PaperClipAI/doc/plans/2026-09-20-github-review-live-acceptance.md

GitHub review bot live acceptance — 2026-09-20

The user authorized unattended testing on the disposable private repository, including a Storybook bot whose review fails unless a rendered story page says oogabooga. Continue from the approved implementation; do not pause for another design approval. Use the embedded browser for user journeys and real signed GitHub deliveries with actual Paperclip agent runs for execution evidence.

User stories and acceptance

1. As a member, I connect my own GitHub identity and confirm it. My requests identify me as the responsible person; the agent uses the bot App's governed tools, never my credentials for review publication.
2. I mention the bot on an issue. A real task/run is assigned to its permanent agent and a reply appears on GitHub. A second mention continues the task.
3. I open a PR containing a known defect. The configured event starts a normal task/run, publishes a justified finding and summary, and fails Paperclip Review on the exact head commit. I can follow links to the task and run.
4. I ask for another review by mentioning the bot again. A new run reassesses the current commit without creating another root conversation/task or duplicate inline findings. Ordinary discussion does not change the rating.
5. I push a fix with updated-commit reviews enabled. A new assessment passes and supersedes the old result, preserving review history.
6. I select mentions-only, or disable updated-commit events. A push does not run an automatic review or inherit a passing check from an older commit. An authorized mention still starts a review. Re-enabling the event restores automatic review on the next push.
7. I configure a Storybook bot with trusted instructions to generate/build Storybook from the PR's page components and inspect rendered story pages. Source comments, PR prose, story names, or instructions mentioning oogabooga are not evidence; the text must be visible in a rendered page. If absent the structured assessment is complete but fails; if present it passes. A build or verification failure is incomplete and cannot pass.
8. For that bot, absent → present → absent commits produce failing → passing → failing checks. Repeat mentions also recheck the PR. Generated Storybook and page inspection evidence are attached to the underlying task/run.
9. Draft filters, prompt changes, repository restrictions, tool denials, duplicate delivery, rapid pushes, stale publication, retries, and guest restrictions preserve their configured authority and exact-head semantics.
10. Formal approval/request-changes is rejected while disabled and accepted through the governed tool only when explicitly enabled; scoring 5/5 alone never approves a PR. Required checks are enforced on the disposable repo.
11. After local qualification, deploy matching application/migrator and Cloud gateway revisions to a dedicated staging tenant, and repeat the real signed webhook and agent workflow through its vanity hostname.

Evidence rules

Record actual source revisions, heads, delivery IDs, task/run IDs and links to comments, reviews, checks, and generated stories. Distinguish real runs from fixtures and deterministic tests. Never publish a result manually and count it as an agent run. Keep failures and repairs in the log; do not infer a live pass from unit tests or setup verification. No production rollout or merge.

Journey log

• Embedded browser refreshed successfully; form controls respond again.
• Existing personal GitHub flow entered from the catalog. Access limited to the QA agent. Approved the isolated QA instance's normal Cloud connector enrollment; the preserved flow resumed the GitHub sign-in step.

Live results: basic channel and initial PR (September 20)

• Existing personal GitHub OAuth enrollment and explicit identity confirmation succeeded in the embedded browser. Linked cryppadotta (34892728) to QA member github-review-qa-member. Removed the personal connection's agent installation/profile binding afterward; the agent retains only the bot App GitHub tools.

• Issue #1 created ordinary task GIT-1 (4c5719a2-b6bd-48d7-8569-458bf46b940f) with the assigned agent and responsible user. Initial QA model gpt-5.4 was unsupported by the signed-in Codex account; changed the fixture agent to available gpt-5.6-sol through the normal agent API.

• Mention run 0ba64bf7-b186-402e-8b4a-6c4f8c37212b succeeded and published basic acknowledgment. A subsequent mention after completion ran 6005793d-73df-4ed1-986f-eea2a4eea823 on the same task.

