":\"diagnose-windows-sandbox-acl\"}"}}},{"type":"chunk","time":1788699430688,"chunk":{"type":"usage","usage":{"inputTokens":100,"outputTokens":20,"cacheReadTokens":0,"reasoningTokens":5}}},{"type":"chunk","time":1788699430688,"chunk":{"type":"finish","reason":{"kind":"tool-calls"}}}]},"surfaceOp":"append"} {"type":"tool/call","data":{"turn":1,"step":1,"callId":"call_skill_load","name":"skill","arguments":"{\"name\":\"diagnose-windows-sandbox-acl\"}"}} -{"type":"tool/result","data":{"turn":1,"step":1,"message":{"role":"tool","source":{"kind":"tool","callId":"call_skill_load"},"toolCallId":"call_skill_load","content":[{"type":"text","text":"\n\nBase directory for this skill: {{cwd}}/.dsh/skills/diagnose-windows-sandbox-acl\nResolve relative paths mentioned by this skill against the base directory before using them. Load referenced resources only as needed.\n\n\n\n# Diagnose Windows sandbox ACL failures\n\n**Write every approval request in plain words, in the user's language.** The prompt is all the user reads before widening access, so it must stand alone: which folder the script touches, that it adds the signed-in user's full-control entry where a right is missing and removes foreign package entries, that file contents and owners are unchanged, and that the printed recovery command undoes each change. Keep error codes, `WRITE_DAC`/`WRITE_OWNER`, `S-1-15-2-*` SIDs, `icacls`, ACL, ACE, verdict names and switches out of the request: write Windows file permissions, not ACL or ACE. Ask once, for one command.\n\n## One command diagnoses and repairs\n\nThe script has no modes. `-Path`, `-AllowRoot` and `-Out` read the path and every ancestor and repair what the observations prove, in the same run:\n\n- directories on that chain that lack effective `WRITE_DAC` or `WRITE_OWNER` receive a full-control allow ACE for the signed-in user, because DSH cannot provision its workspace grant without them;\n- explicit AppContainer package allow ACEs (`S-1-15-2-*`, except the well-known groups ending in 1 or 2) are removed at their sources, ancestor first, which also removes those packages' access;\n- a directory the sandbox cannot provision — the state its provisioning error reports on the workspace root — or the authorized root itself also has its subtree searched for those ACEs in the same run, so a deeper entry needs no second request. That bounded walk reports `truncated` and unreadable directories; if truncated, pass the still-failing deeper path once more.\n\nEvery change is backed up, then verified by re-reading it. `-AllowRoot` bounds all of it: an object is changed only when it is that directory or strictly inside it, so a workspace root can repair itself. Never split this into a diagnostic call and a repair call, never ask twice for one repair, and never pass a mode switch that does not exist.\n\n```powershell\n& '\\scripts\\diagnose-windows-sandbox-acl.ps1' -Path '' -AllowRoot '' -Out ''\nexit $LASTEXITCODE\n```\n\nSubstitute full, quoted paths. Keep `-Out` persistent and user-owned, preferably beside the failing workspace; never use the skill resource directory, which is deleted when the skill unloads. Pass one path per invocation.\n\n## Run it\n\nRepeat the failing operation once confined, then the command above confined too. When sandbox setup itself fails — the error names the workspace root or `SetNamedSecurityInfoW` — every confined call fails before the command runs, so do not retry confined: request approval for that one command and run it unconfined. Unconfined does not elevate the Windows token.\n\nDiagnose unexpected denials of workspace writes, listing, or plainly readable paths; explain expected ones instead: writes outside the workspace, any write in `read-only`, piped grandchild `spawn EPERM`, ConstrainedLanguage errors. If approval is refused or unavailable, report the path as undiagnosed and stop.\n\n## Read the output\n\nEvery record reaches stdout, the `acl-report-*.jsonl` file under `-Out`, and the final `RECAP` line, which carries the verdicts, changes, verifications, refusals and scans. Tool output keeps only its tail, so read the recap first, and read specific records from the report when it is not enough. Trust `verification` records, never `completed` actions. Decide from `details.nextAction`:\n\n| `nextAction` | Meaning |\n|---|---|\n| `verify_original_confined_operation` | Repairs verified; repeat the original operation confined. |\n| `stop` | Nothing repaired, or a refusal ended the run. Report and stop. |\n| `restore_pending_then_stop` | Rollback unverified. Run the printed recovery commands in order, then stop. |\n\nA deny ACE's presence alone does not establish causation, and the script never removes one; DSH's `S-1-4-*` grants and the Everyone `DeleteSubdirectoriesAndFiles` deny are expected, not conflicts. A deny that blocks the repair ends the run without a repair (`REPAIR_REFUSED`, or `GRANT_FAILED`) after restoring what it attempted. Each change leaves two files in `-Out` (`acl-backup-.json` and its `.ps1`), and the run prints the matching `ROLLBACK` command.