Current CI status: #3910 is merged. #3911 at 87e8e211 has all 19 checks completed: 17 SUCCESS, one SKIPPED and one NEUTRAL. Coverage, Integration, repository guards and Android/Apple/Linux smoke succeeded. See B2-87e8-ci-final.json. Local validation and live limits below remain unchanged. #3325 awaits final merge and acceptance/admin verification.
Current B1a head: b920c2333b03031d3221a120b1040a6d9de9f8e2, 33 files and 1923 gross lines from base 5f6dfb3e6f0992958cdaa0081fcfde3ada24c0cc. The only change from published a6fd is .github/workflows/linux.yml (+7/-4). It adds UTC update/install start and completion markers and ordinary apt output, while preserving the six-minute bound and failure exits. It is a diagnostic change, not a proven repair of the unknown apt stall.
The new B1a-b920-* files contain the full exact clean run-1vTiRz receipt (834 related files/6474 tests), full independent Sol and Claude Sonnet review results with their limits, stubbed apt shell checks, and the read-only recurring-failure assessment. The receipt manifest records original and screened SHA-256 separately. No real apt or Linux smoke ran locally. The earlier 56c0 and a6fd setup failures remain failed historical CI; the one successful 56c0 targeted retry applies only to 56c0. Final b920 CI completed: 17 SUCCESS, one SKIPPED and one NEUTRAL. Coverage, integration, repository guards and Android/Apple/Linux smoke succeeded. Linux APT setup and both smoke evidence steps executed successfully. See B1a-b920-ci-final.json. This does not identify or repair the earlier unknown stall cause.
The B1a production tree matches reviewed and Android-verified a6fd byte for byte; the earlier gate, reviews, and live proof keep a6fd attribution. B2 #3911 remains separately owned and needs exact b920 dependency integration and its own validation. #3325 remains open.
This evidence supplements the merged request-deadline core and shared ownership preparation. It does not claim that #3325 is closed or that either guard PR has merged.
Published PRs: #3910 โ operations and nested tests and #3911 โ lifecycle and claim recovery, based on #3910. B1a b920 CI is complete; corrected B2 validation and publication remain pending. No merge or full issue acceptance is claimed.
The original 56c0 and B2 artifacts below retain their exact historical SHA attribution. Current B1a corrective candidate: a6fd20a7577c4475bf64730db5e8e6dbaa4914a9 on base 5f6dfb3e6f0992958cdaa0081fcfde3ada24c0cc, 32 files and 1912 gross lines. The retry fix stops a bounded nested resource refusal after the first attempt for both CLI and manifest retries. Ordinary resource-free failures still retry.
The new B1a-a6fd-* files contain the full exact clean run-zNMTPj receipt (834 related files/6474 tests), exact full Sol and Claude Sonnet review results with their limits, unchanged 56c0 negatives (three attempts/three prefixes), the historical failed c080 gate and artifact-dependent fixture ledger, and exact-SHA Android live evidence. The two live bounded routes each had one attempt and one journaled home; the ordinary failure had three attempts and three journaled prefixes. Journal counts do not establish a universal native syscall count. The a6fd reviews predate its later gate and live run; their stated limits are retained.
PR #3911's B2 integration and newly approved IME policy correction remain pending separate validation and publication. The older B2 artifacts below retain their own SHA and do not establish that later integration. #3325 remains open. CI for a6fd must be read on that remote head after publication; the earlier 56c0 CI result does not transfer.
| Layer | Base | Head | Diff |
|---|---|---|---|
| Operations and nested test | 5f6dfb3e6f0992958cdaa0081fcfde3ada24c0cc |
56c0f4775ab67b5242f45172f23a534b7bf816d5 |
25 files, +1735/-26 = 1761 gross |
| Lifecycle and claim recovery | 56c0f4775ab67b5242f45172f23a534b7bf816d5 |
37c681cf23c30833998e4dda52eeb0ac4f247b71 |
22 files, +2060/-96 = 2156 gross |
Merged core 1793 + merged preparation 2056 + operations 1761 + lifecycle/claims 2156 = 7766 gross across the delivered concerns. Historical additive preparation accounting is 50 lines higher because its correction commits overlap. Each remaining parent-child diff is below 2500 gross.
