✔ Messages builds native system/tool history, bounds images, and seals a distinct mixed-pool variant (7.926ms) ✔ Messages URL and native parameter contracts preserve other protocol paths and hard output budget (0.5547ms) ✔ native usage includes both cache buckets once, retains raw and rejects invalid counters (0.1555ms) ✔ real HTTP Messages JSON executes one tool, retains opaque history, actual URL and cache accounting (7050.3755ms) ✔ real HTTP Messages SSE executes one tool, retains opaque history, actual URL and cache accounting (10008.0698ms) ✔ Messages api-key authentication is explicit; mismatched model remains blocked with original usage (7.3906ms) ✔ Messages bearer authentication is explicit; mismatched model remains blocked with original usage (3.9648ms) ✔ Messages redirect is not followed, either credential stays at the original server (3.916ms) ✔ strict native SSE reads the nested start model and retains mismatch source without retries (2.7818ms) ✔ a closed native tool block cannot execute before message_stop, then executes exactly once (11383.4232ms) ✔ mixed Chat/Messages pool seals both variants and preserves one completed tool on failover (13134.4843ms) ✔ native JSON rejects incomplete stops, duplicate tools and hosted content (0.2814ms) ▶ strict SSE validates all blocks, terminal, event names, UTF8 and tool JSON before dispatch ✔ reject native stream 0 (1.0664ms) ✔ reject native stream 1 (0.2813ms) ✔ reject native stream 2 (0.1847ms) ✔ reject native stream 3 (0.1762ms) ✔ reject native stream 4 (0.1568ms) ✔ reject native stream 5 (0.1662ms) ✔ reject native stream 6 (0.1673ms) ✔ reject native stream 7 (0.1998ms) ✔ reject native stream 8 (0.2068ms) ✔ strict SSE validates all blocks, terminal, event names, UTF8 and tool JSON before dispatch (3.2986ms) ✔ upstream SSE error retains original type/request ID and initial usage, never retries (4.1501ms) ✔ stop cancels the owned stream and retains partial consumption without running tool or replaying (55.4438ms) ✔ saved Messages import preserves auth choice, rejects endpoint duplication, and keeps source file untouched (2259.6322ms) ✔ large source receipts use cursor-preserving invalidation without filling the event stream or changing the session (5.3974ms) ✔ small event updates retain the current session and actual cursor (0.1153ms) ✔ the bundled local service enforces host, origin, token and JSON object contracts (4065.5302ms) ✔ an in-flight connection probe prevents configuration changes and another probe until its actual result returns (4504.391ms) ✔ API review applies verified content, detects conflicts, exports sessions and keeps keys in memory (9120.176ms) ✖ a real bundle restart preserves proposals, interrupts old approvals and requests fresh consent on resume (21946.3358ms) ✔ runtime diagnostics and model-response validation use actual records and preserve the saved connection (4813.5141ms) ✔ persistent event replay is read-only and keeps run/artifact identity across reconnects (16036.4609ms) ✔ real browser HTTP plans, original field pages and PNG bytes preserve ownership and survive host restart (14081.017ms) ✔ self harness browser tools require approval and feed verified actual page fields back to the original Task (15337.231ms) ✔ browser host recovery retains unknown root outcome, blocks replay, and requires an explicit owned acknowledgement (4771.232ms) ✔ late HTTP stop of a finished browser job cannot interrupt a later job in the same Task (8705.0707ms) ✔ actual failed page blocks source-passed delivery; a new passing page permits a bound commit and preserves both original records (21612.863ms) ✔ a build-only change after Git review invalidates the actual browser evidence and leaves HEAD and the protected index intact (10707.4397ms) ✔ actual DOM controls include validated unique CSS selectors across repeated parent buttons and escaped labels (9749.5584ms) ✔ browser field coverage distinguishes sections and Run windows while retaining original source identity (0.5398ms) ✔ repeated original browser pages pause a live Task without re-opening the browser or breaking tool pairing (11968.8276ms) ✔ browser plan rejects outside builds, invalid interaction scope and unbounded steps before execution (0.9144ms) ✔ real owned Edge interactions, assertions, raw PNG and stopped process survive exact result reuse and host reopen (7815.1652ms) ✔ failed real page assertion preserves actual value and forbids silent retry of the original plan (11996.5816ms) ✔ build/source mutation and sealed step tampering reject before browser dispatch; earlier task cannot execute plan (1539.2795ms) ✔ outside network never reaches another local service, and original JavaScript/console failures remain failures (4042.1913ms) ✔ stopping a real in-flight browser preserves finished steps, closes its owned root and never replays (3387.9742ms) ✔ 原PNG损坏使原结果校验失败,不能以已有通过状态复用 (3994.7021ms) ✔ fixed browsing contracts reject local host APIs, credential URLs, protocols, cross-origin and arbitrary execution (0.9861ms) ✔ attested desktop queue preserves same-tab observations, original PNG bytes and approval-bound one-shot actions (3103.5236ms) ✔ page data cannot grant domains, interaction or other Task scope; nonce and ownership remain fixed (1782.0292ms) ✔ unknown dispatched submission blocks even after a new observation, rejects late success and retains explicit acknowledgement (2876.988ms) ✔ downloads require exact approved address, remain original binary artifacts, and cannot reuse consumed observation (1560.5363ms) ✔ revoking shared persistent profile closes all its tabs and clears only that profile after original desktop acknowledgement (2355.8889ms) ✔ service recovery never reopens old tabs or replays dispatched actions and keeps unknown original receipts (2393.7594ms) ✔ agent browser flow waits for original Task/Run approval, dispatches once, and returns paired original results (8812.4758ms) ✔ changed approved input and stopped dispatched agent action cannot be accepted or replayed (6431.7347ms) ✔ browser inventory gives bounded metadata and original references instead of repeating full website histories (3194.2235ms) ✔ receipt storage failure settles the original dispatched action as unknown and never hangs or accepts late success (3560.4832ms) ✔ input that fits the engine window but exceeds remaining credit diagnoses the credit limit (0.579ms) ✔ window overflow and historical pauses retain distinct and honest diagnoses (0.0809ms) ✔ R observed reservation shortage keeps the window-fit diagnosis separate from Run/Task credit (0.1076ms) ✔ a budget pause persists concrete text/tool/history/image breakdown and never dispatches an oversized request (6419.0989ms) ✔ explicit frozen preview resumes the same Task in a new Run, preserves completed tools and does not change global config or selected model (13985.8148ms) ✔ insufficient budgets, silently clamped capacity and arbitrary config fields reject without a physical request or mutating task budget (9024.7211ms) ✔ changed session, changed registry, expiry and altered frozen preview all require a new explicit preview (7386.2967ms) ✔ preview preserves unknown invocation status and increasing Task allowance is committed only with confirmed new Run (7089.8261ms) ✔ batch APIs enforce ownership and review fingerprints, apply once, and persist the completed batch across restart (12644.9042ms) ✔ delivery reports bind actual batch identity and reject corrupted snapshot evidence without changing project files (7890.5666ms) ✔ a compatible model can seal proposals through the real HTTP loop while application stays behind user review (10528.9458ms) ✔ a sealed batch at the step limit hands off and continues the same Task into actual checks without replaying saved tools (43924.0808ms) ✔ applied batches cannot turn cancellation, other budgets, restart or unknown commands into a review continuation (7030.3656ms) ▶ complete sealed batch hands off after all tools settle with no next model request ✔ complete (8187.5853ms) ✔ later-error (10988.0182ms) ✔ complete sealed batch hands off after all tools settle with no next model request (19176.4526ms) ▶ partial, changed, failed or cancelled batches cannot claim complete review handoff ✔ partial (18749.404ms) ✔ new-proposal (15525.2008ms) ✔ failed (13836.1382ms) ✔ conflict (15478.9001ms) ✔ cancelled (15345.1998ms) ✔ unknown (16846.4675ms) ✔ tampered (18567.4648ms) ✔ partial, changed, failed or cancelled batches cannot claim complete review handoff (114350.1302ms) ✔ actual process exit at file-batch prepared preserves durable identity and never replays writes (5844.4412ms) ✔ actual process exit at file-batch application_started preserves durable identity and never replays writes (6279.9749ms) ✔ actual process exit at file-batch write_intent preserves durable identity and never replays writes (5794.6037ms) ✔ actual process exit at file-batch staged preserves durable identity and never replays writes (6237.952ms) ✔ actual process exit at file-batch replaced preserves durable identity and never replays writes (7524.2532ms) ✔ actual process exit at file-batch file_recorded preserves durable identity and never replays writes (7428.1676ms) ✔ actual process exit at file-batch finished preserves durable identity and never replays writes (9816.1178ms) ✔ actual process exit at file-batch rollback_started preserves durable identity and never replays writes (9386.8227ms) ✔ actual process exit at file-batch restore_intent preserves durable identity and never replays writes (11192.6789ms) ✔ actual process exit at file-batch restore_staged preserves durable identity and never replays writes (11560.1703ms) ✔ actual process exit at file-batch restored preserves durable identity and never replays writes (12592.1243ms) ✔ actual process exit at file-batch restore_recorded preserves durable identity and never replays writes (15098.0539ms) ✔ actual process exit at file-batch rollback_finished preserves durable identity and never replays writes (7881.5973ms) 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 (9626.4556ms) Switched to a new branch 'owner-branch' ✔ a conflict in the last file or Git branch prevents writing every file in the batch (11329.3457ms) ✔ duplicate canonical paths, ancestor collisions, links, dependencies and Windows aliases never form a batch (7722.602ms) ✔ fingerprints, changed proposals and corrupt snapshot artifacts cannot write or restore any target (8966.8643ms) ✔ a failure after replacement restores only batch bytes and never retries an old application (3015.7526ms) ✔ rollback preserves owner edits, blocks other project writers, and explicit acknowledgement releases recovery (1986.5036ms) ✔ an unattempted owner-created file identical to a suggestion remains external and is never removed (2265.3507ms) ✔ active batch requests are serialized and cancellation leaves a verifiable ended batch (2009.5482ms) ✔ batch application and rollback preserve unrelated Git index and worktree bytes (3939.1981ms) ✔ recovery reuses exact staged bytes but retains a conflicting temporary file for owner inspection (2674.1553ms) ✔ one Task can seal pending proposals from earlier Runs while snapshots retain each source Run (2337.7991ms) ✔ engineering inspection preserves a dirty Git baseline and treats project rules as data (2355.1465ms) ✔ real npm tests and build produce counted results, log artifacts and persistent verification (3692.4431ms) ✔ an actual failing test retains the nonzero exit code and failed count (2800.2864ms) ✔ unchanged automatic checks return original failure evidence, while explicit user reruns and changed source still execute (5156.8449ms) ✔ agent loops cannot hide identical failing checks behind new plan IDs or unrelated reads (12296.0667ms) ✔ zero tests, skipped tests and unrecognized stdout cannot become verification passes (10049.139ms) ✔ a successful build with no selected tests does not verify the task (2149.8955ms) ✔ stale check plans and forged script names execute no project commands (1745.9328ms) ✔ applying a later proposal invalidates the prior passed engineering snapshot (3633.7013ms) ✔ stopping an actual check terminates its child process and blocks concurrent jobs (2233.7377ms) ✔ check timeout records a failure with its actual termination result (2553.4919ms) ✔ source edits during a successful check make the result stale (3826.7829ms) ✔ bounded large logs retain the test footer and readable artifact instead of inventing counts (2997.3221ms) ✔ model-selected checks require a run-bound decision and denial creates no check job (4654.3115ms) ✔ check APIs expose owned plans and jobs and preserve the authoritative results across restart (7402.7251ms) ✔ reviewed changes continue the original task without reopening a generic completed run (4935.7283ms) ✔ concurrent review-continuation API requests accept one run and retain the goal after restart (12002.1713ms) ✔ the model reads owned failure logs without running the check again (4560.4643ms) ✔ process exit at check queued preserves results and never replays commands (1939.8857ms) ✔ process exit at check result_committed preserves results and never replays commands (2450.6425ms) ✔ model code tools keep Task/Run ownership, index evidence and exact proposals across storage recovery (7517.9501ms) ✔ bundled code APIs isolate project results, validate queries and create review-only patches with conflicts (3684.8438ms) ✔ code tool arguments reject malformed ownership-independent inputs before side effects (0.2751ms) ✔ observe-only device authority cannot read, configure or clear project parsing caches (1068.1361ms) ✔ pairing is short lived, single-use, generation-bound and validates exact permissions (1001.1785ms) ✔ device transaction failure neither consumes the ticket nor leaves a credential (705.7614ms) ✔ devices have project and action boundaries; rotation, expiration, revocation and host changes reject old authority (771.8585ms) ✔ persistent registry stores hashes and a bounded