Open-source AI browser agent for Chrome and Firefox (monorepo) ðŸ§
Native chat durable resume verification
An ongoing conversation must be able to wait through a browser alarm and worker restart without sending an already delivered reply again. A running resume job now creates a separate future successor. The scheduler records the actual running predecessor privately; repeated future jobs with the same intent still coalesce. That predecessor relationship also survives storage restoration.
For supported native Telegram conversations, the chat kernel stores a bounded delivery receipt only after an independently observed outgoing message matches the guarded recipient, exact authored body, and a new native message ID. The receipt binds the genuine user turn, exact original task digest, canonical peer, page URL, originating run/action, message ID, and body digest. It stores no transcript or full message body. A new user turn with identical wording cannot inherit it.
Native Telegram sends and delivery acknowledgement require the verified latest viewport, not merely the bottom of an older virtual slice. A shared mounted_native_id_floor validates all mounted IDs and peers and retains the highest 200 numeric IDs, including empty and media-only rows. The private floor does not claim full-history completeness and cannot be lowered by the authored-message cache. The one-use dispatch token binds that baseline; structured chat_send also refuses any difference between its original observation and the actual recipient probe before pending state is created or persisted. Incomplete, changed or wrong-thread evidence requires fresh observation, without rebasing the send.
A same-task alarm can finish from that receipt only after reading the current native conversation again. The same message must still exist with the same authored body and peer, the composer must be empty, and no pending outbound message, OTP pause, unrelated current page action, or verification debt may remain. A provider, task, recipient, or page change during the read invalidates it. Internal terminal reads never consume incoming deltas: a newly arrived question blocks completion until the model explicitly observes it. The normal done bookkeeping step preserves its own freshly verified proof; a later tool invalidates that sequence. This does not add a fictional action to the resumed run's execution counter.
An unanswered permission or ordinary clarification is recovered from its live Agent entry when the panel reopens, even after replay acknowledgement. The descriptor remains bound to the current request, run, task, provider and steering revision; settled entries, Stop, replacement tasks and worker loss cannot revive it. When a run ends (finished, failed or stopped), the panel closes the cards it still showed: a permission card is removed and a question is locked, so a card left on screen after Stop can no longer be clicked; a scheduled task's cards stay with that task. Hydration rechecks current state, preserves the original deadline and never answers or renews the question. Accepted mid-run corrections wake obsolete permission waits without recording a grant or dispatching the earlier action, then continue through the normal steering boundary. The paired owned-fixture suites contain 54 hydration and 22 permission-steering checks; this is UI and control-flow proof, not native delivery or external-service endurance proof.
The authenticated source22 native-panel reopen restored the same unanswered permission card, and a genuine public correction woke that wait without granting it; the old click returned no-dispatch. This phase passed, but the whole 300-second task failed after passive page updates discarded two arithmetic-only plain replies before any calculator call and led to a new unrelated permission prompt. A separate source22 first-main correction calculated the exact result, then discarded five plain replies and stopped as page-unstable. These failed runs remain separate from the successful card-restoration phase (auth-telegram-source22-permission-steering-content-free-diagnosis.json and the retained native reopen/first-main artifacts).
Source24 recognizes a single genuine arithmetic-only task using the calculator's shared lexer and bounded parser. Exact token whitespace may vary; split numbers, extra inputs and mixed work cannot qualify. A pre-calculation empty-tool reply may survive passive DOM updates only to reach the normal actual-calculator nudge; it cannot create a receipt, evidence or success. A real calculation receipt can preserve its later plain reply, while explicit done, current private authority, Stop, hard feedback and all action-debt checks remain mandatory. The paired actual Agent chat/stream suite passes 124 checks and 950 neighbors; pure-calculation-before-tool-source-final.json records the owned unit/protocol qualification, not a live-provider or authenticated UI pass. Source24 complete coverage, 120-minute endurance and current authenticated retakes remain pending.