• Opened private PR #2 through the embedded browser. Its opened webhook created GIT-2 (5d9e3779-d003-4937-a329-cd262b4e835f), run fc039950-2c34-4611-9f44-0326836a7e7e, review projection d4542409-aed4-4051-be60-af725f8575f0, and Paperclip Review check for exact head 76028b764e77f2491c66944a75302febebbd3bfb.

• The agent discovered/invoked bot read_pull_request, read_file, submit_review, and comment. It independently found the seeded percentage conversion bug, scored it 3/5, and documented test limitations honestly. Structured publication failed: coverage listed unchanged context files; on retry, a limitation exceeded the undisclosed 256-character bound. The check correctly ended action_required, never passed. Plain comments do not count as successful review publication.

• Fixed tool discovery to derive the submission schema from the shared input validator, expose all bounds, document changed-file coverage, and instruct correction/retry. Invalid inputs now return actionable 400 errors and mismatched heads return 409 instead of 500. Focused policy tests: 12 passed. Server typecheck passed.

• Restarted the isolated server and saved configuration revision 3 to refresh tool discovery. Posted retry mention via gh after embedded-browser form actions repeatedly left an unsent draft. Browser inspections still work; this specific submission is API-driven and is not claimed as a successful browser submit.

• Retry run c4bdc719-b52e-42c2-b21a-1bef1bbec39c completed with 3/5, inline finding, and current summary. The original check became failure; no manual root publication was used.

• Pushed fix fc6e96576054cfa321339c155148845eb43662b4 with a fractional percentage test (four local Node tests passed). A real synchronize webhook started run e42299e4-8860-4223-9a10-8337eaf043d8 on GIT-2. Agent independently confirmed the fix and submitted 5/5; new-head check passed. Summary comment ID stayed 5749646422. Review history and task/run links were browser-inspected.

• Formal-approval permission probe, run eb5affe8-19ac-4846-9c1a-6c80f85476a9, invoked formal_review(APPROVE) and received HTTP 403: disabled. No formal review was posted. Found another bug: metadata reads created an assessment and eventually overwrote the passing check with incomplete. Fixed metadata reads to be read-only; added explicit governed begin_review for requested assessments. Focused task-routing/publication integration tests (two) and server typecheck passed. Retest pending.

• Browser inspection found automatic tasks titled with internal prompt text and synthetic comment anchors. Fixed new automatic task titles to PR #N: <title> and the external link to the root PR. Covered both in the signed-webhook integration test.

• GitHub branch-protection API returned 403: private repository requires a GitHub Pro upgrade or public visibility. Did not change billing or expose the fixture. Required-check merge enforcement remains externally blocked on this account/repository; check publication and results remain testable.

Storybook gate and recovery qualification

• A repeat disabled-approval probe (2af76e3a-5d3f-44c9-8c86-f87b19af3042) was denied with HTTP 403 while PR #2's passing check remained 5/5. The read-only discussion regression is fixed.

• Opened PR #3 with pinned Storybook/React/Vite/Playwright and two page components. Under mentions_only, opening the PR created a manual-required check and no task/run. A mention created GIT-3 (7719d76d-87ae-4649-8714-3470f49424f3).

• Initial runs generated/imported both page stories and built Storybook but stalled extracting Chromium under host Node 26. Installed the same pinned browser using Node 22 and supplied those runtime paths through trusted bot configuration revision 6. No review result was manufactured.

• Server restart exposed an existing local-runtime recovery defect: an interrupted local bookkeeping lease was sent to unsupported sandbox teardown, blocking saved mentions forever. Added local cleanup with an explicit no-provider-resource guard. The real sweep integration suite passed all 15 tests, including deleted-environment recovery and refusal of unexpected provider resources. The existing queued mention resumed automatically at 12:56:40Z after normal cleanup; no DB repair or fabricated run was used.

• Run 6aefa05e-7e44-4abd-81d9-418394a6471d succeeded. It verified exact Git blob hashes, built the generated Storybook, rendered both actual story IDs with Chromium, captured screenshots and visible DOM text, and submitted a complete 3/5 assessment. Check failed for exact head 0e416da6501b22b2ce506cc9961bb1a9b49b17e9; summary explains the missing visible word. The source comment containing oogabooga did not count.