audit, never raw credentials (799.8425ms) ✔ real TLS verifies IP SAN, CA chain, encrypted restart and public fingerprint (1070.8548ms) ✔ TLS sealing failure leaves listener closed and no plaintext or successful TLS registration (922.9892ms) ✔ HTTP pairing is concurrent single-use; secure cookie and strict origin/host checks (1566.9948ms) ✔ gateway never exposes management routes, arbitrary assets or unpaired project data (1461.7499ms) ✔ authorized project events paginate from original seq without leaking other project events or payloads (1767.3003ms) ✔ remote task projection excludes raw model history, host configuration, tool input and credentials (1122.0948ms) ✔ permission checks apply to direct remote mutation, not only UI controls (1123.9322ms) ✔ approval fingerprint and exact task/revision are bound; duplicate requests never execute twice (1475.7389ms) ✔ unknown original operation is recorded and never replayed after response failure (1024.4278ms) ✔ revocation and gateway closure reject old access without invoking runtime stop (1366.0682ms) ✔ rate limits reject pairing brute force and do not persist candidate secrets (830.4089ms) ✔ real Windows CurrentUser DPAPI persists only a sealed CA and reopens it (2035.2396ms) ✔ mobile approval executes the original Runtime command exactly once and rejects the later desktop decision (3356.092ms) ✔ desktop rejection wins a concurrent mobile request and original command remains unexecuted (4435.8827ms) ✔ phone stop waits for the original process exit and prevents later heartbeats and command replay (3450.2528ms) ✔ Run replacement during parent-stop wait is rejected rather than stopping the new Run (1065.7337ms) ✔ diff pages preserve source SHA; report reads verify blobs and exclude unrelated artifact kinds (1779.4031ms) ✔ hot upgrade refuses an active gateway and restores closed gateway ownership on failure (644.8366ms) ✔ damaged device permissions or expiry fail closed and never overwrite the original registry (1050.1985ms) ✔ device revocation destroys its actual authenticated TLS connection (1075.3849ms) ✔ companion observes only current Task screenshots, verifies original PNG bytes and loses access immediately on revoke (1443.8132ms) ✔ mobile browser observation is current Task/project scoped, preserves original PNG and exposes no browser control or profile data (1190.4144ms) ✔ actual config HTTP failure preserves the committed selection, config and volatile credential before a successful retry (1642.9537ms) ✔ controlled installation pins the native kernel and bridge independently of the existing connection-probe identity (1333.272ms) ✔ controlled profiles disable every native project/terminal provider and native model/retry route using the actual include patch contract (0.5147ms) ✔ owned tool catalogues seal full schemas canonically and reject duplicate or malformed declarations (4.9438ms) ✔ bridge authentication, current Run, process, SDK session and browser-origin checks reject before dispatch (1466.8716ms) ✔ bridge duplicates, unknown tools, malformed arguments and source changes are refused before a second side effect (928.2891ms) ✔ bridge model requests bind the exact catalogue, route and output bound and preserve raw request/result identity (948.3985ms) ✔ stopped and oversized bridge requests never reach a local model or tool callback (690.7854ms) ✔ a Windows virtual data-directory name binds new storage and credentials to its actual physical root (2220.5068ms) ✔ physical data-root adoption still rejects explicit junctions at the directory or any ancestor (873.2275ms) ✔ actual Windows CurrentUser DPAPI roundtrips owned UTF-8 and rejects damaged bytes and wrong entropy (1348.5353ms) ✔ only explicitly supplied keys persist; new provider restores original revisions and public views contain no key (3386.934ms) ✔ registry and cipher commit failure restores original routing, keys, block and every database row (2086.8197ms) ✔ a two-key import with a protection failure rolls back all registry and cipher entries and retains its preview (2550.8761ms) ✔ connection revision changes retain the original saved block and report a source mismatch on restart (2751.0184ms) ✔ disable skips real restart loading, explicit enable restores, delete preserves memory and forget removes both (3317.3097ms) ✔ corrupt saved block and another SID do not load; failures are visible without returning ciphertext or plaintext (3207.8473ms) ✔ legacy persistence requires a new explicit key and binds config and credential revisions; stale requests cannot save (2974.3489ms) ✔ owned plaintext keys never reach SQLite pages, WAL, config or public execution exports (1759.5992ms) ✔ reviewed application invalidates old checks, names current SHA and advances through a new approved check to Git preview (8408.9225ms) ✔ skip-tests, foreign task, unknown operations, diagnostic and parent review retain their boundaries (4038.5091ms) ✔ before the first change, real baseline failures persist across later reads; passed baselines and changed sources cannot certify delivery (9457.6641ms) ✔ actual process exit at Git queued never replays a commit and can reconcile the durable result (13753.1587ms) ✔ actual process exit at Git preparing never replays a commit and can reconcile the durable result (20784.4179ms) ✔ actual process exit at Git index_locked never replays a commit and can reconcile the durable result (18665.0157ms) ✔ actual process exit at Git prepared never replays a commit and can reconcile the durable result (16507.2619ms) ✔ actual process exit at Git ref_prepared never replays a commit and can reconcile the durable result (16257.5446ms) ✔ actual process exit at Git publishing never replays a commit and can reconcile the durable result (16183.4231ms) ✔ actual process exit at Git committed never replays a commit and can reconcile the durable result (11020.4125ms) ✔ actual process exit at Git index_published never replays a commit and can reconcile the durable result (12137.7426ms) ✔ actual process exit at Git finished never replays a commit and can reconcile the durable result (11107.3857ms) ✔ an unknown Git result cannot overwrite an externally changed index or delete an unrelated lock (10669.244ms) ✔ known publication with a later source edit retains the commit and marks the delivery for inspection (11124.3288ms) ✔ CRLF normalization, multiple applied revisions and unrelated special index flags preserve exact bytes and ownership (13607.2766ms) ✔ adding new durable record kinds to an existing hot Store preserves all existing Maps and records (7458.0611ms) ✔ an unrelated owner deletion is fingerprinted and preserved while linked task paths are rejected (13619.7579ms) ✔ actual failed tests become a reviewed local commit and evidence report while preserving unrelated index and files (21915.2163ms) ✔ selected preexisting unstaged, staged and untracked owner edits cannot enter a task commit (23711.3011ms) Switched to a new branch 'owner-branch' ✔ source, unrelated stage and