\n\n## After a repair, and when it stops\n\nRepeat the original failed operation confined. A verified repair is not undone because the original operation fails for a further reason: continue from the new observations, and if a deeper path is still denied, run the same one command there — again one call, one approval.\n\n**Stop after any failed or refused repair, or failed verification.** The script already restored that invocation's changes; do not repeat it or start another repair, and never remove a deny ACE by hand.\n\n**A stopped run still owes the user a decision.** Name the blocking object and the right or ACE it lacks, what changed and what was rolled back, the recovery commands in order, and the report path. Only the user can lift confinement, so ask them to switch this session to **full access** (Chinese UI: 完全权限) temporarily, and say what that opens: keep working under it, or ask you to keep investigating this file-permission problem, since the report already records the mechanism and the evidence. Then ask them to send this session as feedback, quoting the report records that matter, so the unhandled scenario reaches us — only what reaches this conversation travels with it. For example:\n\n> 这个工作区被 Windows 文件权限挡住了：`<对象>` 上 `<缺哪个权限 / 哪条权限项>`。你可以把本会话切到「完全权限」(full access) 临时继续工作；切完之后既能直接干活，也可以让我继续排查这个文件权限问题（报告里已经有机制和证据）。也麻烦把这个会话作为反馈发出去，好让我们补上这个场景。\n>\n> This workspace is blocked by Windows file permissions: `` ``. Switch this session to full access (完全权限) to keep working meanwhile; then either continue your work or have me keep investigating this file-permission problem from the report. Please also send this session as feedback so we can cover the case.\n\n## Never\n\n- edit ACLs by hand, change an owner, erase a deny, or replace child permissions recursively;\n- run the script elevated, through UAC or `runas`;\n- widen `-AllowRoot` to reach an ancestor, or repeat a denied or failed call.\n\nReport in the user's language: analyzed paths, changes, verification, recovery commands, next step. Label an authorized unconfined run as such, not as a sandbox repair.\n\n"}],"isError":false,"id":"{{message:6}}"}},"sourceEventSeqs":[15],"surfaceOp":"append"} +{"type":"tool/result","data":{"turn":1,"step":1,"message":{"role":"tool","source":{"kind":"tool","callId":"call_skill_load"},"toolCallId":"call_skill_load","content":[{"type":"text","text":"\n\nBase directory for this skill: {{cwd}}/.dsh/skills/diagnose-windows-sandbox-acl\nResolve relative paths mentioned by this skill against the base directory before using them. Load referenced resources only as needed.\n\n\n\n# Diagnose Windows sandbox ACL failures\n\n**Write every approval request in plain words, in the user's language.** The prompt is all the user reads before widening access, so it must stand alone: which folder the script touches, that it adds the signed-in user's full-control entry where a right is missing and removes foreign package entries, that file contents and owners are unchanged, and that the printed recovery command undoes each change. Keep error codes, `WRITE_DAC`/`WRITE_OWNER`, `S-1-15-2-*` SIDs, `icacls`, ACL, ACE, verdict names and switches out of the request: write Windows file permissions, not ACL or ACE. Ask once, for one command.\n\n## One command diagnoses and repairs\n\nThe script has no modes. `-Path`, `-AllowRoot` and `-Out` read the path and every ancestor and repair what the observations prove, in the same run:\n\n- directories on that chain that lack effective `WRITE_DAC` or `WRITE_OWNER` receive a full-control allow ACE for the signed-in user, because DSH cannot provision its workspace grant without them;\n- explicit AppContainer package allow ACEs (`S-1-15-2-*`, except the well-known groups ending in 1 or 2) are removed at their sources, ancestor first, which also removes those packages' access;\n- a directory the sandbox cannot provision — the state its provisioning error reports on the workspace root — or the authorized root itself also has its subtree searched for those ACEs in the same run, so a deeper entry needs no second request. That bounded walk reports `truncated` and unreadable directories; if truncated, pass the still-failing deeper path once more.\n\nEvery change is backed up, then verified by re-reading it. `-AllowRoot` bounds all of it: an object is changed only when it is that directory or strictly inside it, so a workspace root can repair itself. Never split this into a diagnostic call and a repair call, never ask twice for one repair, and never pass a mode switch that does not exist.\n\n```powershell\n& '\\scripts\\diagnose-windows-sandbox-acl.ps1' -Path '' -AllowRoot '' -Out ''\nexit $LASTEXITCODE\n```\n\nSubstitute full, quoted paths. Keep `-Out` persistent and user-owned, preferably beside the failing workspace; never use the skill resource directory, which is deleted when the skill unloads. Pass one path per invocation.\n\n## Run it\n\nRepeat the failing operation once confined. The script writes permissions, which the confined token cannot do — run it there and it would report the sandbox's own restriction as a missing right — so request approval for that one command and run it unconfined. A confined run of the script is never useful. Unconfined does not elevate the Windows token.