✔ Messages builds native system/tool history, bounds images, and seals a distinct mixed-pool variant (7.5352ms) ✔ Messages URL and native parameter contracts preserve other protocol paths and hard output budget (0.7113ms) ✔ native usage includes both cache buckets once, retains raw and rejects invalid counters (0.1684ms) ✔ real HTTP Messages JSON executes one tool, retains opaque history, actual URL and cache accounting (3768.657ms) ✔ real HTTP Messages SSE executes one tool, retains opaque history, actual URL and cache accounting (3165.2554ms) ✔ Messages api-key authentication is explicit; mismatched model remains blocked with original usage (5.6114ms) ✔ Messages bearer authentication is explicit; mismatched model remains blocked with original usage (4.5747ms) ✔ Messages redirect is not followed, either credential stays at the original server (3.948ms) ✔ strict native SSE reads the nested start model and retains mismatch source without retries (2.8382ms) ✔ a closed native tool block cannot execute before message_stop, then executes exactly once (3031.2422ms) ✔ mixed Chat/Messages pool seals both variants and preserves one completed tool on failover (4108.912ms) ✔ native JSON rejects incomplete stops, duplicate tools and hosted content (0.5529ms) ✔ Messages rejects empty, whitespace and thinking-only answers without rejecting valid text or tool calls (0.393ms) ✔ empty Messages JSON fails once with original response and 6692/0 usage, before verification or tools (2388.731ms) ✔ empty Messages SSE fails once with original response and 6692/0 usage, before verification or tools (2236.7692ms) ▶ strict SSE validates all blocks, terminal, event names, UTF8 and tool JSON before dispatch ✔ reject native stream 0 (1.1658ms) ✔ reject native stream 1 (0.2455ms) ✔ reject native stream 2 (0.1854ms) ✔ reject native stream 3 (0.1828ms) ✔ reject native stream 4 (0.1607ms) ✔ reject native stream 5 (0.1679ms) ✔ reject native stream 6 (0.1496ms) ✔ reject native stream 7 (0.1262ms) ✔ reject native stream 8 (0.1797ms) ✔ strict SSE validates all blocks, terminal, event names, UTF8 and tool JSON before dispatch (3.2387ms) ✔ upstream SSE error retains original type/request ID and initial usage, never retries (4.5943ms) ✔ stop cancels the owned stream and retains partial consumption without running tool or replaying (53.916ms) ✔ saved Messages import preserves auth choice, rejects endpoint duplication, and keeps source file untouched (598.865ms) ✔ large source receipts use cursor-preserving invalidation without filling the event stream or changing the session (5.4902ms) ✔ small event updates retain the current session and actual cursor (0.1333ms) ✔ the bundled local service enforces host, origin, token and JSON object contracts (2596.4617ms) ✔ an in-flight connection probe prevents configuration changes and another probe until its actual result returns (2577.4142ms) ✔ API review applies verified content, detects conflicts, exports sessions and keeps keys in memory (5211.4745ms) ✔ a real bundle restart preserves proposals, interrupts old approvals and requests fresh consent on resume (12037.7562ms) ✔ runtime diagnostics and model-response validation use actual records and preserve the saved connection (2656.2461ms) ✔ persistent event replay is read-only and keeps run/artifact identity across reconnects (8398.5137ms) ✔ real browser HTTP plans, original field pages and PNG bytes preserve ownership and survive host restart (8386.4094ms) ✔ self harness browser tools require approval and feed verified actual page fields back to the original Task (13468.9904ms) ✔ browser host recovery retains unknown root outcome, blocks replay, and requires an explicit owned acknowledgement (3551.089ms) ✔ late HTTP stop of a finished browser job cannot interrupt a later job in the same Task (7395.801ms) ✔ actual failed page blocks source-passed delivery; a new passing page permits a bound commit and preserves both original records (19309.9019ms) ✔ a build-only change after Git review invalidates the actual browser evidence and leaves HEAD and the protected index intact (9628.7245ms) ✔ actual DOM controls include validated unique CSS selectors across repeated parent buttons and escaped labels (7575.0662ms) ✔ browser field coverage distinguishes sections and Run windows while retaining original source identity (0.451ms) ✔ repeated original browser pages pause a live Task without re-opening the browser or breaking tool pairing (5543.4574ms) ✔ browser plan rejects outside builds, invalid interaction scope and unbounded steps before execution (0.9011ms) ✔ real owned Edge interactions, assertions, raw PNG and stopped process survive exact result reuse and host reopen (4002.9149ms) ✔ failed real page assertion preserves actual value and forbids silent retry of the original plan (8539.0896ms) ✔ build/source mutation and sealed step tampering reject before browser dispatch; earlier task cannot execute plan (1782.0619ms) ✔ outside network never reaches another local service, and original JavaScript/console failures remain failures (3274.4592ms) ✔ stopping a real in-flight browser preserves finished steps, closes its owned root and never replays (2782.7547ms) ✔ 原PNG损坏使原结果校验失败,不能以已有通过状态复用 (3122.4693ms) ✔ fixed browsing contracts reject local host APIs, credential URLs, protocols, cross-origin and arbitrary execution (1.008ms) ✔ attested desktop queue preserves same-tab observations, original PNG bytes and approval-bound one-shot actions (1688.6062ms) ✔ page data cannot grant domains, interaction or other Task scope; nonce and ownership remain fixed (1237.5537ms) ✔ unknown dispatched submission blocks even after a new observation, rejects late success and retains explicit acknowledgement (1582.8256ms) ✔ downloads require exact approved address, remain original binary artifacts, and cannot reuse consumed observation (927.1039ms) ✔ revoking shared persistent profile closes all its tabs and clears only that profile after original desktop acknowledgement (2150.2181ms) ✔ service recovery never reopens old tabs or replays dispatched actions and keeps unknown original receipts (1220.5736ms) ✔ agent browser flow waits for original Task/Run approval, dispatches once, and returns paired original results (5066.8215ms) ✔ changed approved input and stopped dispatched agent action cannot be accepted or replayed (1412.2295ms) ✔ browser inventory gives bounded metadata and original references instead of repeating full website histories (1098.0591ms) ✔ receipt storage failure settles the original dispatched action as unknown and never hangs or accepts late success (1530.3448ms) ✔ input that fits the engine window but exceeds remaining credit diagnoses the credit limit (0.4889ms) ✔ window overflow and historical pauses retain distinct and honest diagnoses (0.0804ms) ✔ R observed reservation shortage keeps the window-fit diagnosis separate from Run/Task credit (0.094ms) ✔ a budget pause persists concrete text/tool/history/image breakdown and never dispatches an oversized request (2459.1563ms) ✔ explicit frozen preview resumes the same Task in a new Run, preserves completed tools and does not change global config or selected model (4590.4763ms) ✔ insufficient budgets, silently clamped capacity and arbitrary config fields reject without a physical request or mutating task budget (2733.5784ms) ✔ changed session, changed registry, expiry and altered frozen preview all require a new explicit preview (3522.9859ms) ✔ preview preserves unknown invocation status and increasing Task allowance is committed only with confirmed new Run (4695.3813ms) ✔ batch APIs enforce ownership and review fingerprints, apply once, and persist the completed batch across restart (6600.9151ms) ✔ delivery reports bind actual batch identity and reject corrupted snapshot evidence without changing project files (4543.0338ms) ✔ a compatible model can seal proposals through the real HTTP loop while application stays behind user review (8688.0691ms) ✔ a sealed batch at the step limit hands off and continues the same Task into actual checks without replaying saved tools (19438.3173ms) ✔ applied batches cannot turn cancellation, other budgets, restart or unknown commands into a review continuation (2350.1324ms) ▶ complete sealed batch hands off after all tools settle with no next model request ✔ complete (5539.0166ms) ✔ later-error (5832.3898ms) ✔ complete sealed batch hands off after all tools settle with no next model request (11372.1548ms) ▶ partial, changed, failed or cancelled batches cannot claim complete review handoff ✔ partial (5081.4276ms) ✔ new-proposal (5175.7006ms) ✔ failed (4963.8489ms) ✔ conflict (5303.5908ms) ✔ cancelled (5120.9529ms) ✔ unknown (5754.5617ms) ✔ tampered (5718.0502ms) ✔ partial, changed, failed or cancelled batches cannot claim complete review handoff (37119.2538ms) ✔ actual process exit at file-batch prepared preserves durable identity