branch changes after review invalidate a commit plan without changing HEAD (24336.802ms) Switched to a new branch 'delivery-branch' ✔ a later Git metadata change keeps matching source evidence valid but a real source edit requires checks (13101.6545ms) ✔ a corrupt passed-check log and a pending proposal block Git delivery instead of inventing evidence (5977.9169ms) HEAD is now at c9905ed Sealed delivery fixture baseline ✔ Git filters, merge state, detached HEAD and old task ownership are explicitly rejected (23369.1805ms) ✔ local snapshot commits disable repository hooks and signing and isolate the Git child environment (10197.326ms) ✔ stopping before publication cancels safely; stopping after publication preserves the actual completed commit (19992.1095ms) ✔ delivery API confirms exactly one owned plan, serves verified reports and preserves results across restart (16101.7733ms) ✔ the model can prepare an owned Git review and evidence report while publishing remains a separate user action (9529.5292ms) ✔ HTTP Git review pairs deferred calls, survives reload and never requests more work until the saved preview is resolved (9955.0042ms) ✔ a reviewed Git preview can be revised or withdrawn without publishing or losing the current Task (11935.4077ms) ✔ actual Electron debug argument ordering retains exactly the compiled entry and rejects extra or repeated actions (1.1489ms) ✔ native bridge refuses other origins, subframes, other windows and unsupported inputs before connecting (1537.4058ms) ✔ desktop preview opening requires a currently registered project preview and rejects iframe callers (1646.2497ms) ✔ file reveal binds the current project, task, path, UTF-8 bytes and original SHA without executing the file (1695.0535ms) ✔ file changed during asynchronous task recheck is not revealed as the old bytes (1267.2412ms) ✔ desktop browser profile and aliases stay outside project inventory and file access (1067.6108ms) ✔ native picker cancel and origin change after dialog publish no project registration (596.8926ms) ✔ pending native action is not dispatched twice, and current task review returns its original seq (564.0459ms) ✔ native stop cancellation and changed original host never stop a replacement process (477.1295ms) ✔ window register rejects host PID, stale installation, extra fields and missing creation identity (0.4047ms) ✔ a rebuilt source archive binds actual generated client assets and retains its original running identity after disk changes (2027.1581ms) ✔ host challenge proof rejects a stale challenge, changed identity and changed signature (0.7991ms) ✔ stop confirmation binds the original instance, creation and loaded runtime (0.1397ms) ✔ state paths reject escaping, linked ancestors, ambiguous hardlinks and damaged JSON (773.8496ms) ✔ failed registration publishes no initialized memory and can complete after its actual inputs are supplied (2679.1892ms) ✔ host shutdown cancels an un-signalled HTTP model request and refuses later requests until recovered (1.3975ms) ✔ host shutdown cancels model-pool acquisition and releases queue ownership without inventing attempts (1842.4886ms) ✔ full candidate validation returns the exact loaded-input host identity and rejects undeclared or altered files (1194.1615ms) ✔ fast per-file scan still rejects a junctioned subdirectory, a hard-linked file and a stream path with the original bytes intact (676.3638ms) ✔ installation requires separate absolute program/data directories and a candidate in its own versions namespace (813.7336ms) ✔ inventory rejects a new hardlink on an optional file after its full-manifest scan and accepts the untouched candidate after removal (1711.1354ms) ✔ staging retains the original data and pointer without querying or restarting a host (4756.8933ms) ✔ a post-pointer failure restores every original pointer byte and persistent row; a first-install failure removes only its published pointer (11965.4256ms) ✔ changed active manifest and concurrent pointer changes are rejected without overwriting the unknown record (5287.7243ms) ✔ an unregistered non-HTTP listener blocks preflight even when service-info is absent (2536.8126ms) ✔ uninstall archives its original pointer and preserves every data row/artifact instead of recursively deleting data (8428.84ms) ✔ preflight rejects an unrelated program root or changed native bootstrap while preserving its original bytes (5270.0855ms) ✔ a changed owned bootstrap follows its candidate and a post-pointer fault restores both original bytes (21729.1492ms) ✔ an external bootstrap change after preflight is preserved and never overwritten as an update or rollback (9276.3947ms) ✔ owned startup retries only read-only observation after a slow health response and attests the same original child (5741.4697ms) ✔ unverified startup expires with its original live child and never treats timeout or an unrelated health PID as proof (5421.8103ms) ✔ diagnostic failures and repeated questions retain the original execution error, task, checks and read-only authority (3958.7704ms) ✔ diagnostic binding survives reopen, rejects foreign Task references and detects original error or artifact tampering (1657.3492ms) ✔ a new Task never adopts a previous Task error from session.lastFailure (2022.6424ms) ✔ verified research and explicit no-test scopes do not demand an unrelated test job or claim engineering success (2953.1052ms) ✔ unknown scope and unknown operations remain actionable even after a diagnostic answer or read proof (1303.5396ms) ✔ delivery status HTTP is read-only, selection-bound and equal to the saved report after host restart (4798.4465ms) ✔ routine engineering facts remove duplicate success excerpts while complete diagnostics and errors remain available (1544.7396ms) ✔ a question about the previous workbench error receives local error context and current change states (1906.9583ms) ✔ an unknown model request stops, and explicit resume never replays the approved command (5404.3301ms) ✔ exhausted refused connections preserve tool results, and explicit resume adds no new user turn or replay (5220.1496ms) ✔ stopping during model backoff cancels further requests and clears the running request state (1794.2644ms) ✔ HTTP authentication errors are classified, redacted and never automatically retried (18.4827ms) ✔ temporary HTTP service errors retry and recover with recorded attempt counts (16.5453ms) ✔ invalid and oversized protocol responses fail once with a useful response-format diagnosis (5.5412ms) ✔ HTTP rate limits expose their source and obey the service retry interval before recovery (1013.8058ms) ✔ Windows commands preserve Chinese through a PowerShell pipeline into Node (988.1443ms) ✔ Windows shell syntax errors stay readable and native command exit codes are retained (1131.3793ms) ✔ review and completed states do not turn a generic continue click into a new model request (644.0571ms) ✔ old session data receives a truthful historical connection diagnosis without rewriting history (979.463ms) ✔ engine HTTP checks bind the detected installation and current task, page original bytes and preserve them across host restart (7115.2653ms) ✔ HTTP stop seals a pending SDK call, rejects automatic replay and keeps failures attached to the original session (5339.4642ms) ✔ sealed r1-off failure fits its original 16K config with authentic latest status and readable source (12.6734ms) ✔ sealed r1-on failure fits its original 16K config with authentic latest status and readable source (9.9615ms) ✔ sealed r2-on failure fits its original 16K config with authentic latest status and readable source (7.7069ms) ✔ sealed r2-off failure fits its original 16K config with authentic latest status and readable source (7.4636ms) ✔ archived native call pairs require exact saved arguments/results rather than model-written summaries (8.2225ms) ✔ full system/user requirements and mandatory current failure can refuse rather than be trimmed (12.1818ms) ✔ small native input is byte-preserving; developer instructions remain mandatory after compaction (0.1782ms) ✔ older failing checks and pending proposals remain visible after a later large file read (14.3888ms) ✔ bounded result keeps every check status and cuts Unicode only on whole characters (0.4672ms) ✔ controlled engine HTTP freezes one child, pages original owned input/report, rejects foreign reads and prevents changed-source reuse (20689.2027ms) ✔ a sealed whole-batch review pauses the owned native session without certifying engine completion or losing the saved host result (18954.7303ms) ✔ a changed review seal still closes the owned native process and cannot claim a verified review handoff (13579.0209ms) ✔ native text and call IDs preserve original result pairing while unsupported content fails closed (0.4209ms) ✔ one engine child freezes real installation identity before creating tasks and rejects a changed source (8753.8292ms) ✔ owned native tools return saved results, use real approvals/checks and account each actual request once across new engine Runs (28433.0875ms) ✔ forged native scope cannot change a parent file and a final completion claim cannot replace worker proof (15011.8852ms) ✔ duplicate native IDs are rejected without replaying a saved tool (11410.2452ms) ✔ native input exceeding the context budget is retained and sends no model request (10229.4092ms) ✔ bounded native sends verify original tool provenance and reject forged views or missing versions (12915.0855ms) ✔ stop retains a late reported model usage once and executes no tool from that output (11509.8947ms) ✔ a source change after the original model intent stops before native tools execute (11387.0856ms) ✔ an engine source change while approval is pending invalidates the old permission before any check starts (14376.4051ms) ✔ missing or noncompleted turn end cannot be adopted even with an owned receipt and assistant text (0.5691ms) ✔ closed root with an original live descendant remains unknown and retains its process tree (0.2752ms) ✔ host recovery reserves original pending attempts exactly once and explicit review preserves unknown outcomes (14726.0361ms) ✔ corrupted model reservation cannot reduce recovered budget or permit a new engine request (12550.5771ms) ✔ engine detection seals the actual CLI, SDK, profile and Node and rejects mixed versions without launch (1096.8016ms) ✔ owned engine probe waits for real ready identity, verifies six native responses and saves immutable task evidence (2051.946ms) ✔ an invalid startup nonce is rejected before initialize and preserves the process close and partial report (1783.3695ms) ✔ SDK identity and session directory mismatches retain raw responses and never become verified (4261.7607ms) ✔ cancel during a pending initialize closes only the owned process without sending an unsupported MCP cancellation (1635.6436ms) ✔ cancel before readiness sends no protocol request and a user retry creates a different process and record (1995.556ms) ✔ source or scratch changes and unexpected model traffic fail after preserving the actual observed result (4527.4536ms) ✔ unowned idle notifications stay quarantined and server requests receive a rejection without granting operations (3342.4715ms) ✔ engine probe environment excludes parent secrets and NODE_OPTIONS while preserving operating system paths (1397.3949ms) ✔ foreign sessions and altered reports or task bindings cannot reuse verified engine evidence (1267.8589ms) ✔ restart never replays an active engine check or stops a historical PID, and keeps prior task state (1416.7569ms) ✔ model argument validation and runtime activity do not grant external engine prompt or project execution (654.6532ms) ✔ a follow-up question receives the actual failed engine check from its original Task and keeps its report unchanged (2399.775ms) ✔ unfixed dependency URLs and invalid integrity are reported without exposing URL credentials or allowing a plan (788.4866ms) ✔ hot upgrade refuses an active dependency operation or evaluation before changing runtime objects (0.5209ms) ✔ read-only environment diagnosis identifies actual runtime, missing dependencies and sealed inputs without installing or running scripts (6160.7626ms) ✔ diagnosis rejects unsupported managers, lock integrity and credential configuration without reading config content (6001.194ms) ✔ approved real npm ci runs only in owned worktree, skips scripts, retains logs and binds later actual tests (10789.794ms) ✔ explicitly allowing lifecycle scripts executes sealed root script and preserves the separate preparation result (8341.6406ms) ✔ stale lockfile, tampered plan/config, cross-session identities and mismatched download sources never dispatch npm (7384.5067ms) ✔ registry failure keeps original npm exit/log and performs no automatic fallback or replay (8675.0625ms) ✔ stopping a hanging dependency download records only its owned process and preserves stopped logs (7825.6254ms) ✔ restart retains unknown preparation, blocks writes and checks, and requires explicit read-only reconciliation (6667.5715ms) ✔ HTTP environment flow survives refresh, requires explicit approval and preserves task and artifact ownership (11928.189ms) ✔ Python environment diagnosis reads declarations and runtime identity without executing project code or installing packages (1093.3307ms) ✔ empty build cache directories keep dependency identity while added executable files invalidate it (978.5454ms) ✔ dependency snapshots reject external links and incomplete budgets instead of claiming environment equivalence (1087.1711ms) ✔ model-requested environment preparation uses the original approval boundary and never treats preparation as passed engineering checks (12957.4771ms) ✔ timeout and offline cache miss retain actual failures and new attempts require newly sealed plans (12701.9241ms) ✔ both evaluation worktrees can approve fixed dependency preparation and different actual dependencies reject candidate execution (27953.3601ms) ✔ evaluation denominator distinguishes