The genuine user turn's private webbrainResumeTaskId stays in serialized local state. The shared provider-message projection omits the reserved top-level webbrain* application namespace from Chat Completions and Responses requests. This also keeps applied steering IDs and private prior-dispatch receipts local. Ordinary roles, names, tool call IDs, multimodal contents, quoted user words, custom vendor fields, and provider-specific replay fields retain their existing handling. Both browser builds exercise actual applied steering and durable serialization at this boundary in test/provider-message-privacy.mjs.
The optional installed-MV3 fixture is test/quality/native-chat-soak.mjs. It uses real wall-clock time, real chrome.alarms, the normal background/Agent pipeline, actual storage, two genuine worker stops/restarts, and Stop while a model reply is pending. Only the provider and the native Telegram-shaped application are owned local fixtures. Network access outside the local provider and intercepted application is blocked. It checks exactly one reply per fresh numbered question, durable waits, bounded stored history/kernel size, no late send after Stop, and successful terminal completion after the final alarm. It does not prove authenticated Telegram/WhatsApp behavior, real-provider reply quality, or an unattended external conversation.
Run the optional test in PowerShell:
$env:WEBBRAIN_SOAK_MINUTES = '60'
$env:WEBBRAIN_QUALITY_ARTIFACTS = 'ci/artifacts/native-chat-soak-one-hour'
node test/quality/native-chat-soak.mjs
The three-minute validation completed in 189,297 ms with three questions answered exactly once, three scheduled waits, both genuine worker restarts, a verified Stop/continue boundary, and successful completion. Its maximum persisted session size was 93,208 bytes.
The earlier real one-hour run passed from 2026-10-07 19:34:16.178 UTC through 20:34:55.902 UTC: 3,639,724 ms, 59 actually arrived numbered questions answered exactly once, 59 scheduled waits, both genuine worker stops/restarts, and independently verified successful completion with no pending jobs. Stop settled in 1,133 ms; releasing the delayed model result produced no further reply before Continue. The maximum persisted session was 516,565 bytes with 541 messages; the bounded kernel reached 119 seen message IDs and 59 sent keys. There were 359 provider requests, and the largest request was 585,354 bytes. Evidence is ci/artifacts/quality-audit-2026-10-07/native-chat-soak-one-hour/summary.json. This archive predates the later editor, native contact, debugger and terminal-classification corrections.
That hour uses a configured 1,000,000-token context and distinct numbered reply bodies. It proves the wall-clock alarm/restart/Stop flow within its measured storage bounds; it does not prove finite-context compaction or an in-flight acknowledgement race between identical reply bodies. Those require separate regressions. The optional fixture also accepts WEBBRAIN_SOAK_CONTEXT_WINDOW, WEBBRAIN_SOAK_PROMPT_TIER, and WEBBRAIN_SOAK_MAX_OUTPUT_TOKENS: its finite-context mode reports a messages-plus-tool-schemas character estimate as usage, reserves the actual requested output, checks exact original-task preservation, and requires a witnessed history compaction. This estimate is not a vendor tokenizer or a claim that a vendor rejected the request.
The earlier snapshot11 hour passed from 2026-10-08 02:00:56.023 UTC through 03:01:38.465 UTC: 3,642,442 ms, or 60 minutes 42.442 seconds. All 59 fresh numbered questions received exactly one reply, with 59 actual scheduled waits and genuine worker heap loss at approximately minutes 10 and 40. Stop settled in 1,142 ms; releasing the delayed model result caused no late reply, and Continue resumed the same task once. Final live and independently read raw stored journals both report completed, successfulDone: true, and runSucceeded: true, with matching request IDs, running: false, no pending outbound and zero active jobs. The loaded extension and native fixture retained all 602 source hashes unchanged; no endpoint or product-source patch was applied. The immutable archive is ci/artifacts/quality-audit-2026-10-07/native-chat-soak-60min-final-source-snapshot11/summary.json with its adjacent source-verification.json. This archive predates the follow-up editor and monitor corrections.
