You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
Protected main is exact 87c4daa1830bac5a5228b6036752ad5633232085 (verified signed). OriginWeave still has no immutable GitHub Release. Protected AGENTS.md permits scheduled agents to create bounded reviewed PRs but explicitly forbids them from altering .github/workflows/**, merging, tagging, or publishing. This issue therefore remains the canonical handoff for MV3 workflow-owner action rather than a reason to weaken browser isolation in a leaf product branch.
Three active workflow-owned decisions remain distinct and must not be conflated.
PR #43 (test/mv3-downloads) is open/Draft/mergeable at exact 6f3134d18d3118aab33d28048671dc71a5f47b77. Its recorded base is stale main@c789b802...; protected main is now 87c4daa....
The leaf generation contains the three validated helper mechanics in .github/workflows/mv3-compatibility.yml:
root ownership for the pinned archive's chrome_sandbox;
mode 4755; and
CHROME_DEVEL_SANDBOX bound to that helper.
Those mechanics executed successfully on exact #43: MV3 33866932365 / job 101003808036, CI 33866932436, Security Scan 33866932442, and Semgrep 33866932380 are terminal success. Artifact 9938906825 (sha256:8644c2d16ce9d91ff7bb46075793b313507d3f37b590965a686f9b05a61ce519) reports pinned Chrome 150.0.7871.129 and 3/3 real-browser trials. This establishes technical feasibility for that leaf generation only; it does not make the leaf workflow canonical.
PR #148 is open/Draft/mergeable at exact 0135984f1bc1f68d89d7777f49c4999474105a12 on exact #1473dff28d9bf2dd27b72507e39979d51b8bf140fb4. Its consumer runner contains zero --no-sandbox launch overrides across ordinary MV3, ordinary Agent Task, forced-close, and browser-crash paths, and it retains bounded failure_stage/reason_code plus primary-vs-cleanup causal preservation.
That source boundary has now been exercised on the exact head rather than inferred. Repository CI 33990522263 is terminal success, including exact numerical production coverage of 415 functions / 3,555 lines / 4,444 regions / 476 branches = 100%. Manifest V3 Compatibility 33990522248, job 101371812631, is terminal failure after successfully downloading pinned Chrome/ChromeDriver and reaching the real fixture. Artifact 9977680352 (sha256:5430103b6bae59acc0d62d3c21cce559b27a15eadc6f049a857551d9a92ef643) reports Chrome 150.0.7871.129 / r1639810, complete profile cleanup, and 0/3 in every real-browser lane: ordinary MV3, ordinary Agent Task, forced-close Agent Task, and browser-crash Agent Task. The browser-crash lane identifies failure_stage=session_create, failure_type=RuntimeError, reason_code=runtime_error; browser_passes is empty. isolation_complete=true is cleanup bookkeeping, not browser success.
This paired evidence is stronger than the earlier differential: when the inherited --no-sandbox escapes were removed under the unchanged current workflow, every real Chromium launch failed closed; the separately authorized #43 helper generation proves the same Chrome generation can start sandboxed when the helper mechanics are present. Do not transfer #43 GREEN to #148 or claim causality beyond that paired boundary, but do not blind-rerun #148 under the unchanged workflow either.
Required owner action: reconstruct/adopt the validated helper mechanics, or an equivalent reviewed least-privilege implementation, against the current protected workflow generation. Do not copy an obsolete workflow blob. Preserve harden-runner/egress, immutable actions, lifecycle/concurrency controls, least privilege, evidence artifacts, and fail-closed behavior. After canonical adoption, consumers must adopt/restack non-destructively and regenerate exact-head sandboxed Linux browser evidence. Do not restore --no-sandbox, reduce trial counts, hide session_create, or reinterpret cleanup as task success.
PR #290 (codex/avoid-mv3-draft-transition-runs) remains open/Draft/mergeable at exact ebeefcd534db4324498fdb18046ebc6255ddcdf2, but it is based on #245a769f484e2c110e0523b3b28cd21573f43867562, not protected main.