and never replays writes (2130.9515ms) ✔ actual process exit at file-batch application_started preserves durable identity and never replays writes (2084.4481ms) ✔ actual process exit at file-batch write_intent preserves durable identity and never replays writes (2351.3987ms) ✔ actual process exit at file-batch staged preserves durable identity and never replays writes (2455.7ms) ✔ actual process exit at file-batch replaced preserves durable identity and never replays writes (2524.961ms) ✔ actual process exit at file-batch file_recorded preserves durable identity and never replays writes (3216.226ms) ✔ actual process exit at file-batch finished preserves durable identity and never replays writes (2740.304ms) ✔ actual process exit at file-batch rollback_started preserves durable identity and never replays writes (3539.5528ms) ✔ actual process exit at file-batch restore_intent preserves durable identity and never replays writes (3274.1664ms) ✔ actual process exit at file-batch restore_staged preserves durable identity and never replays writes (3832.3605ms) ✔ actual process exit at file-batch restored preserves durable identity and never replays writes (3264.8252ms) ✔ actual process exit at file-batch restore_recorded preserves durable identity and never replays writes (3177.7021ms) ✔ actual process exit at file-batch rollback_finished preserves durable identity and never replays writes (3227.4966ms) warning: in the working copy of 'README.md', LF will be replaced by CRLF the next time Git touches it warning: in the working copy of 'src/two.mjs', LF will be replaced by CRLF the next time Git touches it ✔ sealed batches preserve bytes until review, then apply and explicitly restore BOM/CRLF/new files (3287.3111ms) Switched to a new branch 'owner-branch' ✔ a conflict in the last file or Git branch prevents writing every file in the batch (3217.57ms) ✔ duplicate canonical paths, ancestor collisions, links, dependencies and Windows aliases never form a batch (1190.6273ms) ✔ fingerprints, changed proposals and corrupt snapshot artifacts cannot write or restore any target (2699.0279ms) ✔ a failure after replacement restores only batch bytes and never retries an old application (2279.5923ms) ✔ rollback preserves owner edits, blocks other project writers, and explicit acknowledgement releases recovery (2276.9629ms) ✔ an unattempted owner-created file identical to a suggestion remains external and is never removed (1940.903ms) ✔ active batch requests are serialized and cancellation leaves a verifiable ended batch (1550.5976ms) ✔ batch application and rollback preserve unrelated Git index and worktree bytes (3737.6643ms) ✔ recovery reuses exact staged bytes but retains a conflicting temporary file for owner inspection (2343.0169ms) ✔ one Task can seal pending proposals from earlier Runs while snapshots retain each source Run (1532.9081ms) ✔ engineering inspection preserves a dirty Git baseline and treats project rules as data (2961.4516ms) ✔ real npm tests and build produce counted results, log artifacts and persistent verification (3163.1878ms) ✔ an actual failing test retains the nonzero exit code and failed count (2497.3883ms) ✔ unchanged automatic checks return original failure evidence, while explicit user reruns and changed source still execute (4878.2722ms) ✔ agent loops cannot hide identical failing checks behind new plan IDs or unrelated reads (11096.6936ms) ✔ zero tests, skipped tests and unrecognized stdout cannot become verification passes (9429.7766ms) ✔ a successful build with no selected tests does not verify the task (2556.0986ms) ✔ stale check plans and forged script names execute no project commands (1834.5862ms) ✔ applying a later proposal invalidates the prior passed engineering snapshot (3165.9654ms) ✔ stopping an actual check terminates its child process and blocks concurrent jobs (1991.5313ms) ✔ check timeout records a failure with its actual termination result (2776.7217ms) ✔ source edits during a successful check make the result stale (3304.3391ms) ✔ bounded large logs retain the test footer and readable artifact instead of inventing counts (2534.164ms) ✔ model-selected checks require a run-bound decision and denial creates no check job (4533.7187ms) ✔ check APIs expose owned plans and jobs and preserve the authoritative results across restart (6924.6756ms) ✔ reviewed changes continue the original task without reopening a generic completed run (2954.4643ms) ✔ concurrent