\n\nDiagnose unexpected denials of workspace writes, listing, or plainly readable paths; explain expected ones instead: writes outside the workspace, any write in `read-only`, piped grandchild `spawn EPERM`, ConstrainedLanguage errors. If approval is refused or unavailable, report the path as undiagnosed and stop.\n\n## Read the output\n\nEvery record reaches stdout, the `acl-report-*.jsonl` file under `-Out`, and the final `RECAP` line, which carries the verdicts, changes, verifications, refusals and scans. Tool output keeps only its tail, so read the recap first, and read specific records from the report when it is not enough. Trust `verification` records, never `completed` actions. Decide from `details.nextAction`:\n\n| `nextAction` | Meaning |\n|---|---|\n| `verify_original_confined_operation` | Repairs verified; repeat the original operation confined. |\n| `stop` | Nothing repaired, or a refusal ended the run. Report and stop. |\n| `restore_pending_then_stop` | Rollback unverified. Run the printed recovery commands in order, then stop. |\n\nA deny ACE's presence alone does not establish causation, and the script never removes one; DSH's `S-1-4-*` grants and the Everyone `DeleteSubdirectoriesAndFiles` deny are expected, not conflicts. A deny that blocks the repair ends the run without a repair (`REPAIR_REFUSED`, or `GRANT_FAILED`) after restoring what it attempted. Each change leaves two files in `-Out` (`acl-backup-.json` and its `.ps1`), and the run prints the matching `ROLLBACK` command.\n\n## After a repair, and when it stops\n\nRepeat the original failed operation confined. A verified repair is not undone because the original operation fails for a further reason: continue from the new observations, and if a deeper path is still denied, run the same one command there — again one call, one approval.\n\n**Stop after any failed or refused repair, or failed verification.** The script already restored that invocation's changes; do not repeat it or start another repair, and never remove a deny ACE by hand.\n\n**A stopped run still owes the user a decision.** Name the blocking object and the right or ACE it lacks, what changed and what was rolled back, the recovery commands in order, and the report path. Only the user can lift confinement, so ask them to switch this session to **full access** (Chinese UI: 完全权限) temporarily, and say what that opens: keep working under it, or ask you to keep investigating this file-permission problem, since the report already records the mechanism and the evidence. Then ask them to send this session as feedback, quoting the report records that matter, so the unhandled scenario reaches us — only what reaches this conversation travels with it. For example:\n\n> 这个工作区被 Windows 文件权限挡住了：`<对象>` 上 `<缺哪个权限 / 哪条权限项>`。你可以把本会话切到「完全权限」(full access) 临时继续工作；切完之后既能直接干活，也可以让我继续排查这个文件权限问题（报告里已经有机制和证据）。也麻烦把这个会话作为反馈发出去，好让我们补上这个场景。\n>\n> This workspace is blocked by Windows file permissions: `` ``. Switch this session to full access (完全权限) to keep working meanwhile; then either continue your work or have me keep investigating this file-permission problem from the report. Please also send this session as feedback so we can cover the case.\n\n## Never\n\n- edit ACLs by hand, change an owner, erase a deny, or replace child permissions recursively;\n- run the script elevated, through UAC or `runas`;\n- widen `-AllowRoot` to reach an ancestor, or repeat a denied or failed call.\n\nReport in the user's language: analyzed paths, changes, verification, recovery commands, next step. Label an authorized unconfined run as such, not as a sandbox repair.\n\n"}],"isError":false,"id":"{{message:6}}"}},"sourceEventSeqs":[15],"surfaceOp":"append"} {"type":"step/end","data":{"turn":1,"step":1}} {"type":"step/start","data":{"turn":1,"step":2}} {"type":"assistant/message","data":{"turn":1,"step":2,"message":{"role":"assistant","content":[{"type":"reasoning","text":"The skill is loaded."},{"type":"text","text":"DONE"}],"source":{"kind":"model","provider":"deepseek-official","model":"deepseek-v4-flash"},"id":"{{message:7}}"},"usage":{"inputTokens":180,"outputTokens":10,"cacheReadTokens":0,"reasoningTokens":4},"stream":[{"type":"chunk","time":1788699430709,"chunk":{"type":"block-start","index":0,"blockType":"reasoning"}},{"type":"reasoning-chunks","time0":1788699430709,"index":0,"dt":[],"texts":["The skill is loaded."]},{"type":"chunk","time":1788699430709,"chunk":{"type":"block-start","index":1,"blockType":"text"}},{"type":"text-chunks","time0":1788699430709,"index":1,"dt":[],"texts":["DONE"]},{"type":"chunk","time":1788699430709,"chunk":{"type":"block-end","index":0,"block":{"type":"reasoning","text":"The skill is loaded."}}},{"type":"chunk","time":1788699430709,"chunk":{"type":"block-end","index":1,"block":{"type":"text","text":"DONE"}}},{"type":"chunk","time":1788699430709,"chunk":{"type":"usage","usage":{"inputTokens":180,"outputTokens":10,"cacheReadTokens":0,"reasoningTokens":4}}},{"type":"chunk","time":1788699430709,"chunk":{"type":"finish","reason":{"kind":"stop"}}}]},"surfaceOp":"append"}