The user chose simplifying only the new resource-refusal guidance. A refused step retains typed reason, resource shape, public command name, dispatch status, completed progress, and replay source. Guidance asks the agent to run the original operation separately with its original options and the same session/private context. No executable CLI string is reconstructed.
Existing replay takeover remains: failed-step/source context, repair hints, resume.from, resume.alternateFrom, planDigest, and existing resume safeguards. After directly completing the refused operation, use alternateFrom when offered and a tail remains; from would retry the same bounded operation. Inspect completed steps before retrying.
Normal outer authentication, lease renewal, in-flight protection, locks, events and activity remain. The refused step performs no own admission, binding, cleanup or native dispatch. Stable claim previews refuse before effects; if another daemon changes a claim after preview, the actual recovery callback can run after facts/binding and prevents the recovery gateway. No cross-daemon atomicity or universal zero-bind/native-stop guarantee is claimed.
- B1a
run-ZsVDa1: exact clean pass, 831 related files / 6458 tests. All ten runnable checks passed; two checks were deduplicated and eight lanes skipped with their declared CI/manual owners. - B2
run-s7H6US: exact clean pass, 36 runnable checks passed, 25 skipped, two deduplicated. See its full sanitized receipt for command and test counts. - Both receipts name their exact base/head and
checkoutChanged:false. Sanitization changes receipt bytes;artifact-manifest.jsonrecords original and screened SHA-256 values separately. - B1a full independent Sol and Claude Sonnet reviews clear
56c0with no P0โP2 findings. Their source-only limits are preserved in the full review files. - B2 independent Sol and Claude Sonnet reviews clear exact
37c681with no new P0โP2 findings. Sol traced the full diff and current gate. Sonnet ran 140 focused tests and 853 eager-closure cases, and disclosed that it did not re-read every historical brief. Their complete results and limits are included; no aggregate mutation score is claimed because its narrative count is inconsistent. Remaining P3 validation-order/coverage notes have no demonstrated current resource mutation and were not used to start another code round. - Earlier failed and cancelled gates remain historical. In particular, B1a's earlier two timeouts were not explained by the later controlled pass. Unchanged base and candidate both passed the same 213-test provider suite; that comparison did not prove contention.
B1a's unchanged-main operation test fails without the guard. The old 481c test-only comparison detects three nested-hint failures and one typed shape/name redaction collision, with other tests passing. Corrected source passes them.
B2's unchanged 56c0 comparison detects the missing live-foreign-claim recovery refusal and provider-releasing close refusal. Other copied-test failures caused by absent B2 fixture APIs are invalid comparisons and are not behavior evidence. Later guard mutations detect active-provider bypass and existing-session reopen behavior; their full dispositions accompany final B2 review evidence.
The former full executable-command reconstruction tests are obsolete under the explicit user decision. Their removal is a product-scope change, not a resource-policy test cut.
B1a exact 56c0 on an Android 15 emulator, source/helper 0.21.31-dev, proved:
- batch
homethen refusedlogs start, truthful prior progress, no post-refusal capture, session/IME retained, and successful direct start/stop in the same context; - replay
home/logs start/home, source-line refusal, same-session direct repair, digest-guardedalternateFrom, no repeated prefix and tail once; - an unbound outer batch's nested test opened the app, refused perf memory snapshot, kept original-test guidance, restored the IME and retired its attempt. A separate ordinary direct test passed.
The first nested run correctly refused at open because its outer session already held the device. That setup result is retained, not misreported as resource proof. Extra durable attempt-resource cleanup and direct native perf parity remain unit-only.
B2 exact 37c681 live verification proved owned-IME caller close refusal/direct restoration, app-log close/reopen refusal, byte-preserving dead-owner claim refusal/direct resource recovery and snapshot, and Maestro app-only close. See B2-android-live.md and selected response/proof JSON. Physical devices, all backend resource families, provider cloud release/expiry, live-reboot races, universal native-stop and no-overlap guarantees are not claimed. All B1a-owned sessions, claims, daemons and IME changes were cleaned; the borrowed emulator remained booted and its panel was detached without shutdown.