That earlier hour retains the 1,000,000-token Full configuration and 4,096 requested output tokens. Its measured maximums are 527,460 persisted bytes, 541 messages, 119 seen IDs and 59 sent keys; 359 provider requests reach a maximum request size of 586,237 bytes. It uses an owned local provider and intercepted Telegram-shaped application, sends no external messages and creates no accounts. It proves the measured one-hour installed-browser lifecycle, not hours of authenticated Telegram/WhatsApp service use or finite-context compaction. The separate 64K proof below covers actual history compaction. The 01:28:52 UTC hour attempt remains an interrupted earlier-source run; it was stopped before the foreground terminal and defensive scheduled-callback corrections and is not counted as a passing hour.
The follow-up snapshot14 hour passed from 2026-10-08 14:13:19.441 UTC through 15:13:58.914 UTC in 3,639,473 ms. It answered 59 fresh questions exactly once through 59 scheduled waits, two actual worker reboots, Stop/Continue and 21 recorded context compactions with a 65,536-token Full fixture budget. Stop settled in 1,154 ms with zero late sends. Both live and independently read stored journals finish completed with both success flags, matching request IDs, no pending outbound and zero active jobs. Source verification passed at 15:13:59.504 UTC for all 602 unchanged product/fixture inputs. Evidence is ci/artifacts/quality-audit-2026-10-07/native-chat-soak-60min-final-source-snapshot14/summary.json and its adjacent source-verification.json. Its measured maxima are 359,106 stored bytes, 155 messages and 185,304 request bytes. The provider and Telegram-shaped application remain owned local fixtures; the context estimate is not a vendor tokenizer.
A separate genuine 120-minute snapshot15 run began at 14:35:15.490 UTC and failed terminal completion at 16:35:36.911 UTC after 7,223,106 ms. It delivered 117/117 replies exactly once, passed both real worker restarts and Stop/Continue, and retained all 602 source hashes. Its unchanged failed archive is ci/artifacts/quality-audit-2026-10-07/native-chat-soak-120min-final-source-snapshot15. Beyond 200 mounted native rows, the private send-floor baseline was incorrectly marked incomplete, preventing the latest verified delivery from installing a completion receipt. The unchanged installed three-minute native-chat-seeded-117-before-red reproduction confirms the defect in 192,810 ms, including three exact replies, both restarts and Stop/Continue. A later source19 retake was intentionally stopped after 25 minutes with 19/25 questions answered exactly once when a separate virtual-history false acknowledgement was reproduced. Its preserved native-chat-soak-120min-final-source-snapshot19 archive is not a terminal pass.
The final native latest-position, body-independent shared floor and observe-to-probe refusal checks pass 492 tests; proof is ci/artifacts/quality-audit-2026-10-07/telegram-native-latest-final-proof.json. Cold read-only preparation now adopts its private owner only inside a current active capture and preserves its exact reference and one-use token through the actual click handler. Its installed case passes in 5,917 ms with 600 source files unchanged, one read-only native click and zero message mutations; the task honestly ends partial with both success flags false because its alias remains unbound (telegram-cold-preparation-preprobe-final/evidence.json). The five installed native chat paths pass with 26 exact replies, ten completion judges, successful terminal completion and 602 unchanged product/fixture inputs (native-chat-preprobe-final/summary.json). These are local provider/native-application fixtures, not authenticated user-service endurance proofs. The source20 120-minute run was deliberately interrupted after 23 minutes 54 seconds and 23 exact question/reply pairs, with all 602 inputs stable, before the generic question-restoration changes. It is not a terminal pass. Full coverage attempt21 was also interrupted before those changes. The source22 120-minute retake was deliberately interrupted after 16 minutes 41 seconds and 16 exact pairs with 602 stable inputs, before the arithmetic response corrections; coverage attempt23 was interrupted at the same boundary. Complete source24 coverage and endurance remain pending.