Its valid intent is to isolate pull-request concurrency by workflow + repository + PR number, keep manual dispatch independent from PR cancellation, admit useful PR synchronization/Ready events, and avoid spending hosted runners on Draft PR jobs. That intent should be preserved.
The predecessor-relative workflow patch itself is not safe to promote verbatim. Protected #286 already admits opened, synchronize, reopened, ready_for_review, converted_to_draft, and closed, with a closed-event fail-closed guard. #290 would remove converted_to_draft and closed from the trigger list and simplify the job guard, regressing protected lifecycle/capacity semantics. Reconstruct the useful concurrency/manual-dispatch intent on the current protected workflow while retaining #286's converted-to-draft/closed handling, permissions, harden-runner egress restrictions, immutable pins and browser evidence.
PR #245 (feat/block-web-audio-fingerprinting-v3) remains open/Draft/mergeable at exact a769f484e2c110e0523b3b28cd21573f43867562 on prerequisite feat/stealth-profile-coherence@ba364f63a5cfe61270b1b4553102deaf98724382, 14 commits ahead / 0 behind. Native CI 33278168850 and MV3 33278168852 are successful on that unchanged head.
Its product/privacy slice is valid, but the branch also changes .github/workflows/mv3-compatibility.yml to add Web Audio trigger paths, execute scripts/ci/run_web_audio_privacy.py, retain web-audio-privacy.json, and update the immutable actions/upload-artifact pin. Native GREEN proves execution, not workflow ownership. The authorized workflow owner must independently decide the smallest canonical integration without losing the browser/privacy regression.
Canonical owner acceptance
Adopt the sandbox-helper setup in the current protected MV3 workflow generation without disabling Chromium sandboxing.
Regenerate exact-head repository CI, 100% owned-production coverage/rustdoc, security/governance checks, and real pinned-Chromium evidence before any merge/release claim.
Historical authorization and differential evidence remain in this issue's comments. They are useful lineage, not transferable current-generation authority. Keep this issue open until the active consumers use an authorized current workflow contract and their claimed browser evidence executes on those exact heads.
Current governance boundary
Protected
mainis exact87c4daa1830bac5a5228b6036752ad5633232085(verified signed). OriginWeave still has no immutable GitHub Release. ProtectedAGENTS.mdpermits scheduled agents to create bounded reviewed PRs but explicitly forbids them from altering.github/workflows/**, merging, tagging, or publishing. This issue therefore remains the canonical handoff for MV3 workflow-owner action rather than a reason to weaken browser isolation in a leaf product branch.Three active workflow-owned decisions remain distinct and must not be conflated.
#43 / #148 — Chromium sandbox helper
PR #43 (
test/mv3-downloads) is open/Draft/mergeable at exact6f3134d18d3118aab33d28048671dc71a5f47b77. Its recorded base is stalemain@c789b802...; protected main is now87c4daa....The leaf generation contains the three validated helper mechanics in
.github/workflows/mv3-compatibility.yml:chrome_sandbox;4755; andCHROME_DEVEL_SANDBOXbound to that helper.Those mechanics executed successfully on exact #43: MV3
33866932365/ job101003808036, CI33866932436, Security Scan33866932442, and Semgrep33866932380are terminal success. Artifact9938906825(sha256:8644c2d16ce9d91ff7bb46075793b313507d3f37b590965a686f9b05a61ce519) reports pinned Chrome150.0.7871.129and 3/3 real-browser trials. This establishes technical feasibility for that leaf generation only; it does not make the leaf workflow canonical.PR #148 is open/Draft/mergeable at exact
0135984f1bc1f68d89d7777f49c4999474105a12on exact #1473dff28d9bf2dd27b72507e39979d51b8bf140fb4. Its consumer runner contains zero--no-sandboxlaunch overrides across ordinary MV3, ordinary Agent Task, forced-close, and browser-crash paths, and it retains boundedfailure_stage/reason_codeplus primary-vs-cleanup causal preservation.That source boundary has now been exercised on the exact head rather than inferred. Repository CI