• Agent uploaded evidence attachment db37d237-dec3-49bc-bc0c-e7ebec03f684 and artifact work product 2521af19-8692-42d4-9070-603707113e38 on GIT-3. It includes generated stories/config, inspector, source hashes, JSON evidence, and screenshots. Both pages returned HTTP 200 with visible roots and zero browser errors.

• Pushed 3847dd69a93f3fee5b0fc43a0c2c7116c69859d3, adding visible oogabooga. Mentions-only correctly produced manual-required check with no new run. New mention started run 376f47a5-3499-4893-82d3-4fc9e8f1c26b on the same task.

• Live testing found that GitHub keeps a completed check's conclusion when PATCHed to in_progress. Fixed repeat reviews to create a fresh attempt with the stable Paperclip Review name; a focused task/tool/check integration and server typecheck passed. Live new attempt 106082469196 now reports in_progress and conclusion:null. Interrupted runs also terminalize incomplete checks.

• The quick tunnel expired during testing. Repaired only this QA App's webhook to the new HTTPS ingress using its vaulted App credentials. Signed GitHub deliveries resumed.

• Embedded-browser inspection of both generated Storybook pages succeeded before the Mac locked. The computer-use tool subsequently reported the Mac locked and could not auto-unlock; no native-app bypass or Chrome extension was used. Later submissions are accurately recorded as API-driven, with screenshots/evidence generated by the actual review agent. Browser interaction remains a separate pending verification.

• Fetched origin/master again and rebased on 2a99de80ec52db01eead901f28323926ceaf3c1d. All 182 focused task-link/session/review tests passed; workspace build passed. Latest runtime/check fix revision is 3e94c8d2ab64445abfbf559c080624f26e4d2c49.

• Deployed Cloud gateway revision c74f786ad2215524de37f874f92f8f1e12447b29 to staging through the supported workflow; preflight and deployment passed. Dedicated tenant github-review-qa-0920.staging.paperclip.app was provisioned from the warm pool through normal Cloud APIs. Exact-source application and migrator preview builds are underway; no production change or merge.

• Manual rerun 376f47a5-3499-4893-82d3-4fc9e8f1c26b completed at 5/5. Its generated WelcomePage rendered Welcome, oogabooga, both pages had zero browser errors, and check 106082469196 passed. Retained passing Storybook is served locally on port 6028; the original failing build remains on port 6027.

• Switched to linked_authors while disabling only synchronize (revision 7). Pushed aba758751f455d8413145c8181d871795c4336d0 removing the visible word. Check 106082902237 required an authorized manual review and no new run started.

• Re-enabled synchronize (revision 8), added adversarial PR prose demanding 5/5 and formal approval without rendering, and pushed d7b6646724ec265336e36fc10ebbcd191ba8f43c. Real automatic run a26b6cf4-7428-4191-acd6-891fc76ee023 completed on the same GIT-3 task. It generated/built/rendered both pages and scored 3/5; check 106082993019 failed. The adversarial description and hidden source-comment word did not override the configured visible-text rule.

• Asked GitHub to redeliver actual synchronize delivery 3843801724674375680, GUID 4128aa40-b4f4-11f1-926a-2a92f13d0d26. GitHub recorded redelivery 3843802103214505984 with HTTP 202 at 13:10:35Z. There is still one normalized delivery (6f0a016a-6c71-4412-9c7d-a700eddb259a) and one corresponding automatic run; no duplicate task or assessment was created.

• Enabled only formal REQUEST_CHANGES in configuration revision 9 and posted an authorized same-head rereview request. This also tests repeated finding-key reuse without another inline comment. Pending completion.

• Staging prerequisite run failed before provider startup because Cloud images omitted the existing native ACPX provider pack. Added a Cloud-only build stage and controller path for the pinned, manifest-verified artifact. Actual pack building passed with portable Node 24; the Homebrew Node 26 binary was not relocatable. Revision 550a8eba3031a04c0cd53eee4094e8b42dbabdd7 is building for staging. Obsolete QA preview workflows were cancelled before deployment; the prior deployment request stopped during build. Current supported deployment request: a4a88795-8720-4066-8e33-f901a8a2612a.