Full default coverage attempt17 ended in a fake-provider fixture error: its phase advanced on a generated read before the runtime consumed it. The Stop case preserved one real send, durable pending state and false success flags. The fixture now requires the exact current-phase observation tool-call ID, with at most three read-only retries of a discarded proposal; the full 15-case c8 retake passed in 230,907 ms with 601 stable source/fixture inputs. Full default attempt18 passed with exit0 from 15:47:11.639 through 16:06:23.954 UTC, including 2,447/2,447 main tests and 60/60 security checks. Its 17,526 captured hashes matched at qualification; independent evidence is ci/artifacts/quality-audit-2026-10-07/coverage-followup-attempt18-summary.json. The later planner and native-chat product corrections require fresh complete coverage. Neither passing coverage nor an earlier hour substitutes for the failed two-hour result.
The separate installed same_body_race case sends ok twice for two different incoming questions. Question two arrives before the first send's outgoing bubble appears after a 900 ms delay. Both deliveries retain their original incoming anchor through persisted state, complete the two ledger rows, and finish successfully. test/chat-reply-dedup-race.mjs also covers restart, the 200-entry cache bound, old native IDs, wrong-peer ambiguity, and edited body rejection. Evidence is ci/artifacts/quality-audit-2026-10-07/native-chat-mv3/same_body_race-evidence.json and same-body-race-persisted-bindings.json.
Normal context management preserves the latest complete assistant/tool-result group while summarizing earlier groups. A native observation may already have consumed its incoming delta in the durable kernel; summarizing that unseen result would lose the question before the next model call. When fixed schemas plus that live group exceed the proactive soft threshold but fit the hard window with the requested output reserved, compaction uses that remaining headroom. An unshrinkable live native observation remains intact and is reported explicitly. test/context-live-tool-group.mjs covers both browser builds, original task identity, exact incoming delta, complete multi-tool grouping, and the output reserve. A 32K full-tier fixture exceeded the conservative request estimate even without meaningful history; changing capabilities or claiming vendor tokenizer compatibility is outside this result.
Before each ordinary main model request, both Agent paths pass the exact already assembled tool list into context management. Without a provider usage reading, the estimate includes these schemas once, including enabled WebMCP and custom skill definitions. A tool-free request contributes zero schema overhead. Once provider-reported prompt usage exists, it already includes the schemas and the existing reported-usage-plus-message-growth projection adds no second schema count. A fixed prompt/schema/authority estimate that cannot fit with requested output reserved is reported as fixed_prompt_above_context_budget; it does not delete the genuine task, downgrade capabilities, or hard-reject a provider request using the character estimate. The paired test file now has 18 cases, including first request, restored history without usage, no double count, tool-free requests and unchanged authority at a too-small fixed budget.
Context size and the selected tool tier remain separate configuration requirements. In the bounded 32,768-token fixture, Full reached an estimated 41,453 prompt tokens before a browser action; Mid later reached 32,663 prompt tokens plus 512 requested output tokens. Those numbers use serialized characters divided by four, not a vendor tokenizer. The current exported Full system prompt and normal tools alone estimate 30,277 tokens, while Mid estimates 21,466; the normal 4,096-token output request adds to that requirement before user, planner or site context. Router providers default to Full and local providers default to Mid even when the configured window is smaller. History compaction cannot remove this fixed cost. Compact omits native chat and durable scheduling, so silently choosing it would remove the capabilities exercised by this chat test. The passing 64K full-tier run below establishes its measured configuration; it does not establish a universal 16K or 32K minimum.