review-continuation API requests accept one run and retain the goal after restart (10402.8488ms) ✔ the model reads owned failure logs without running the check again (3456.2405ms) ✔ process exit at check queued preserves results and never replays commands (1691.0745ms) ✔ process exit at check result_committed preserves results and never replays commands (1956.5706ms) ✔ model code tools keep Task/Run ownership, index evidence and exact proposals across storage recovery (6061.1508ms) ✔ bundled code APIs isolate project results, validate queries and create review-only patches with conflicts (3190.0548ms) ✔ code tool arguments reject malformed ownership-independent inputs before side effects (0.4198ms) ✔ observe-only device authority cannot read, configure or clear project parsing caches (889.7012ms) ✔ pairing is short lived, single-use, generation-bound and validates exact permissions (866.2099ms) ✔ device transaction failure neither consumes the ticket nor leaves a credential (902.673ms) ✔ devices have project and action boundaries; rotation, expiration, revocation and host changes reject old authority (905.7276ms) ✔ persistent registry stores hashes and a bounded audit, never raw credentials (997.6635ms) ✔ real TLS verifies IP SAN, CA chain, encrypted restart and public fingerprint (950.5456ms) ✔ TLS sealing failure leaves listener closed and no plaintext or successful TLS registration (799.7289ms) ✔ HTTP pairing is concurrent single-use; secure cookie and strict origin/host checks (1399.609ms) ✔ gateway never exposes management routes, arbitrary assets or unpaired project data (916.6665ms) ✔ authorized project events paginate from original seq without leaking other project events or payloads (1732.4956ms) ✔ remote task projection excludes raw model history, host configuration, tool input and credentials (896.3314ms) ✔ permission checks apply to direct remote mutation, not only UI controls (1088.2767ms) ✔ approval fingerprint and exact task/revision are bound; duplicate requests never execute twice (1030.8732ms) ✔ unknown original operation is recorded and never replayed after response failure (1291.8353ms) ✔ revocation and gateway closure reject old access without invoking runtime stop (915.0914ms) ✔ rate limits reject pairing brute force and do not persist candidate secrets (1455.1545ms) ✔ real Windows CurrentUser DPAPI persists only a sealed CA and reopens it (2386.0571ms) ✔ mobile approval executes the original Runtime command exactly once and rejects the later desktop decision (2770.9563ms) ✔ desktop rejection wins a concurrent mobile request and original command remains unexecuted (2742.615ms) ✔ phone stop waits for the original process exit and prevents later heartbeats and command replay (2895.2393ms) ✔ Run replacement during parent-stop wait is rejected rather than stopping the new Run (687.7078ms) ✔ diff pages preserve source SHA; report reads verify blobs and exclude unrelated artifact kinds (1112.6832ms) ✔ hot upgrade refuses an active gateway and restores closed gateway ownership on failure (944.9518ms) ✔ damaged device permissions or expiry fail closed and never overwrite the original registry (712.0815ms) ✔ device revocation destroys its actual authenticated TLS connection (1148.3148ms) ✔ companion observes only current Task screenshots, verifies original PNG bytes and loses access immediately on revoke (1346.7096ms) ✔ mobile browser observation is current Task/project scoped, preserves original PNG and exposes no browser control or profile data (1393.1087ms) ✔ actual config HTTP failure preserves the committed selection, config and volatile credential before a successful retry (1310.3125ms) ✔ controlled installation pins the native kernel and bridge independently of the existing connection-probe identity (1013.7264ms) ✔ controlled profiles disable every native project/terminal provider and native model/retry route using the actual include patch contract (0.5478ms) ✔ owned tool catalogues seal full schemas canonically and reject duplicate or malformed declarations (5.0612ms) ✔ bridge authentication, current Run, process, SDK session and browser-origin checks reject before dispatch (652.9015ms) ✔ bridge duplicates, unknown tools, malformed arguments and source changes are refused before a second side effect (733.763ms) ✔ bridge model requests bind the exact catalogue, route and output bound and preserve raw request/result identity (682.4089ms) ✔ stopped and