attempted unsealed candidates from unexecuted samples (0.3445ms) ✔ evaluation summary retains sealed outcomes and stopped or failed original attempts (0.4539ms) ✔ evaluation contract fixes bounded samples, source paths, checks and aggregate budget (688.5547ms) ✔ same-source workspaces seal real commands before execution and preserve original staged and unstaged bytes (14147.6065ms) ✔ actual original failures and candidate success are both preserved, then adoption requires parent review, checks and Git (56092.7033ms) ✔ candidate cannot shrink tests and changing original test bytes rejects execution before any model request (20263.5727ms) ✔ missing upstream usage remains unknown and does not become a free run or invented cost (36870.5705ms) ✔ missing raw check logs preserve unknown result and original sample denominator (39212.9025ms) ✔ source and model changes never silently replace an existing batch input (3749.6239ms) ✔ model revision change before preparing a fixed sample cannot silently use the replacement connection (4124.291ms) ✔ HTTP evaluations require host token, original fingerprint and session ownership; worker API cannot bypass evaluation flow (16512.7271ms) ✔ sample queue cannot skip or reset attempted samples, and reports freeze further writes (27951.7919ms) ✔ stop cancels only current candidate and preserves partial results instead of replaying a request (24393.914ms) ✔ restart recovers interrupted evaluations as unknown and keeps completed checks and plan identity (29216.9587ms) ✔ JSON migration survives process exit before commit and retains exact legacy backups (2208.3387ms) ✔ process exit after_state rolls back state, event and outbox together (1781.6833ms) ✔ process exit after_event rolls back state, event and outbox together (2479.2113ms) ✔ crash recovery at accepted retains execution identity and never replays side effects (2040.4234ms) ✔ crash recovery at queued retains execution identity and never replays side effects (2548.0301ms) ✔ crash recovery at approval_wait retains execution identity and never replays side effects (2785.7936ms) ✔ crash recovery at approved retains execution identity and never replays side effects (2797.2788ms) ✔ crash recovery at effect_unrecorded retains execution identity and never replays side effects (3207.2363ms) ✔ crash recovery at result_committed retains execution identity and never replays side effects (3276.2616ms) ✔ crash recovery at before_next_model retains execution identity and never replays side effects (3813.4398ms) ✔ state and event failures publish no success and preserve the last committed memory snapshot (864.4882ms) ✔ tool result artifacts and event cursors retain project ownership across restart (2125.6299ms) ✔ token budgets pause before tools execute and continuation keeps original task and accumulated usage (1981.8572ms) ✔ a time budget interrupts an unresponsive model and persists a resumable run (1154.2296ms) ✔ a corrupted large result artifact cannot become a recovered success (1000.6833ms) ✔ hard-link aliases cannot expose protected bytes through reads, proposals, inventory or automatic project rules (828.7666ms) ✔ ambiguous project hard links are explicit and ordinary independent file copies remain readable (622.5921ms) ✔ current-file HTTP receipts bind Task/Run and bytes, persist on restart and end without altering original runs or user Git data (4546.3801ms) ✔ file targets exclude referenced inputs and prohibited subsequent modifications (1.475ms) ✔ full actual read and existing receipt complete a no-test file task without fabricating changes, checks or commits (2726.2716ms) ✔ partial pages, partial lines, wrong hashes and failed reads cannot be adopted; contiguous full paging can (1287.0461ms) ✔ user edits and new requirements invalidate receipts, and a second task must explicitly adopt a source in its own project/session (992.4479ms) ✔ adopting an unchanged file never satisfies an explicit modification or a new modification instruction (1116.3624ms) ✔ required checks and stale passing flags cannot be bypassed by current file receipts (824.4809ms) ✔ results plan proofs bind current file receipts while changes proofs still demand real applied change sets (825.99ms) ✔ duplicate receipts cannot create progress indefinitely or reset no-progress reads (3107.5564ms) ✔ manual completion preserves the failed original run and refuses unknown operations (764.6375ms) ✔ completion refuses active work, files changed during inspection and a newly selected task (860.2224ms) ✔ unrelated files cannot stand in for named output paths, nor can a late apply satisfy newer instructions with an old proposal (979.0161ms) ✔ delivery and mobile projections distinguish adopted bytes, new changes, no tests and no commit; mutations cannot rewrite receipts (1185.0765ms) ✔ proposals preserve project bytes until review; apply backs up the exact original bytes (714.7979ms) ✔ new nested files are created only on apply and unchanged content creates no proposal (628.7276ms) ✔ stale reads and changed baselines are rejected without overwriting external edits (759.033ms) ✔ file paths reject traversal, absolute paths, credentials and outside junctions (759.9338ms) ✔ read and search expose bounded text scopes and reject non UTF-8 / oversized files (700.7874ms) ✔ project-local junctions cannot expose or create protected Git and credential files (583.4485ms) ✔ Windows trailing-dot and trailing-space path aliases cannot bypass protected paths (537.5872ms) ✔ a starter is copied into user data once; restarting from another program version preserves existing projects and editable bytes (1199.6793ms) ✔ startup does not insert a demo or rewrite/reorder an existing user project registry (821.7477ms) ✔ PNG bytes are fully decoded, metadata-free and bounded; JPEG EXIF orientation is applied to the sent PNG (103.0655ms) ✔ forged type, short headers, CRC damage, incomplete chunks, trailing payloads, animation and dimension bombs reject (24.8421ms) ✔ draft image ownership, withdrawal, reload and atomic Task binding retain original SHA and original bytes (1189.341ms) ✔ a failed session transaction restores draft status and cannot leave a Run consuming the attachment (683.4283ms) ✔ draft limits, duplicate selection, source changes and cross-Task association reject before dispatch (1124.6534ms) ✔ demo and legacy paths refuse image dispatch without creating a Task or consuming the draft (667.806ms) ✔ chat-completions sends one owned image per actual request, pairs original tools and verifies exact sealed bytes (2467.5554ms) ✔ responses sends one owned image per actual request, pairs original tools and verifies exact sealed bytes (2769.4587ms) ✔ mixed pool excludes unknown vision before lease and sends zero HTTP to the incompatible member (1733.9437ms) ✔ vision probe requires correct randomized pixel order and evidence expires when the connection revision changes (927.0969ms) ✔ HTTP upload/read/withdraw uses session authorization, returns byte SHA and refuses