The separate finite-context installed case passed in 188,851 ms with a 64,000-token full-tier configuration and 512 requested output tokens. Before the first genuine worker restart, the owned fixture seeded 64 complete synthetic prior readonly tool groups into the actual persisted conversation, preserving the exact genuine task/task ID, pending state, and delivery receipt. Normal worker restoration and context management summarized 108, 10, 4, and 8 older messages in four compactions. All three fresh questions were answered exactly once; three real scheduled waits, both worker stops/restarts, Stop/Continue, and final verified completion passed. The largest estimated prompt was 47,645 tokens plus the 512-token output reserve, and the largest actual request was 191,138 bytes. The seeded 346,404-byte/140-message session fell to 93,634 bytes/36 messages at minute three. Evidence is ci/artifacts/quality-audit-2026-10-07/native-chat-soak-finite-64k-seeded/summary.json.
An installed retest after first-request schema accounting passed in 190,280 ms. It answered all three fresh questions exactly once, scheduled three real waits, restarted the worker twice with verified fresh globals, preserved Stop/Continue with a 1,138 ms Stop latency and no late send, and finished successfully. Five actual compactions summarized 108, 8, 4, 6 and 10 older messages, with one model summary request. The largest estimated prompt was 44,998 tokens plus the 512-token output reserve; the largest actual request was 180,563 bytes. The seeded 346,401-byte/140-message session measured 102,321 bytes/40 messages at minute three. Evidence is ci/artifacts/quality-audit-2026-10-07/native-chat-soak-finite-64k-final/summary.json. The earlier four-compaction proof remains preserved separately.
A later strengthened installed test exposed a separate UI bookkeeping defect: the final native done and alarm job succeeded, but the restored UI journal retained the earlier pause's false success flags. Scheduled updates now carry their original job/execution owner into the shared journal, and terminal flags are classified before storage captures the snapshot. Late cancelled callbacks, newer requests, same-request resumes, accepted steering and delayed reads of cleared storage cannot restore an old task's success. The paired lifecycle regression has 34 cases; this changes the UI record rather than delivery permission or native receipt evidence.
After that correction, the strengthened 64K full-tier test passed in 189,531 ms with three exact-once replies, five actual context compactions, both real worker restarts and a 1,156 ms Stop boundary. The current live journal and independently read raw session-storage journal both report completed, successfulDone: true, and runSucceeded: true, with matching request IDs, no pending outbound message and no active scheduled jobs. Evidence is ci/artifacts/quality-audit-2026-10-07/native-chat-soak-finite-64k-terminal-final/summary.json. The earlier one-hour proof predates this UI correction and does not establish these newly strengthened journal assertions.
The same strengthened finite-context case passed again after the native contact bootstrap correction in 188,897 ms, with 602 source hashes unchanged. It retains three exact-once replies, five actual compactions, both worker restarts and Stop/Continue with no late reply. Stop settled in 1,135 ms; the final live and raw stored journals both report successful completion with matching request IDs, no pending send and no active jobs. Evidence is ci/artifacts/quality-audit-2026-10-07/native-chat-soak-finite-64k-bootstrap-terminal-final/summary.json and its source-stability wrapper. This remains a local Telegram-shaped fixture with a 64K full-tier configuration and 512 requested output tokens.
After the later drawer, current-transcript and prepared-repair changes, the finite test passed in 189,346 ms with 602 source files unchanged. Three questions received exactly one reply each, five actual compactions preserved fresh observations, both worker restarts passed, and Stop settled in 1,141 ms without a late reply. Both live and raw saved journals finished successfully with matching request IDs, no pending outbound and no active jobs. Evidence is ci/artifacts/quality-audit-2026-10-07/native-chat-soak-finite-64k-prepared-terminal-final/summary.json and source-verification.json. The previously recorded one-hour run remains separately qualified to its earlier source.
A later paired callback regression exposed stale completion after accepted steering: an earlier successful done could leave the scheduled job successful when the corrected task failed, and foreground UI classification could retain an earlier failure after the corrected task succeeded. Accepted steering_applied now resets only the current terminal evidence and error classification; prior event history and uncertain side-effect checkpoints remain intact. Queued steering alone does not erase current evidence. The scheduled callback interleaving is defensive regression coverage, while foreground mid-run correction uses the reachable Agent/background/UI path. The paired lifecycle file now has 58 tests. The finite proof above predates this terminal-classification correction; current-source endurance and foreground integration evidence are recorded separately.
