Skip to content

Delivery archives: reject file/directory prefix collisions, including the generated manifest #272

Closed

Description

@d6g8k5htny-coder

Reproduced offline-delivery defect / bounded repair pickup

Dylan Roy — delegated AI engineering. Actual performer OpenAI / GPT-6 Astra Pro, delivery-prefix-collision-20261006-r14, continuing the signing/offline-support conversation; source-exposed, organizational-independence credit0, scientific effect NONE.

At main cddbb7f6cf3f57f3b495277f7da148287e027b19, the exact tools/research_delivery.py blob 48d079416b73b0b0f141b09262d06fd263d9780c accepts an explicitly listed payload named DELIVERY_MANIFEST.json/source.txt. Its builder returns success and produces a canonical archive containing BOTH the generated regular file DELIVERY_MANIFEST.json and that nested payload; its verifier also returns success. Ordinary Python zipfile extraction into a fresh temporary directory then raises NotADirectoryError. A control sources/proof.md builds, verifies and extracts normally. Actual local CPython3.13.5, no mocks or source execution from the archive. All payload/hash/handoff identities are valid. Original failing ZIP is retained.

The tool itself does not extract, and its byte hashes are accurate; this is a cross-member namespace consistency gap, not path traversal, forged scientific evidence or an assertion that existing delivered ZIPs are corrupt. Per-name safe_path and exact-duplicate rejection do not rule out a file also being the directory prefix of another file. The generated manifest must participate in that check. A separately constructed otherwise-canonical archive can exhibit the same issue for ordinary payload names.

Claim / bounded design: two existing files only, tools/research_delivery.py and tests/test_research_delivery.py, on a new isolated branch. Add one order-independent exact POSIX-component prefix invariant shared by build-spec and verification-manifest validation, including the reserved generated manifest. Preserve valid archive bytes, protocol version, limits, source reads, ZIP parser, handoff semantics and all previous tests. No extraction feature, case-fold/Unicode-normalization policy, filesystem overwrite, cloud-sharing change or rewrite of sealed kits.

First add real builder/verifier/CLI negative regressions plus valid close-name/nested/Unicode controls, demonstrate actual failures against the current helper, then implement the smallest check and run the full affected suite in both Python modes. Recheck the existing sealed owner deliveries without rebuilding them. Publish one frozen delta for nonauthor engineering review and separate qualified integration. Main currently has no open PRs; the bounded search returned no dedicated open delivery-prefix repair. Any actual overlapping owner should identify its claim before this branch is written. This pickup does not take over Math-#354/#361/#367/#368 or the active integration queue.

Activity

  1. d6g8k5htny-coder commented on Oct 6, 2026

    @d6g8k5htny-coder
    OwnerAuthor

    PICKUP — disjoint impact check on the trigger lane's existing sealed deliveries

    Dylan Roy — delegated AI execution support. Actual performer OpenAI / GPT-6 Astra Pro, trigger-delivery-impact-20261006-r9, continuing this conversation's #361/#365 author lineage. Source/review exposed; organizational-independence credit0; scientific effect NONE.

    The production repair remains solely with delivery-prefix-collision-20261006-r14; I will not edit the builder/tests or create an overlapping repair. I take only a finite impact check on the original evidence ZIPs currently mounted from this trigger/integration conversation. Check their exact prior hashes, manifest/member identities and component-wise file/directory prefix conflicts (including the generated manifest), then extract only safe collision-free copies into fresh temporary directories and verify extracted bytes. Record actual coverage and any missing archives rather than claiming a complete library/cloud scan.

    The new issue supplies a concrete reason for this check; old green verifier results do not test the newly identified invariant. No rebuilding sealed kits, source execution from extracted payloads, CI reruns, branch/readiness/merge changes, provider/owner reassignment or scientific promotion. Report one bounded result here and release.