unknown ownership (672.9452ms) ✔ image accounting excludes Base64 as text but seals every original request byte and reserves image input separately (3.0714ms) ✔ actual model comes from owned consumption, excludes other Tasks and deduplicates the current Run (1.4817ms) ✔ mixed routes retain each original model/protocol/outcome and unknown use is never reported as zero (0.1357ms) ✔ zero dispatch, missing ledger, missing model and truncated Run history never borrow configured model identity (0.202ms) ✔ repeated attempt identities across Runs and contradictory usage stay explicit instead of inflating successes (0.1044ms) ✔ observed route never certifies returned identity and retains mismatches and original source (0.1426ms) ✔ directory picker reads one level, preserves empty folders, skips credential and linked folders (676.8417ms) ✔ HTML preview serves encoded names and relative modules on an isolated loopback origin; closing revokes it (719.4371ms) ✔ HTML preview refuses credentials, hidden inventories, traversal, external junctions, hardlinks and unsupported file types (641.9154ms) ✔ runtime shutdown closes local preview ports and pending starts cannot escape shutdown (615.7329ms) ✔ idle file previews expire and release their loopback servers (706.7203ms) ✔ directory and preview HTTP controls require workbench authentication; project pages cannot read bootstrap (2118.3926ms) ✔ MCP registration stays disabled until an explicit real stdio probe and schema-based tool selection, with actual process cleanup (2157.8802ms) ✔ MCP probe holds the project operation lock until its owned process and report are settled (1614.8247ms) ✔ MCP validates the declared JSON schema subset and rejects unsupported schemas before external dispatch (0.6464ms) ✔ actual model MCP call approves long schema-valid arguments, validates the sent catalog and stores one original result without leaking environment secrets (4462.1937ms) ✔ denying an MCP call never starts another process or executes the external tool (3601.5581ms) ✔ actual handshake drift rejects tools/call even if program entry hash is unchanged (4521.1109ms) ✔ cancelled dispatched external side effect stays unknown, closes its owned process and cannot replay on reopen (3914.7119ms) ✔ changed executable input, foreign project IDs and corrupted original contracts block enable and dispatch (1690.8564ms) ✔ registered MCP HTTP API uses fixed execution contracts and project-owned probe records with no model requests (3490.7819ms) ✔ unknown MCP pauses later tool calls and all project writers; fingerprinted acknowledgement clears only the review gates and explicit resume does not replay (6162.2155ms) ✔ MCP acknowledgement persistence failure restores both job and tool review gates; corrupt evidence cannot be acknowledged (4507.7334ms) ✔ actual host exit after an external effect restores unknown jobs without a fabricated report, allows source-bound review and never replays the call (4239.2371ms) ✔ changed enabled MCP programs can be explicitly disabled without launching or trusting their changed source (1784.7826ms) ✔ stdio preserves split UTF-8 frames and final replies drained after process exit (1758.0835ms) ✔ server-initiated model requests are refused instead of inheriting provider privileges (834.9918ms) ✔ invalid or over-budget protocol invalid closes only its owned process and is not retried (844.1788ms) ✔ invalid or over-budget protocol oversized closes only its owned process and is not retried (829.044ms) ✔ cancel and timeout do not resend a dispatched request and pre-cancel has no request trace (856.4381ms) ✔ startup failure remains undispatched and has no process PID to terminate (549.1342ms) ✔ REA launch filters a fixture credential canary and rejects changed entry identity before a new connection (862.1811ms) ✔ schema1 connections remain readable without inventing observed capabilities or migrating original state (680.6309ms) ✔ protocol, unsupported capabilities and stream choice persist; client claims cannot forge verified evidence (712.4425ms) ✔ ability evidence belongs to exact connection and credential revision; unknown and unsupported reject required routing (0.2836ms) ✔ invalid declarations and inconsistent stream settings reject before registry, revision and credentials mutate (693.9453ms) ✔ model registry keeps original credentials/endpoint intact, rejects stale edits, and forgets only volatile keys on restart (870.0487ms) ✔ config import previews exact revisions, keeps secrets volatile, syncs only when enabled and preserves manual edits on conflicts (857.8976ms) ✔ real HTTP pool fails over 429 to another model and never repeats completed tools; attempts seal actual request bytes and Run identity (3992.3893ms) ✔ pool rotation and weights select members on new Runs; disabled and missing credentials are excluded without requests (5157.0418ms) ✔ 401 blocks the credential revision; updating the key restores availability without leaking error echoes (3039.4895ms) ✔ queued request cancellation and in-flight disable release only owned leases and do not dispatch another HTTP request (1466.0748ms) ✔ compiled model-center HTTP API persists public profiles, blocks forged origins, and local operations never invoke a model (2653.281ms) ✔ model registry disk failures restore routing, pool profiles and credential references without partial updates (2142.8302ms) ✔ model selection and saved configuration commit together and failures preserve the original mode and endpoint (2337.7713ms) ✔ configuration imports roll back the entire batch and retain one usable preview after validation or commit failure (928.1413ms) ✔ HTTP quota retains native code/type/param/request id, bounded original and allowlisted headers, never repeats that connection (2757.7124ms) ✔ Retry-After above 90 seconds remains exact and a short local budget cannot trigger an early retry (1.8288ms) ✔ Retry-After date, expired date, huge value and invalid header retain explicit semantics (1.1279ms) ✔ unclassified 429 remains unspecified and every bounded actual retry retains its own original error (2987.9141ms) ✔ oversized HTTP error saves exactly the bounded prefix and labels its incomplete range (2310.7477ms) ✔ Chat SSE error preserves upstream fields, original failed event, partial tool and reported usage without executing or retrying (2231.3688ms) ✔ Responses failed SSE retains original event/response identity and usage (1996.0481ms) ✔ Responses incomplete SSE retains original event/response identity and usage (1677.4627ms) ✔ unsupported successful Responses output retains original JSON and usage with a local protocol origin (1540.3835ms) ✔ HTTP 200 native error is still an upstream failure rather than a fabricated successful answer (0.8044ms) ✔ JSON-escaped credential echoes are redacted in the original preview and capture as well as parsed fields (0.5845ms) ✔ quota failover uses another independent pool member and never replays a failed stream (2459.695ms) ✔ native parameters