oversized bridge requests never reach a local model or tool callback (761.9142ms) ✔ a Windows virtual data-directory name binds new storage and credentials to its actual physical root (1879.8795ms) ✔ physical data-root adoption still rejects explicit junctions at the directory or any ancestor (741.4235ms) ✔ actual Windows CurrentUser DPAPI roundtrips owned UTF-8 and rejects damaged bytes and wrong entropy (1297.4389ms) ✔ only explicitly supplied keys persist; new provider restores original revisions and public views contain no key (3022.9618ms) ✔ registry and cipher commit failure restores original routing, keys, block and every database row (1773.4233ms) ✔ a two-key import with a protection failure rolls back all registry and cipher entries and retains its preview (2181.7461ms) ✔ connection revision changes retain the original saved block and report a source mismatch on restart (2495.0308ms) ✔ disable skips real restart loading, explicit enable restores, delete preserves memory and forget removes both (2669.7261ms) ✔ corrupt saved block and another SID do not load; failures are visible without returning ciphertext or plaintext (2790.2703ms) ✔ legacy persistence requires a new explicit key and binds config and credential revisions; stale requests cannot save (2899.1186ms) ✔ owned plaintext keys never reach SQLite pages, WAL, config or public execution exports (1573.9556ms) ✔ reviewed application invalidates old checks, names current SHA and advances through a new approved check to Git preview (8790.0795ms) ✔ skip-tests, foreign task, unknown operations, diagnostic and parent review retain their boundaries (3773.0723ms) ✔ before the first change, real baseline failures persist across later reads; passed baselines and changed sources cannot certify delivery (8779.0426ms) ✔ actual process exit at Git queued never replays a commit and can reconcile the durable result (13583.9221ms) ✔ actual process exit at Git preparing never replays a commit and can reconcile the durable result (14162.7925ms) ✔ actual process exit at Git index_locked never replays a commit and can reconcile the durable result (15975.5443ms) ✔ actual process exit at Git prepared never replays a commit and can reconcile the durable result (16063.5577ms) ✔ actual process exit at Git ref_prepared never replays a commit and can reconcile the durable result (16923.2613ms) ✔ actual process exit at Git publishing never replays a commit and can reconcile the durable result (17122.1499ms) ✔ actual process exit at Git committed never replays a commit and can reconcile the durable result (12859.1287ms) ✔ actual process exit at Git index_published never replays a commit and can reconcile the durable result (13486.3737ms) ✔ actual process exit at Git finished never replays a commit and can reconcile the durable result (8627.3641ms) ✔ an unknown Git result cannot overwrite an externally changed index or delete an unrelated lock (8436.9194ms) ✔ known publication with a later source edit retains the commit and marks the delivery for inspection (9722.1788ms) ✔ CRLF normalization, multiple applied revisions and unrelated special index flags preserve exact bytes and ownership (11242.7322ms) ✔ adding new durable record kinds to an existing hot Store preserves all existing Maps and records (7931.8402ms) ✔ an unrelated owner deletion is fingerprinted and preserved while linked task paths are rejected (13372.1161ms) ✔ actual failed tests become a reviewed local commit and evidence report while preserving unrelated index and files (16151.2154ms) ✔ selected preexisting unstaged, staged and untracked owner edits cannot enter a task commit (19683.3082ms) Switched to a new branch 'owner-branch' ✔ source, unrelated stage and branch changes after review invalidate a commit plan without changing HEAD (23028.3159ms) Switched to a new branch 'delivery-branch' ✔ a later Git metadata change keeps matching source evidence valid but a real source edit requires checks (12231.5204ms) ✔ a corrupt passed-check log and a pending proposal block Git delivery instead of inventing evidence (6104.0081ms) HEAD is now at c0e9adc Sealed delivery fixture baseline ✔ Git filters, merge state, detached HEAD and old task ownership are explicitly rejected (24570.6056ms) ✔ local snapshot commits disable repository hooks and signing and isolate the Git child environment (8499.0624ms) ✔ stopping before publication cancels safely; stopping after publication preserves the