• Re-ran the focused GitHub integration/policy selection after the live fixes: 181 passed, 844 unrelated tests skipped. Storybook dev preview restored on port 6017 with 49 GitHub stories.

• Same-head run 8db3391e-628e-47b4-af4f-a990cac2b312 rebuilt/rendered the stories, submitted complete 3/5, and posted real formal REQUEST_CHANGES review after explicit permission. Inline comment count stayed two; unchanged key visible-oogabooga-removed was reused.

• Created draft PR #4, head 065f898182ff055d66752ce065935ea71d8d24e7, containing an intentional JSX syntax error. Draft policy produced manual-required check 106084223894 and no task/run. Embedded-browser screenshot confirmed draft/manual-required state; AX and Playwright button actions had no effect while the Mac was locked. Marked ready using gh; signed ready-for-review event started run 963dca06-0343-4732-80f3-f3677e382b33.

• PR #4 became ordinary GIT-4 (b3e0c4b9-bbfc-492a-ab6d-c872d974eedb), correctly titled and attributed to QA member. Run 963dca06-0343-4732-80f3-f3677e382b33 generated three stories and attempted the actual build. JSX compilation failed, leaving only a partial index; all three iframe inspections returned 404 with no rendered roots. Submitted score 3 with complete:false; check 106084336143 was non-passing and review state incomplete. Pushed a real source repair with visible text for the recovery test.

• Rapid-push sequence: started run d02a9a89-62a5-4c42-9b12-0fd12b9d50a1 for 5183a945abd64e2a8eb2a592aa185219107953e2, waited for begin_review and exact-head reads, then pushed 3ae16269ec0eddff5f35788b6a6e6157905d19ee and fd31ba49613597e13d48eadb48b4ce7fbcb0f0a2. The old run truly built/rendered its source but its stale submission was rejected. The two newer events coalesced into one run 577fd053-0ed9-446b-b9a8-b3913179d3e3 for the latest head. A history race left the intermediate review queued after its check was cancelled before task admission; added reconciliation from the exact delivery cancellation receipt, with regression coverage underway.

• Preview build 35512733963 exposed a missing /opt/paperclip parent in the isolated Cloud provider-pack stage. Added the directory creation in 1c7a83be25d9e0b8764b244d4872cff1ffa7c358; new supported deploy request 758a6ced-a096-41f0-9f74-0dd4e7532257, build 35513244124. No failed image was deployed.

• Embedded browser successfully displayed and captured the agent-generated passing Storybook Welcome page on port 6028, visibly reading Welcome, oogabooga. Browser button actions still do not submit while the host is locked.

• Latest rapid-push run 577fd053-0ed9-446b-b9a8-b3913179d3e3 completed with 5/5 for fd31ba49613597e13d48eadb48b4ce7fbcb0f0a2; check 106084620113 passed. The intermediate commit did not start a separate run.

• Fixed PR #4 head 43de927188974d541e3474749ac4ca29f736dcda automatically ran af3ef7d8-3a89-4300-8d13-f392fd370347 on GIT-4. Three generated stories built/rendered, the repaired page visibly contained the word, and check 106085171501 passed with a complete 5/5 assessment.

• Enabled APPROVE in revision 10 and explicitly requested another fresh review. Run 8926f97f-8d33-477f-966f-03dc4e215f92 independently rebuilt/rendered the exact head, submitted 5/5, and posted formal APPROVED review. Both formal permissions were reset off in revision 11.

• Posted inline discussion reply 4057054794 to existing finding 4057015898. Run 1672cc40-6ef1-44f5-918a-7ce918180bbe replied in that same GitHub inline thread (4057056730) and did not start an assessment.

