Repository navigation
Conversation
|
Intent: Preserve generated SharePoint item locators on policy-owned exact trusted tenants without weakening higher-confidence credential redaction or markerless detection elsewhere. |
First Principles Review (Opus 5.5, fork) — 🔴 BLOCKPremise-level review of First-Principles-Verdict: BLOCK The whole attestation mechanism ships turned off. The only supplier returns Notes
What this change shipsInventory (8 items) — 0 justifiedIntent: keep provider-generated SharePoint links readable instead of showing them as
BlockersDormant seam: the fix ships switched off — zero consumers [FIRST-PRINCIPLES-REVIEWED] 5d9c69e Description read[DESCRIPTION-READ] 20f9792136e2ac08badaf6c0030429d751a7d8328c8177c11cf2a4d7b6408cf4 That sha256 names the PR description this verdict read. Recompute it with |
Design Review (Opus 5.5, fork) — ✅ PASSDesign-level review of Design-Verdict: PASS Nothing blocks: the empty public policy keeps every path failing closed, and the identity-cache split keeps KAS-only refusal semantics intact. Notes
[DESIGN-REVIEWED] 5d9c69e Description read[DESCRIPTION-READ] 3e442264b4138dae1b510a511e6faefaec816609c5b44e691a21c39e7b2e2a9c That sha256 names everything this verdict read: the PR description, plus the 1 evidence file(s) the workflow collected from it and handed this review. It covers that evidence because the media strip replaces every attachment URL with the same placeholder, so an edit swapping one attachment for another would leave the description's captured text identical while what this review judged changed. Recompute it with |
GPT 6.1 Review (fork) — 🔴 changes requested (blocking)Reviewed 1 of 1 blocking finding(s) are security-class and were withheld from adjudication, so the blocking verdict stands. BLOCKING -- test/test_chat_utils.py:522 -- Drain the locator loader before home isolation unwinds
[BLOCK-MERGE] 5d9c69e Adjudication (Opus 5.5) — is blocking on each finding proportionate?The adjudicable block has no findings, and the one fenced finding (F1) stays blocking as UPHOLD-FENCED. I did not FLAG it: a FLAG needs the conditions to be rare in practice, and the repository has already recorded this exact pattern causing a real CI flake. F1 — UPHOLD-FENCED. Each step it needs is real, in the trusted base or the patch:
|
Opus 5.5 Review (fork) — ✅ no blocking findingsReviewed |
Security Scope Review (fork) — ✅ PASSWhich legitimate operations does adjudicated, zero confirmed regressions Scope-Verdict: PASS Reason: The change does not tighten the shell deny composite. It touches no TIGHTENING: gateway-side locator-attestation admission (and the readability cache it depends on) -- it refuses to treat a SharePoint link that an agent wrote, or echoed from the call input, as a provider-generated locator: ADJUDICATED: 13 candidates, 0 confirmed newly-refused PLATFORM COVERAGE: posix, windows and any. Posix and windows rows cover the path and executable spellings near the touched area: the new and touched test modules through UNADJUDICATED: Three guards are outside the shell classifier:
[SCOPE-REVIEWED] 5d9c69eSecurity scope review -- adjudicated candidates
No candidate flipped from allowed to refused. |
|
🤖 Kiro Crew [operator: NicholasRBowers#a942f9ca]: This PR has been inactive for 7+ days. I reviewed the blockers but they require your input:
The remaining items are mechanical and can wait for that ruling: re-porting the pass-3 decline into #12684's When you've addressed these, the pipeline will re-assess on its next cycle. |
|
PR triage (open PRs against I checked the head What changed on main since the last triage note
Functional gap I reproduced on this head (advisory, not P0). I trusted Not raising a new blocker. The open GPT 5.6 BLOCK (forged trusted-tenant locator skips pass 3) already covers the security concern. There is no breaking change: the default Recommendation: REBASE onto main, then merge discussion with the owners of #15792 / #16952 about one shared pass-3 link-decline mechanism. After that, continue development on the late-separator gap. |
85e06a1 to
7b1d3ad
Compare
|
Rebuilt on current main at 37449b3 with the accepted durable-attestation design. URL text alone grants no exemption: only gateway-resolved, companion-declared provider fields can persist bounded SHA-256 locator digests in the sealed leaf, and redaction.py no longer imports exfil or platform. Parent verification passed 123 attestation/display/flake tests; the broader focused run passed 533. Body accounting, push guard, and the fresh lease passed on 7b1d3ad. CI and security review are pending. |
UX Review (Opus 5.5, fork) — ⏭️ skippedRevision |
7b1d3ad to
a2f03f4
Compare
|
|
Pushed 2941b0b, rebased onto 7796e2a.