The current-source finite test after that correction passed in 189,010 ms, with 602 product/fixture files unchanged. All three questions were answered exactly once; five actual compactions, both native worker restarts and a 1,140 ms Stop boundary passed without a late send. Both the live and raw stored journals report successful completion with matching request IDs, no pending outbound and no active scheduled jobs. Evidence is ci/artifacts/quality-audit-2026-10-07/native-chat-soak-finite-64k-scheduler-steering-final/summary.json and source-verification.json. Its source snapshot includes the loaded Chrome extension and native fixture; later changes to two unrelated test-harness imports do not change this product-source evidence. This remains a three-minute local native conversation, not an authenticated service endurance claim.
The separate installed foreground producer test accepts a genuine public correction after the earlier task's actual done. A real provider error on the corrected task leaves both live and raw stored flags false with zero panel celebrations; a fresh successful control finishes with both flags true and one celebration. Evidence is ci/artifacts/quality-audit-2026-10-07/foreground-terminal-steering-real-producer-final/summary.json. Scheduled callback correction remains defensive coverage because public scheduled steering is disabled; the ordinary hour soak does not claim to exercise that accepted-correction interleaving.
To reproduce that optional finite-context case in a fresh PowerShell session:
$env:WEBBRAIN_SOAK_MINUTES = '3'
$env:WEBBRAIN_SOAK_CONTEXT_WINDOW = '64000'
$env:WEBBRAIN_SOAK_PROMPT_TIER = 'full'
$env:WEBBRAIN_SOAK_MAX_OUTPUT_TOKENS = '512'
$env:WEBBRAIN_SOAK_SEED_PRIOR_HISTORY = '1'
$env:WEBBRAIN_QUALITY_ARTIFACTS = 'ci/artifacts/native-chat-soak-finite-64k-seeded'
node test/quality/native-chat-soak.mjs
The audited working tree now follows 0f1b21a8ef9389612325353a00a62e634349b33c (40.0.8), with these QA fixes applied. The initial hour started before the version update and later acknowledgement-provenance, finite-context and general-contact fixes. Snapshot11 covers its recorded source; snapshot14 covers the later editor and monitor corrections. Their immutable archives remain separate.
The user has loaded the current unpacked 40.0.8 build in the existing Chrome session. The existing OpenRouter GPT-5.4-mini route was saved and selected with a 128,000-token context, and Test Connection passed. At about 21:43 UTC, a real extension UI task sent one authored Unicode QA message to the user's authenticated WhatsApp self-chat, showed native blue Read checks and completed successfully. The initial typing step reported an error before the successful click; that delivery alone does not certify the later editor fixes. The private screenshot is retained at ci/artifacts/quality-audit-2026-10-07/auth-whatsapp-40-0-8-self-terminal.jpg. Two self-chat QA messages were sent in total, one under the old 36.8 extension and one under 40.0.8.
With the user's explicit authorization, the fifth general-contact UI attempt on the frozen prepared-repair build sent one authored QA message to the user's mother and completed successfully. The four preceding attempts sent nothing; their owned drafts were cleared. Independent native inspection at 2026-10-08 01:24:58 UTC confirmed exactly one outgoing bubble, the exact 119-character body's SHA-256, a valid native message ID, Delivered status and an empty composer. A separate native contact-drawer inspection at 01:26:16 UTC matched the canonical hashed phone identity. Evidence is ci/artifacts/quality-audit-2026-10-07/auth-whatsapp-mom-prepared-native-delivery-proof.json and auth-whatsapp-mom-prepared-independent-identity-proof.json. No extra retry or additional contact message was sent. These single-message authenticated checks remain separate from the isolated endurance fixtures.