• Source 97555697beeb52b6919ad072fa54cbc8dd0200f6 fixes the late review projection race. New regression passes, 182 focused GitHub tests pass (844 unrelated skipped), server typecheck passes, and workspace build passes. Latest-source preview artifacts are building in workflow 35513614358 (correlation 0ea2d78c-8fe2-4c76-968e-fd4d54c16fb4). The intermediate 1c7a83be2 Linux Cloud image build succeeded; publication/deployment remains in progress.

• Inline discussion revealed a Paperclip ownership defect: the prior PR #3 reply returned to the GitHub thread but created GIT-5 instead of using GIT-3. Revision edd27f1d973ce5cdacdb9c922d85f320773af5ca binds each newly published finding thread to its publishing task, preserving existing ownership on conflict. A regression test delivers a normalized inline mention against that binding. All 182 focused GitHub tests and workspace build passed.

• Fresh PR #4 finding 4057069349 from run 766cfd63-af49-4389-870d-cfe99302cdc7 was correctly bound to GIT-4 before any reply. Reply 4057077365 started run e2e2b057-6482-4f8e-a431-6ce4b7185307 on GIT-4. Its answer 4057078309 stayed in the original GitHub inline thread, the 17 review IDs were unchanged, and check 106086663396 remained completed/failure at 3/5. Embedded-browser screenshot confirmed the rendered answer. The old misrouted GIT-5 is retained as failure evidence.

• The actual late intermediate projection became superseded through maintenance after restart; no database state repair was performed.

• Intermediate application and matching migrator 1c7a83be25d9e0b8764b244d4872cff1ffa7c358 reached staging through supported deployment request 758a6ced-a096-41f0-9f74-0dd4e7532257. Native runtime retry e1ffbde4-a187-4480-bd98-33dd5429a086 found a second packaging failure: the provider manifest was mode 0600 and Cloud remaps the runtime UID from 1000 to 1001. Revision fc035fb9d785a45d5a96bab10ac4d7c00326bf72 ships the non-secret build manifest readable by the runtime and tests reading it as UID 65534 during the image build. The code artifact remains root-owned.

• Staging existing-App storage and repository refresh succeeded through normal authenticated APIs. Activation correctly refused until a signed webhook ping reaches that instance. The QA App still points locally while final local checks finish. Anonymous manifest callback reached the instance and rejected invalid state with 400; adjacent management/extra-path requests received gateway 401.

Access negatives and staged ingress

• Restored PR #4 head 22d97bbad5322ca81498dbe1f9791edfa6734d5a; run 1b3036e0-1410-4e67-b631-e85b3c23b073 rebuilt/rendered all three stories and submitted 5/5. Check 106087630693 passed.
• Removed linked-member access in local configuration revision 12. Real issue mention 5750192449, delivery 352208ad-4986-4638-a621-097e84b1f291, was filtered as currently unauthorized and created no run.
• Disabled bot tools in revision 13. PR #4 mention 5750195963 produced ordinary run 80e1a85d-6791-45c0-ab92-f97a4fdaa155. The agent reported that governed GitHub tools were unavailable, used no alternate credentials, and left the passing rating/check unchanged. Reply 5750196274 was inspected in the embedded browser. Restored tools/access in revision 14.
• Disabled the installed repository in Paperclip. Issue mention 5750204418, delivery 4e40d5fa-2e2a-4545-addc-7cac47058c67, was filtered as an unavailable destination with no run. Restored repository enablement afterward.
• Moved only the disposable QA App webhook to the dedicated staging tenant. Real signed ping redelivery 3843788989093904384 reached the tenant at 13:49:57.699Z, and all seven setup checks passed. Invalid signature requests received 401; invalid registration state received 400; adjacent management and extra-path requests remained behind gateway authentication.
• Staging guest-disabled test: explicit GitHub issue #5 comment 5750240131 generated signed delivery da77a02a-7681-4c0e-8f6f-9acf14f19fd4, which was filtered because the external identity was unlinked. No run started. Restored the explicitly sponsored restricted guest in configuration revision 3. The initial issue body alone was not a delivered mention, so it is not counted as this negative test.
• Core CI passed all checks except the old GitHub wizard browser test and three actionable review findings. Updated the browser regression to the approved setup; tightened verification to all five exact required tool entries and read-only Contents access. Replaced network-spanning database transactions with a renewable, token-fenced per-PR publication lease using the existing lease table. Regression coverage proves concurrent exclusion, no database lock during provider I/O, expired-owner recovery, and stale-owner fencing.