actual completed commit (15892.5302ms) ✔ delivery API confirms exactly one owned plan, serves verified reports and preserves results across restart (13600.0445ms) ✔ the model can prepare an owned Git review and evidence report while publishing remains a separate user action (9688.4723ms) ✔ HTTP Git review pairs deferred calls, survives reload and never requests more work until the saved preview is resolved (10399.7665ms) ✔ a reviewed Git preview can be revised or withdrawn without publishing or losing the current Task (9285.7517ms) ✔ actual Electron debug argument ordering retains exactly the compiled entry and rejects extra or repeated actions (1.1848ms) ✔ native bridge refuses other origins, subframes, other windows and unsupported inputs before connecting (512.3666ms) ✔ desktop preview opening requires a currently registered project preview and rejects iframe callers (628.521ms) ✔ file reveal binds the current project, task, path, UTF-8 bytes and original SHA without executing the file (1056.0635ms) ✔ file changed during asynchronous task recheck is not revealed as the old bytes (641.9925ms) ✔ desktop browser profile and aliases stay outside project inventory and file access (974.3896ms) ✔ native picker cancel and origin change after dialog publish no project registration (582.7162ms) ✔ pending native action is not dispatched twice, and current task review returns its original seq (624.6805ms) ✔ native stop cancellation and changed original host never stop a replacement process (632.9859ms) ✔ window register rejects host PID, stale installation, extra fields and missing creation identity (0.3833ms) ✔ a rebuilt source archive binds actual generated client assets and retains its original running identity after disk changes (1766.145ms) ✔ host challenge proof rejects a stale challenge, changed identity and changed signature (0.8294ms) ✔ stop confirmation binds the original instance, creation and loaded runtime (0.136ms) ✔ state paths reject escaping, linked ancestors, ambiguous hardlinks and damaged JSON (952.6159ms) ✔ failed registration publishes no initialized memory and can complete after its actual inputs are supplied (2595.2259ms) ✔ host shutdown cancels an un-signalled HTTP model request and refuses later requests until recovered (1.5739ms) ✔ host shutdown cancels model-pool acquisition and releases queue ownership without inventing attempts (1004.7195ms) ✔ full candidate validation returns the exact loaded-input host identity and rejects undeclared or altered files (885.7308ms) ✔ fast per-file scan still rejects a junctioned subdirectory, a hard-linked file and a stream path with the original bytes intact (1017ms) ✔ installation requires separate absolute program/data directories and a candidate in its own versions namespace (1217.1969ms) ✔ inventory rejects a new hardlink on an optional file after its full-manifest scan and accepts the untouched candidate after removal (652.7417ms) ✔ staging retains the original data and pointer without querying or restarting a host (3180.4634ms) ✔ a post-pointer failure restores every original pointer byte and persistent row; a first-install failure removes only its published pointer (11178.8377ms) ✔ changed active manifest and concurrent pointer changes are rejected without overwriting the unknown record (4307.3952ms) ✔ an unregistered non-HTTP listener blocks preflight even when service-info is absent (2550.6911ms) ✔ uninstall archives its original pointer and preserves every data row/artifact instead of recursively deleting data (7273.2218ms) ✔ preflight rejects an unrelated program root or changed native bootstrap while preserving its original bytes (4438.6831ms) ✔ a changed owned bootstrap follows its candidate and a post-pointer fault restores both original bytes (16531.6276ms) ✔ an external bootstrap change after preflight is preserved and never overwritten as an update or rollback (8049.29ms) ✔ owned startup retries only read-only observation after a slow health response and attests the same original child (4337.727ms) ✔ unverified startup expires with its original live child and never treats timeout or an unrelated health PID as proof (5908.0335ms) ✔ diagnostic failures and repeated questions retain the original execution error, task, checks and read-only authority (2143.4679ms) ✔ diagnostic binding survives reopen, rejects foreign Task references and detects original error or artifact tampering (1091.446ms) ✔ a new Task never