The earlier read-only Telegram check at 01:27:45 UTC showed a QR/link-device screen without an authenticated chat shell, and performed no external action. Evidence is ci/artifacts/quality-audit-2026-10-07/auth-telegram-login-final-check.json. After Chrome access returned and the old Telegram tab closed, a new tab's 15:22:44 UTC native baseline found one self identity, one live transcript and one empty editor, with zero exact QA-body matches. The authenticated delivery retake remains pending. Hours of authenticated Telegram/WhatsApp conversation remain unverified.
General one-to-one native WhatsApp contact support now has separate owned-browser proofs. A guarded header click must open a new native drawer whose unique subtitle matches the unchanged header, with a unique native phone field; guarded Close must leave the exact original selected row, main panel, header, title, composer and transcript in place. Phone text stays private and supplies a hashed canonical peer. Grouped incoming rows require continuous native sender/tail evidence. A delivered outgoing row must be new and follow the still-mounted original pre-send baseline; opaque IDs are never ordered lexicographically.
Before sending, the app persists the exact baseline IDs, live tail, private content-document epoch, original task/resume ID and provider configuration digest. Only a private retained floor with the same original node objects can revive after a pure worker reboot, and only after a fresh native drawer recapture. The installed same-task worker case passed in 15,089 ms with one actual authorized outgoing message, a genuine worker restart and Continue, no resend, cleared pending state and successful terminal completion. Evidence is ci/artifacts/quality-audit-2026-10-07/whatsapp-contact-mv3-worker-restored-final-pass/evidence.json. Chrome's extension-owned debugger attachment also survives worker heap loss; the resumed client proves ownership with a harmless protocol command before reusing it, without detaching another debugger.
Actual Stop is a different boundary. The separate installed case stopped after one dispatched pending message, used genuine Continue with the same task and document, and recaptured the contact drawer. It passed in 14,987 ms by retaining pending uncertainty and sending nothing again; it could not revive Stop's revoked floor or claim successful completion. Evidence is ci/artifacts/quality-audit-2026-10-07/whatsapp-contact-mv3-stop-pending-final/evidence.json. New tasks even with identical wording, provider/configuration changes, accepted steering, document loss, remounted or reordered baseline rows and a missing live tail likewise cannot inherit delivery authority. These contact fixtures do not establish authenticated general-contact delivery or hours of WhatsApp service use, and they do not support group chats.
The earlier installed WhatsApp matrix passed all 12 cases in 184,894 ms, with 601 extension/fixture source files unchanged. It includes the native postwrite-debt boundary: a cancelled native input handler writes the exact draft, a scoped Voice-to-Send rerender occurs, and the first outer send still performs no Enter or Send. Fresh observation alone does not clear the uncertain pending send. Genuine scoped reconciliation, guarded clearing and a new verified send produce exactly one delivery. Evidence is ci/artifacts/quality-audit-2026-10-07/whatsapp-contact-matrix-final-source/summary.json; that archive precedes the authenticated native-header bootstrap correction.
The earlier installed matrix passed all 13 cases in 195,285 ms with the same 601-file source-stability check. Every case uses the grounded contact-name carrier and a generic native composer placeholder. Its added bootstrap case rejects the whole header and Send before canonical identity exists, then opens and closes only the exact native controls, sends one generated reply and reaches successful terminal UI state. The 34 paired bootstrap regressions separately cover avatar, call, search and menu refusals. Same-task worker recovery completes without a resend; Stop, document loss and same-wording new tasks retain uncertainty. Evidence is ci/artifacts/quality-audit-2026-10-07/whatsapp-contact-matrix-bootstrap-final-source/summary.json. This archive precedes the later native drawer presentation-change handling and explicit current-transcript scroll repair. This is an owned native application with an intercepted local provider; the authenticated delivery is recorded separately above.