|
|
Pushed e06cdd9, rebased onto 99cced3.
|
|
Pushed 6462323, rebased onto 5f77c3f.
|
|
Rebased onto main Resolution: both budget notes kept. This PR's note now starts from 32,447, and the budget is re-measured at 32,507 (main's 32,447 plus this PR's 60 lines, unchanged). Main's The rebase also picks up two main fixes for Windows reds on the previous head, neither from this diff: #23449 (the Checks on
|
|
Rebased onto main Resolution: both budget notes kept. This PR's note now starts from 32,518, and the budget is re-measured at 32,578 (main's 32,518 plus this PR's 60 lines, unchanged). Main's edits to files this PR also changes merged without conflict: Checks on the rebase:
|
|
Rebased onto main Resolution: main's rewritten paragraph is kept, followed by this PR's SharePoint bullet, byte-identical to the previous head's. Main's other edits to files this PR also changes merged without conflict: Checks on the rebase:
|
|
Rebased onto
Tests:
|
|
Rebased onto
Main's edits to Tests:
|
|
Rebased onto
Main's edits to Tests, run on
Main then moved to |
|
Rebased onto
The rest of the patch merged without conflicts. That includes main's edits to Tests, run on
Main then moved twice more: #20821, #22272 and #22300, then #23538, #23550, #23712, #20410, #21222 and #23725. Neither batch touched a file this PR changes. On each base I re-ran the tests those commits touched plus this PR's tests: 749, then 1,037 passed, with the composition contracts at 65 both times. |
|
Rebased onto
The rest of the patch merged without conflicts. Tests, run on
Main then moved to |
|
Rebased onto
The other 38 paths have the same diff as on Tests:
|
|
Rebased onto
The rest of the patch merged without conflicts, including Tests, run on
|
|
Rebased onto 6b1b65d to clear a conflict with #24392 in
|
|
Rebased onto main to resolve conflicts. Each resolution keeps both sides:
No code of this PR changed. This PR's tests, the tests main changed since the last push, the composition contracts and the static gates pass. The only failures are ones main shows on this host too. |
|
Pushed 8852bf4, rebased onto main On |
|
Rebased onto main |
|
Rebased onto main
No behavior change beyond the merge. |
|
Pushed fc44ffe, rebased onto main |
|
Pushed f7b4389, rebased onto main
Nothing else changed. |
|
Pushed 688a099, rebased onto main |
Code change approver recommendation: A person should decideWhy: the decider raised 1 possible problem and estimates a 60% chance that this change introduces a defect; a person should check it. The code looks careful and fails closed: the public policy is empty, the redactor only skips locators that match a digest, and KAS permission handling still sees a KAS-only view of the verdict cache. But the PR does not reach its stated goal. No supplier declares provider fields, so no SharePoint link is kept in any deployment, and the PR itself says enabling it needs a maintainer decision (label: needs-author-decision). It adds a large dormant mechanism (durable registry, boot preload, revalidation threads, identity-verdict recording for every harness) that delivers nothing yet. Request changes: land it with an approved supplier, or get maintainer sign-off first. Things to check (possible problems the decider raised; at this confidence many turn out to be fine):
An automated recommendation, not a review: a maintainer still decides. It approves on its own only when the estimated chance of a defect is at most 12% and the change has at most 400 changed lines. Decider v3c (217fe607) · commit |
|
Pushed 5d9c69e, a conflict-only rebase onto main |
|
|
Pushed 662c9d4, rebased onto main
The First Principles block is the supplier decision recorded earlier and still needs a maintainer. No other code, test or description changed. |
|
Pushed b276e81, rebased onto main
No other code, test or description changed. |
Problem / Motivation
Goal: Preserve provider-generated SharePoint item locators without allowing forged URL text to suppress credential redaction.
A benign 46-character SharePoint item ID can look exactly like a bare AWS secret and be rewritten as
[REDACTED: credential]. URL shape and a trusted tenant name alone cannot prove that the bytes came from SharePoint.Why it matters
The false positive destroys links in rendered messages and persisted history. A text-only carve-out would create a credential-egress path because a model can forge the same URL shape.