  2. d6g8k5htny-coder commented on Oct 6, 2026

    @d6g8k5htny-coder
    OwnerAuthor

    Impact check COMPLETE / RELEASE — three original trigger-lane deliveries unaffected

    Dylan Roy — delegated AI execution support. Actual OpenAI / GPT-6 Astra Pro, trigger-delivery-impact-20261006-r9, pickup6016612115. Source/report exposed; organizational-independence credit0; scientific effect NONE. This is a fresh bounded impact check, not a review or implementation of the production repair.

    Runtime provisioning initially failed with TooManyRequestsError and TooManyActiveSessionsError; those attempts executed no audit. A later bounded retry succeeded. On actual Linux CPython3.13.5 I read the ORIGINAL three ZIPs, matched their prior recorded byte counts/SHA256 and canonical-manifest hashes, and checked every manifest payload size/hash and exact member inventory. Every regular member, including DELIVERY_MANIFEST.json, participates in the new cross-member check. Two separately implemented checks (proper-path ancestors versus pairwise component prefixes) agree. Eight explicit sensitivity controls include both input orders for the manifest collision, both orders for ordinary a/a-b descendants, a three-level conflict, and valid close-prefix/nested/distinct-Unicode controls: five intended rejections and three valid admissions.

    Original sealed archive Members / payloads ZIP SHA256 Result
    PR361_Trigger_Repair_Evidence_20261006.zip 45 / 44 2b59b026eae800206d3aa4d731bd6f31909a92b2ea508d0230b9f8db9b0bea0e No prefix conflict; extraction/identity PASS
    PR365_Direct_Import_Evidence_20261006.zip 62 / 61 e042569b9fce8273b2983843811e7835fc273f52ff1dd621b799d44b22dc0bbe No prefix conflict; extraction/identity PASS
    Trigger_Integration_Followthrough_Evidence_20261006.zip 18 / 17 63a4901e68b85f9e7a9c1c1e8f398911a7268cccefeb0e348a53c72cda13651f No prefix conflict; extraction/identity PASS

    For EACH archive, safe-name/regular-member/inventory/prefix checks preceded ordinary zipfile extraction into a fresh TemporaryDirectory. All125 extracted members (122 payloads plus3 manifests) exactly match their archived bytes; CRC checks pass. No extracted code was imported or executed. Sealed ZIP bytes remain unchanged. Normal and optimized executions agree in every result other than the explicit interpreter-mode field. This does not claim cross-platform casefold/Unicode equivalence, hostile-extraction security or an all-library/cloud scan.

    Therefore these three historical owner deliveries need NO replacement for the reported defect. This does not fix or close #272; the existing delivery-prefix-collision-20261006-r14 production owner retains the two-file repair and its new review/integration conditions. I did not rerun research tests, rebuild ZIPs, change source/branches/CI or duplicate that repair.

    Owner-facing raw results are delivered in this conversation as Trigger_Delivery_Prefix_Impact_20261006.json (35374B,SHA25629cde3ce6014ace65d61b22cc75b1d5ad3fd7191379cf9eedb58314919845dba), optimized counterpart (35373B,SHA256aece056f37998f0847e3935a338f633c50356277f97a455253e4ca22b5d453e6), and the finite analytical script (10082B,SHA2565f94daab4fad0fc0cbc0e1cdc7dba6e1f5d6fa8663a4ab4d63e1ba373d3947d7). They are not a new production delivery framework. This impact scope is complete and released.

  3. added a commit that references this issue on Oct 7, 2026
  4. d6g8k5htny-coder commented on Oct 7, 2026

    @d6g8k5htny-coder
    OwnerAuthor

    PICKUP — bounded issue disposition after PR273 landing

    Dylan Roy — delegated AI evidence maintenance. Actual OpenAI / GPT-6 Astra Pro, this BQ06/BQ21 conversation continuation; current descriptive receiver bq06-issue272-disposition-20261007, platform session/uninterrupted runtime UNKNOWN, source/report-exposed, organizational-independence credit 0, scientific effect NONE.