33990522263is terminal success, including exact numerical production coverage of 415 functions / 3,555 lines / 4,444 regions / 476 branches = 100%. Manifest V3 Compatibility33990522248, job101371812631, is terminal failure after successfully downloading pinned Chrome/ChromeDriver and reaching the real fixture. Artifact9977680352(sha256:5430103b6bae59acc0d62d3c21cce559b27a15eadc6f049a857551d9a92ef643) reports Chrome150.0.7871.129/r1639810, complete profile cleanup, and 0/3 in every real-browser lane: ordinary MV3, ordinary Agent Task, forced-close Agent Task, and browser-crash Agent Task. The browser-crash lane identifiesfailure_stage=session_create,failure_type=RuntimeError,reason_code=runtime_error;browser_passesis empty.isolation_complete=trueis cleanup bookkeeping, not browser success.This paired evidence is stronger than the earlier differential: when the inherited
--no-sandboxescapes were removed under the unchanged current workflow, every real Chromium launch failed closed; the separately authorized #43 helper generation proves the same Chrome generation can start sandboxed when the helper mechanics are present. Do not transfer #43 GREEN to #148 or claim causality beyond that paired boundary, but do not blind-rerun #148 under the unchanged workflow either.Required owner action: reconstruct/adopt the validated helper mechanics, or an equivalent reviewed least-privilege implementation, against the current protected workflow generation. Do not copy an obsolete workflow blob. Preserve harden-runner/egress, immutable actions, lifecycle/concurrency controls, least privilege, evidence artifacts, and fail-closed behavior. After canonical adoption, consumers must adopt/restack non-destructively and regenerate exact-head sandboxed Linux browser evidence. Do not restore
--no-sandbox, reduce trial counts, hidesession_create, or reinterpret cleanup as task success.#290 — PR concurrency / lifecycle intent
PR #290 (
codex/avoid-mv3-draft-transition-runs) remains open/Draft/mergeable at exactebeefcd534db4324498fdb18046ebc6255ddcdf2, but it is based on #245a769f484e2c110e0523b3b28cd21573f43867562, not protected main.Its valid intent is to isolate pull-request concurrency by workflow + repository + PR number, keep manual dispatch independent from PR cancellation, admit useful PR synchronization/Ready events, and avoid spending hosted runners on Draft PR jobs. That intent should be preserved.
The predecessor-relative workflow patch itself is not safe to promote verbatim. Protected #286 already admits
opened,synchronize,reopened,ready_for_review,converted_to_draft, andclosed, with a closed-event fail-closed guard. #290 would removeconverted_to_draftandclosedfrom the trigger list and simplify the job guard, regressing protected lifecycle/capacity semantics. Reconstruct the useful concurrency/manual-dispatch intent on the current protected workflow while retaining #286's converted-to-draft/closed handling, permissions, harden-runner egress restrictions, immutable pins and browser evidence.#245 — Web Audio privacy workflow integration
PR #245 (
feat/block-web-audio-fingerprinting-v3) remains open/Draft/mergeable at exacta769f484e2c110e0523b3b28cd21573f43867562on prerequisitefeat/stealth-profile-coherence@ba364f63a5cfe61270b1b4553102deaf98724382, 14 commits ahead / 0 behind. Native CI33278168850and MV333278168852are successful on that unchanged head.Its product/privacy slice is valid, but the branch also changes
.github/workflows/mv3-compatibility.ymlto add Web Audio trigger paths, executescripts/ci/run_web_audio_privacy.py, retainweb-audio-privacy.json, and update the immutableactions/upload-artifactpin. Native GREEN proves execution, not workflow ownership. The authorized workflow owner must independently decide the smallest canonical integration without losing the browser/privacy regression.Canonical owner acceptance
Historical authorization and differential evidence remain in this issue's comments. They are useful lineage, not transferable current-generation authority. Keep this issue open until the active consumers use an authorized current workflow contract and their claimed browser evidence executes on those exact heads.