The drawer/current-transcript matrix then passed 14 cases in 209,511 ms, with its own 601-file stable snapshot preserved at ci/artifacts/quality-audit-2026-10-07/whatsapp-contact-matrix-drawer-latest-final-source/summary.json. The prepared-repair 15-case matrix passed in 251,167 ms, again with 601 source files unchanged. Its added prepared-repair case exercises the ordinary Agent pre-inference capture: fresh bounded AX certificates cover only the exact private name carrier, original captured drawer's Close and canonical non-latest transcript. The model deliberately skips additional AX reads for those targets; unrelated Send is refused, then exactly one reply is verified and terminal UI succeeds. No post-inference validation or send authorization is waived. Evidence is ci/artifacts/quality-audit-2026-10-07/whatsapp-contact-matrix-prepared-final-source/summary.json; the older 14-case archive does not establish this added preparation behavior.
After the terminal-classification correction, the complete 15-case matrix passed again in 225,636 ms with all 601 product/fixture source hashes unchanged. Current-source proof is preserved at ci/artifacts/quality-audit-2026-10-07/whatsapp-contact-matrix-scheduler-terminal-final-source/summary.json; the earlier 15-case archive remains separately qualified. The reporting validator reads the corresponding default summary and checks every case's actual reply count, successful terminal flag and durable pending state, rather than relying on a top-level pass alone.
The current scheduler/journal correction also passed the separate installed watch flow in 159,984 ms. Five real 30-second polls cover the baseline, two distinct events, a duplicate after actual worker heap loss and Stop while a response is held. The raw provider-visible alert result rejects the post-reboot duplicate; the server records exactly two acknowledgements. Stop takes 1,538 ms, closes the helper and removes the alarm, with no late mutation. Three further failed checks stop and clean up the failure-case helper. All 1,177 product/runner source hashes remain unchanged. Evidence is ci/artifacts/quality-audit-2026-10-07/watch-mv3-final-scheduled-ui-green/summary.json. This owned watch proof precedes the later authenticated WhatsApp bootstrap correction; it sends no external messages and creates no accounts.
The watch retest after that bootstrap correction passed in 160,074 ms with all 1,177 Chrome, Firefox and runner hashes unchanged. It again verifies five native alarm polls, exactly two acknowledgements, a duplicate rejected after genuine worker heap loss, and three-failure cleanup. Stop settled in 1,541 ms, closed the helper and removed the alarm without a late mutation. Evidence is ci/artifacts/quality-audit-2026-10-07/watch-mv3-native-bootstrap-final/summary.json; start and finish hashes are retained beside it. It precedes the later native drawer presentation-change and explicit scroll-repair fixes and performs no external communication.
A later prepared-source watch attempt under concurrent full coverage retained all 1,177 source hashes but expired the first event's 45-second wait after 54,388 ms overall. The real 30-second alarm fired, one acknowledgement was recorded, and two successful done attempts occurred while the durable job was still running at the deadline. Evidence is retained at ci/artifacts/quality-audit-2026-10-07/watch-mv3-native-prepared-final/failure.json. The later unchanged quiet-window retest is recorded below; this failed attempt does not extend the earlier watch proof to the latest complete source snapshot.
The unchanged watch then passed in a quiet window after the terminal-classification correction, in 160,016 ms with all 1,177 source and runner hashes unchanged. Five actual 30-second alarm polls, genuine worker heap loss, the rejected post-reboot duplicate, exactly two acknowledgements and three-failure cleanup all passed. Stop settled in 1,527 ms, closed the helper and removed the alarm, with no late mutation. Evidence is ci/artifacts/quality-audit-2026-10-07/watch-current-completion-revision-final/summary.json and its adjacent start/finish hashes. No wait threshold or assertion was weakened, and the earlier concurrent-load failure remains preserved. This verifies the ordinary watch lifecycle, rather than claiming that the watch fixture exercised an accepted mid-run task correction.