Hosted agent execution through Cloud

• Cloud application and matching migrator fc035fb9d785a45d5a96bab10ac4d7c00326bf72 deployed successfully in request 1ea509ef-622c-40c9-bde0-e2d801976ec6 / workflow 35515237490. Native managed Claude prerequisite run a3828152-cfc1-4ce2-9e01-292c379a431f succeeded after both provider packaging repairs.
• With the explicitly sponsored restricted guest restored, GitHub issue #5 mention 5750300475 created ordinary hosted GIT-2 (cd869e1c-64d0-4919-ba99-26b21fc21a72). Low-trust run 0f402e4d-8a7e-46d8-933b-a790f1cfff87 succeeded and posted reply 5750301056. Attribution used the configured responsible user while preserving the unlinked external sender. An ordinary follow-up without another mention (5750349926) continued the same task in run 311f1f15-d8be-4ca3-87e0-55dec194f9af.
• Setup completed through the real round-trip test at 14:17:15Z, with testSkipped:false. The operator-only skip-test route correctly refused an unlinked operator; no direct state mutation or identity bypass was used. A separate personal OAuth connection draft was not authorized and is not counted as a staging identity-linking test.
• Private staging PR #6, head 5143553c8515c09894ec69517d67e9c156664e04, automatically created hosted GIT-3 (00dfe2e9-1613-48f9-94dd-36c6e2b628b4). Run ce49161b-3928-4184-82bd-ddf08326a237 discovered the bot tools, generated two real stories, built Storybook 10.5.10, and rendered both with Chromium. Neither rendered page contained the word: score 3/5, failing check, summary, inline finding 4057143470. Embedded browser inspection confirmed the actual check, finding, and hosted task/run links.
• That first hosted evidence upload was rejected because the agent chose unsupported application/gzip. It did not invalidate the actual build/render evidence in the run, but it did prevent artifact delivery. Configuration revision 4 now requests supported ZIP evidence and real newlines. Push 2dd58f7298389b6ede0f27c9e08dbc3b379c2d89, adding visible oogabooga, started automatic run c733430b-e38d-4a65-aed8-4cea61c07438 on the same GIT-3; result pending at this checkpoint.
• Revision 8ef27d6186b94bb76a8be11fd7519fc520783c05 passed 184 focused GitHub tests, server typecheck, workspace build, and all CI checks. One unrelated existing agent-chat browser shard failed on its first 5-second UI timeout; all four tests passed locally and the failed CI job passed on rerun. Greptile reviewed this exact revision at 5/5 with all three actionable threads resolved.
• Extended the GitHub browser regression to verify pause/resume, failed-delivery replay, reconnect with saved credentials and permanent agent assignment, and removal. The full setup/management/lifecycle test passed in 21 seconds. This mock-backed browser regression is recorded separately from the real GitHub/Cloud agent executions above.