No production companion currently supplies the provider-field declaration. The seam remains empty until maintainers decide that companion evidence justifies enabling a specific
(server, tool, field)mapping.Not a goal
What changed (motivation → approach → change)
A gateway-only admission path accepts locators only from companion-declared result fields after successful tool completion, resolved MCP identity, retained complete same-call input, exact trusted-host membership, strict SharePoint route validation, and input-echo rejection. Every shared-runtime ACP harness records an explicit identity-readability verdict. Missing evidence, input, or policy fails closed before a worker is scheduled. The public empty policy schedules no admission work.
Accepted locators are stored as canonical SHA-256 digests in a bounded, owner-only, atomic registry under the existing sandbox-hidden gateway authority root. Dashboard and API-only gateways schedule registry preload on a worker thread after readiness without delaying the listener. Cold readers return an empty snapshot, and other process readers use a bounded background file-identity and generation check to adopt cross-process writes. Event-loop snapshot lookups never resolve the registry path: they compare the in-memory home inputs captured with the snapshot, and a changed data-home override returns an empty snapshot until the background load resolves the new path.
The credential redactor stays free of locator-registry imports and I/O. Gateway, transcript, display, and dashboard tool-result boundaries inject immutable digest snapshots;
redact()hands its snapshot to every credential pass of the composed pass. The empty default keeps stdio and public callers under markerless credential redaction. Gateway result parsing awaits attestation before redaction.docs/system-specs/modules/acp-client.md,platform-context.md, andsecurity.mdsrc/kiro_crew/acp/locator_attestation.py, the_dispatch.pyfacade with its_dispatch_parts/owners, and thesession_handle_*ownerssrc/kiro_crew/platform/interfaces.py,defaults.py, andcontext.py(server, tool)result-field declaration with an empty public default.src/kiro_crew/locator_attestations.pyandsrc/kiro_crew/_durable_vouch.pysrc/kiro_crew/dashboard/server.pywith itsserver_runtime/api_entrypoint.pyandserver_runtime/maintenance.pyowners, thechat_utils.pyfacade with itschat_utils_display.pyowner,display_battery.py, andsrc/kiro_crew/history_append.pysrc/kiro_crew/security/redaction.pywith itsredaction_entropy.pyowner,text_redaction.py, andpass_compose.pysrc/kiro_crew/security/paths_lists.pyvouched-executionsroot; this PR adds no new sealing mechanism.test/test_sharepoint_locator_attestations.py, ACP launch/composition tests,test_session_handle_bindings.py,test_chat_utils.py,test_display_verdicts.py, andtest_tool_call_classification_kas.pyBackwards compatibility
Compatible: the new adapter declaration is optional and fails to an empty mapping. Public and non-gateway behavior remains full markerless redaction until a companion declares exact provider-generated fields. The one-argument document-link span helper and other signatures visible on
mainremain callable as before.Tests
test_dashboard_tool_result_field_keeps_the_link_the_gateway_result_kept,test_credential_beside_an_attested_locator_is_redacted_on_every_composed_pass, andtest_tool_field_reads_the_snapshot_once_and_only_for_sharepoint_text, which fail on the previous head because the dashboard tool-result redactor gave the composed pass no snapshot.test_unattested_or_cold_locator_stays_redacted_through_tool_field_and_plain_redactpins the fail-closed side.main, all in the branch-unchanged ACP projection-stack classTestRuntimeProjectedLaunchDocuments. Without a builtwebsite/dist, all 122 tests in the four server composition contract files pass, with locator preload in the expected post-readiness position.git diff --checkpassed.Manual verification
N/A. Unit tests cover admission, persistence, cold load, cross-process refresh, restart, batch, stream, neighboring-secret, history, and display-cache behavior.
Screenshots / video
Why no screenshot: The change has no rendered UI delta and only changes provider provenance and credential-redaction decisions.
Related Issues
No public issue links this security defect. Maintainer disposition is required before this zero-supplier production seam can be enabled.
Pattern harvest
Rule candidate: review-prompt
Pattern: a credential heuristic can exempt provider bytes only after a non-agent-writable boundary attests exact provenance. URL text and host membership are claims, not proof.
Checklist
Contribution License Agreement
Contributions are accepted under the repository's Apache 2.0 license.