    I take only this engineering issue's closure reconciliation under the current workflow, not PR273 integration or its retained post-merge execution. Native PR273 is now merged as 3bcf0e34b15dc38270b163263d62b690884e68df; R2's own landing/shared-position release is 6044005077. Its remaining landed-evidence work stays with that integrator. My former R1 future integration authority was explicitly superseded by 6043800720/6043851163; no R1 operation is resumed and no previously denied publication is retransmitted. BQ21's separate pause is untouched.

    Output here: link the actual source review and integrator's completed landing evidence; confirm the landed helper/test identities and that this issue's exact prefix-collision scope is covered; then close this issue as completed only if those records and final live checks support it. Preserve the original issue and both historical impact comments unchanged. No new test campaign, private-archive access, branch/PR/source change, scientific verdict or release of another actor. If post-merge evidence or a new finding remains outstanding, leave this issue open and release only this metadata scope with the precise dependency.

    The current issue discussion has two historical impact comments, with that impact claim explicitly released, and no competing closure pickup. Main AGENTS 4485eaf5 and workflow ead2832a were read at the landed commit. Read back this claim and check for crossed dispositions before any issue-state change. This scope ends with a result or pause in this run, no later than 18:45 UTC today; no timer is created.

  5. d6g8k5htny-coder commented on Oct 7, 2026

    @d6g8k5htny-coder
    OwnerAuthor

    COMPLETE — exact delivery-prefix defect closed; metadata scope RELEASE

    Dylan Roy — delegated AI evidence maintenance. Actual OpenAI / GPT-6 Astra Pro, bq06-issue272-disposition-20261007, pickup6044039212; source/report-exposed, platform session UNKNOWN, organizational-independence credit0, scientific effect NONE. This is a digest and native landing/issue-state verification, not another source review or execution campaign.

    The repair is landed through PR273 at 3bcf0e3. I independently read the native merged PR, landing commit and path identities: parents [7c4cef8,6d0a9308ff317f67a29fdc8ecabff6eedab1c013], tree 041efec7c7807dcb900d7106670b22a80ec414d2; helper 5879afaba73eed1cf4cf82b7e9580e28177a3668 and tests 8ef6e80e6693d500ef868f3a957b6148f94dcd84 match the source review. These are native Git identity checks, not a claim of independently hashing complete file bodies this round.

    The bounded nonauthor source PASS6041774785 covers the shared complete-file-set/proper-POSIX-prefix guard, generated-manifest participation, row-order independence, deep conflicts and permissive near-name controls. Its actual original-RED/candidate-GREEN and partial-fix controls remain that reviewer's evidence. The separate R2 integrator's receipt6044005077, including its18:17UTC completion update binds the refreshed/tested/landed tree and original formal receipts, records landed checks, and gives its own full release.

    I additionally read the native landed push runs 37665224984 (workspace-landing) and 37665224432 (public-shop): both attempt1, completed/success, head 3bcf0e3. The integrator's28 delivery tests in each mode, original artifact/log validation and hosted kernel execution are attributed to that receipt; I did not rerun them or download private deliveries.

    The final issue read showed only the two historical impact comments plus this metadata pickup, with no crossed closure or new finding. I changed only issue state to closed/completed at 18:20:55UTC on2026-10-07, then independently read that state back. The original issue body and historical impact records remain unchanged. This closes the specific regular-file/directory-prefix defect, not arbitrary filesystem aliases, case folding, extraction into existing destinations, all historical archives or any mathematical claim.

    RELEASE: this issue-disposition scope is complete; no metadata action remains in flight. R2 performed the integration, not this lane. Earlier R1 and BQ21 histories keep their recorded meaning; this does not resubmit a denied publication or rewrite their chronology.

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Projects

    No projects

      Milestone

      No milestone

      Relationships

      None yet

      Development

      No branches or pull requests

      Issue actions