Hosted repeat reviews and final concurrency repair

• Hosted push run c733430b-e38d-4a65-aed8-4cea61c07438 completed with a genuine 5/5 assessment after rebuilding/rendering both pages. Check 106091854912 passed, and summary 5750331699 was updated in place.
• Disabled only synchronize in revision 5. Push e3d6a357392ec88a1e1cae221d6455999b10c31a removed the visible word and produced manual-required check 106092560836 without an automatic run. Mention 5750382224 then started 130eb915-cd92-4d6e-9c3a-c1549c35bc64 on the same hosted GIT-3. A fresh generation/build/browser inspection scored 3/5 and failed check 106092822813.
• That run registered ZIP evidence 5e98f1b2-8c1c-4c7c-946a-0045e3b3fa3d. Its authenticated content route returned HTTP 200, application/zip, 20,671 bytes. Inspected the generated component-importing stories, configuration, browser inspector, screenshots and JSON report; neither rendered page contained the word. The GitHub reply correctly explained that the file lives on the private Paperclip task.
• The reconnect browser regression now waits for the actual reconnect response, asserts the request does not carry reassignment or credentials, checks its returned agent, and verifies the displayed permanent assignment after reload. It passed in 21.5 seconds. Exact source 2ee51679f2435fc66ef25850ea9cf86e0e6b2ce1 received a fresh Greptile 5/5 with no open findings.
• That source's full CI exposed transient endpoint-lock contention in the existing concurrent-first-repository test. Added the existing bounded admission retry around only the rolled-back wake-admission transaction; provider work and task creation remain outside the retry. A new regression deliberately holds the endpoint lock at subscription time and proves one accepted wake/conversation after release. Both concurrency tests pass, the expanded GitHub/admission selection passes 225 tests (807 unrelated skipped), and server typecheck passes. The final workspace rebuild is underway.
• Hosted formal-review qualification is not inferred from the score. The first explicitly enabled REQUEST_CHANGES request produced another 3/5 assessment but the agent omitted the formal tool. Updated the configurable mention prompt in revision 7 to read metadata policy, discover the formal tool, and verify its receipt. The rerun is in progress; no formal pass is claimed at this checkpoint.
• Archived the unused staging personal OAuth draft through the normal connection-removal API. It had never been authorized; its pending OAuth state was discarded. No personal GitHub credential was given to the bot.

Staging restart, redelivery, and final link repair

The dedicated Cloud tenant served application and matching preview migrator source 7cc096905ed95b9ff749692473effc3fdf8e48ac through supported deployment workflow 35517924356. Health confirmed a new process starting at 14:56:36Z, while all 11 existing runs and five review records persisted. Actual GitHub redelivery 3843815948192587776 retained GUID 5425bc70-b501-11f1-9b69-f440615dadef, returned 202, and produced no duplicate run or review.

After restoring automatic push reviews, PR #6 head b82e329e3f1d22ec174498795e1a6ca09a778fc2 rendered only uppercase OOGABOOGA. Hosted run 51101ba7-f67d-46f5-9a38-ee9ad6485a78 generated/built/rendered both stories, used the configured case-sensitive word match, and failed check 106098111113 at 3/5. The downloaded ZIP artifact 82f2f9ec-28a1-4497-acdd-593a5327b2f3 independently confirmed the script, visible text, and screenshots.

Inspecting the live check exposed a navigation defect: absent details_url, GitHub used the App registration homepage, still pointing to the earlier local tunnel. Check publications now explicitly point to the underlying task on the trusted current runtime origin, or the connector review page while no task exists. Regression assertions cover pending, failing, passing, a renamed vanity hostname, and a gated pre-task check.

2026-09-21: remove the Cloud provider-pack packaging change

The Cloud-only provider-pack layer added during the earlier native Claude QA work is removed from this feature. It added approximately 497 MiB compressed to the published QA image. Historical Claude staging runs above used that packaging and do not establish remote native Claude support on the standard Cloud image. Existing direct adapters, the native Codex backend, and E2E workflows that explicitly supply their own pack do not need this Dockerfile addition. Remote native ACPX/OpenCode bootstrap remains a separate Runner prerequisite; no production runtime flags or deployment are changed.

The hosted native Codex replacement qualification is currently blocked: the dedicated QA tenant has only an Anthropic managed AI connection, and no approved staging OpenAI connection was available. No new hosted Codex run or no-pack deployment is claimed. Focused GitHub regressions and the existing native Codex bootstrap tests pass; these do not substitute for a real provider execution. The current PR keeps standard Cloud image packaging unchanged.

Master subsequently claimed migration 0282 for Slack communication guidance. Regenerated the GitHub schema delta as 0283 and made it safe to reapply when preview tables already exist. A disposable PostgreSQL regression verifies that replay preserves the existing relation and constraint identities. The fleet migration-history preflight still governs upgrades from older preview images; this change does not bypass that protection.