Only allowlisted receipts, requirements, review results and selected evidence are published. Credentials, host configuration, environment files, daemon registration JSON and private wrappers are excluded. Home, host and private artifact paths are redacted. Original evidence remains private with its hashes; the public files are screened copies. The requirements file preserves a historical implementation-plan snapshot; exact current heads and evidence above take precedence over historical checkpoint lines.
Current B2 head is 87e8e2113938b9a324fbdd1fca3ed57d374c069b on merged-main b08b5e796a2e74d4c8cd545a051b304cdf9d03f8: 35 files, +2341/-157 = 2498 gross. The earlier 37c681 and 907bd artifacts above retain their historical SHA attribution. Seven B2 patches range-diff = through the main integration. The exact clean run-wOpmey receipt passed with 36 runnable checks, two deduplications, 25 declared CI/manual skips and 4522 related tests. Current Sol and Claude Sonnet R10 reviews reported no P0โP2 findings; the historical full-context R8 review and its attribution are separate. See B2-87e8-* files and artifact-manifest.json for original and screened hashes.
The user approved ordinary bounded close when Android test IME is the only teardown resource. Bounded batch and authored replay open/home/close restored the original keyboard and retired the owned session/claim on the exact source. Maestro app-only stop retained the IME; a later bounded close restored it. App-log plus IME still refused bounded close with typed app-log reason, no refused-step effects and truthful prior progress. Ordinary reopen, app switch and relaunch preserved the resource manifest. Explicit iOS Settings open reached native snapshot and close; no-app mixed-inventory foreground returned AMBIGUOUS_MATCH, so that route is not claimed successful. An owned stopped Android AVD booted by name and shut down by its serial, then was deleted. The later same-session replay alternateFrom after direct close returned SESSION_NOT_FOUND; the record preserves that limit. Native failed-IME-restore remains routed/unit proof only. The exact source/helper was 0.21.32-dev; panel was 0.21.25. All owned sessions, claims, panels and devices were released; the borrowed Android emulator and iOS simulator stayed booted. Android daemon cleanup required a verified-own-PID graceful TERM fallback after stop --clean declined identity. See the full screened live record.
The first historical fa1f affected gate failed in provider timing with cause unproven; an unchanged rerun passed. Neither result transfers to 87e8. Current PR CI must be read at remote 87e8 after publication, not from old 37c681 green checks. #3325 remains open until the remaining PR is accepted and merged; no issue-level completion or universal native deadline guarantee is claimed.
The earlier iOS open/foreground and stopped-AVD boot runs were direct commands. New B2-87e8-bounded-preview-* files record three actual single-step bounded batch requests at unchanged source 87e8e2113938b9a324fbdd1fca3ed57d374c069b. The installed batch grammar is { "command": "โฆ", "input": { โฆ } }: (a) fresh iOS Settings input.app, (b) no-app input.foreground:true with no UDID, and (c) stopped owned AVD input.device plus input.headless:true. The reviewerโs illustrative positionals/flags fields are not valid for this CLI.
Case A selected the same owned iOS device as a standalone open control and returned a 19-node native snapshot after the batch. Case B ran against a private allowed inventory of one booted iOS simulator plus Android and returned a 19-node native snapshot with single-booted-local selection. Its first setup attempt returned native AMBIGUOUS_MATCH before any step; after rebooting only the disposable simulator and launching Settings as its sole installed running app, the actual bounded request succeeded. Case C booted a new owned stopped AVD by name to its matching running serial in one bounded step. The preparatory input.target parser rejection is retained in the live record.
The five complete screened JSON files and the live report preserve response shape, original raw-result hashes, source SHA, request context with private values replaced, and cleanup evidence. B2-87e8-bounded-preview-screening.md records original versus screened hashes; artifact-manifest.json includes all new files without changing any older evidence file. Both disposable devices were deleted, all owned sessions/claims/daemons/panels were released, and borrowed devices retained their boot states. Earlier direct-run limits remain historically attributed; generic same-session alternate resume after direct close and native failed-IME-restore coverage remain limits. #3325 stays open pending maintainer acceptance and merge.