adopts a previous Task error from session.lastFailure (1012.2443ms) ✔ verified research and explicit no-test scopes do not demand an unrelated test job or claim engineering success (3012.5116ms) ✔ unknown scope and unknown operations remain actionable even after a diagnostic answer or read proof (1047.3595ms) ✔ delivery status HTTP is read-only, selection-bound and equal to the saved report after host restart (5101.1775ms) ✔ routine engineering facts remove duplicate success excerpts while complete diagnostics and errors remain available (655.5671ms) ✔ a question about the previous workbench error receives local error context and current change states (1953.8633ms) ✔ an unknown model request stops, and explicit resume never replays the approved command (3940.5575ms) ✔ exhausted refused connections preserve tool results, and explicit resume adds no new user turn or replay (3115.4404ms) ✔ stopping during model backoff cancels further requests and clears the running request state (1649.6299ms) ✔ HTTP authentication errors are classified, redacted and never automatically retried (17.1097ms) ✔ temporary HTTP service errors retry and recover with recorded attempt counts (17.7165ms) ✔ invalid and oversized protocol responses fail once with a useful response-format diagnosis (5.3243ms) ✔ HTTP rate limits expose their source and obey the service retry interval before recovery (1009.2634ms) ✔ Windows commands preserve Chinese through a PowerShell pipeline into Node (782.5448ms) ✔ Windows shell syntax errors stay readable and native command exit codes are retained (1003.112ms) ✔ review and completed states do not turn a generic continue click into a new model request (745.5677ms) ✔ old session data receives a truthful historical connection diagnosis without rewriting history (683.48ms) ✔ engine HTTP checks bind the detected installation and current task, page original bytes and preserve them across host restart (5621.3275ms) ✔ HTTP stop seals a pending SDK call, rejects automatic replay and keeps failures attached to the original session (4399.3421ms) ✔ sealed r1-off failure fits its original 16K config with authentic latest status and readable source (132.3009ms) ✔ sealed r1-on failure fits its original 16K config with authentic latest status and readable source (77.4701ms) ✔ sealed r2-on failure fits its original 16K config with authentic latest status and readable source (36.6452ms) ✔ sealed r2-off failure fits its original 16K config with authentic latest status and readable source (43.6342ms) ✔ archived native call pairs require exact saved arguments/results rather than model-written summaries (8.5148ms) ✔ full system/user requirements and mandatory current failure can refuse rather than be trimmed (12.85ms) ✔ small native input is byte-preserving; developer instructions remain mandatory after compaction (0.191ms) ✔ older failing checks and pending proposals remain visible after a later large file read (15.1039ms) ✔ bounded result keeps every check status and cuts Unicode only on whole characters (0.4966ms) ✔ controlled engine HTTP freezes one child, pages original owned input/report, rejects foreign reads and prevents changed-source reuse (15251.4503ms) ✔ a sealed whole-batch review pauses the owned native session without certifying engine completion or losing the saved host result (14782.4442ms) ✔ a changed review seal still closes the owned native process and cannot claim a verified review handoff (15007.5049ms) ✔ native text and call IDs preserve original result pairing while unsupported content fails closed (0.4767ms) ✔ one engine child freezes real installation identity before creating tasks and rejects a changed source (8927.2184ms) ✔ owned native tools return saved results, use real approvals/checks and account each actual request once across new engine Runs (26740.4614ms) ✔ forged native scope cannot change a parent file and a final completion claim cannot replace worker proof (13747.1114ms) ✔ duplicate native IDs are rejected without replaying a saved tool (13010.9053ms) ✔ native input exceeding the context budget is retained and sends no model request (11322.004ms) ✔ bounded native sends verify original tool provenance and reject forged views or missing versions (13820.2123ms) ✔ stop retains a late reported model usage once and executes no tool from that output (10855.3183ms) ✔ a source change after the original model intent stops before native tools execute (10640.6312ms)