are isolated by pool member, preserve Responses state and bind the original output reservation (0.3387ms) ✔ unsupported native fields and invalid values reject before dispatch instead of silently dropping or overriding constraints (0.3054ms) ✔ native settings for legacy and registered profiles persist without changing credentials or old records and reject stale edits (1083.4514ms) ✔ compiled native settings API saves/restores exact parameters and rejects reserved fields without mutating the profile (2553.9344ms) ✔ legacy connection test sends its saved native fields and selected output limit field without changing the configuration (2378.5449ms) ✔ identity distinguishes missing, invalid, mixed, exact and explicitly declared aliases without guessing (1.2584ms) ✔ chat-completions JSON rejects a different model before tool dispatch, retaining source and reported usage (2586.562ms) ✔ chat-completions SSE rejects a different model before tool dispatch, retaining source and reported usage (1829.4412ms) ✔ responses JSON rejects a different model before tool dispatch, retaining source and reported usage (1672.3965ms) ✔ responses SSE rejects a different model before tool dispatch, retaining source and reported usage (1777.2455ms) ✔ record policy accepts missing identity but keeps it explicit, and strict declared mapping is frozen in the consumption (1.2107ms) ✔ mixed streamed identities cannot pass strict mode, and interrupted streams retain observed identifiers without authenticating them (1.1646ms) ✔ identity refusal never fails over to another pool member and raw response model cannot leak known credentials (1607.0677ms) ✔ strict refusal remains visible even when the same original response exhausts the budget (1575.7549ms) ✔ strict identity crosses the controlled engine bridge with its original source and zero tool dispatch (11246.0751ms) ✔ frozen direct connection policy and delivery reports preserve response identity without filling historical gaps (2381.5068ms) ✔ saved engine budgets verify all Chat/Responses route variants and reject a mismatched output allowance (738.1008ms) ✔ owned CC Switch SQLite imports the selected TOML Responses wire_api and leaves unknown clients untouched (1082.5056ms) ✔ Responses executes one original tool through Runtime, carries encrypted Items, reports raw usage and seals exact wire bytes (3168.4963ms) ✔ mixed-protocol pool keeps both actual variants and changes route on 429 without re-executing completed tools (3190.3372ms) ▶ Responses incomplete, failed, missing call identity and unsupported hosted outputs never dispatch a tool ✔ Reject {"body":{"id":"resp_owned","status":"incomplete","output":[{"type":"function_call","id":"fc_partial" (1668.0404ms) ✔ Reject {"body":{"id":"resp_owned","status":"failed","output":[{"type":"function_call","id":"fc_failed","cal (1739.2677ms) ✔ Reject {"body":{"id":"resp_owned","status":"completed","output":[{"type":"function_call","id":"fc_bad","cal (1758.462ms) ✔ Reject {"body":{"id":"resp_owned","status":"completed","output":[{"type":"computer_call","id":"not-supporte (1783.6105ms) ✔ Responses incomplete, failed, missing call identity and unsupported hosted outputs never dispatch a tool (6950.1766ms) ▶ invalid JSON in a nonstream Chat or Responses tool group prevents every tool and pool replay ✔ chat-completions (1568.7734ms) ✔ responses (1763.3977ms) ✔ invalid JSON in a nonstream Chat or Responses tool group prevents every tool and pool replay (3332.5375ms) ✔ wire adapter preserves tool IDs but never passes Responses-private Items to Chat or another endpoint (0.2045ms) ✔ connection JSON import accepts explicit Responses protocol and continues to reject unknown protocols without reading client sessions (547.6983ms) ✔ reservation uses actual known usage from the same route, never lowers estimates or rewrites original receipts (1.7136ms) ✔ reservation verification binds original Task/Run consumption and rejects changed usage or foreign provenance (0.6227ms) ✔ known high input calibrates the next request and stops before dispatch while preserving the completed read (1875.1349ms) ✔ one reported overrun stops legacy requests and keeps actual use and excess (1373.5475ms) ✔ one reported overrun stops pool requests and keeps actual use and excess (1699.8085ms) ✔ one reported overrun stops connection requests and keeps actual use and excess (2075.2234ms) ✔ reported usage below the limit prevents an unaffordable legacy retry before another dispatch (1934.9407ms) ✔ reported usage below the limit prevents an unaffordable pool retry before another dispatch (2108.0469ms) ✔ native bridge seals calibrated refusal with original consumption and preserves versioned request verification (15511.568ms) ✔ original R v1/v2 stay verifiable while v3 excludes the isolated input outlier and keeps the ledger intact (4.0072ms) ✔ v3 never excludes a repeated high ratio, the latest receipt, or a short history (3.228ms) ✔ reservation proofs reject reordered latest high receipts, altered history and duplicates while old snapshots survive later consumption (3.0938ms) ✔ input-ratio exclusion preserves independently reported output maxima without rewriting historical v2 (1.7914ms) ✔ capability routing skips text-only members before acquiring a lease and records exact refusal in the original Run (1550.1513ms) ✔ all incompatible members reject with no network request, no lease, preserved Task and explicit non-dispatched ledger (1449.5114ms) ✔ explicit synthetic tool probe verifies only actual matching call, never executes it and loses verification on revision changes (743.2978ms) ✔ plain text probe cannot verify tools; wrong tool nonce and client supplied evidence cannot claim verified support (844.0315ms) ✔ a disabled connection invalidates an in-flight probe even when toggle mutates its original registry object (817.3693ms) ✔ selected connection owns the actual endpoint, credential, wire model and mismatch even when legacy is configured for GPT (2455.4261ms) ✔ budget-only API and demo/real selection preserve routes, reject injected endpoint fields and publish authoritative selection without keys (2887.9622ms) ✔ another session cannot change default model, mode or budgets while a request is active (4600.7444ms) ✔ HTTP download preserves original malformed SSE bytes and artifact SHA instead of decoding invalid UTF8 (2446.0443ms) ✔ terminated SSE adapts empty optional tool arguments, retains the exact wire bytes and executes the validated batch once (4065.3496ms) ✔ reported SSE usage survives rejected required arguments "" without dispatching any tool (2499.4843ms) ✔ reported SSE usage survives rejected required arguments "{\"path\":" without dispatching any tool (2263.7865ms) ✔ reported SSE usage survives rejected required arguments "null" without dispatching any tool (2415.6023ms) ✔ a stream without DONE retains reported usage but cannot adapt or execute even an optional tool (2422.5901ms) ✔ a timeout after reasoning-only fragments creates one unknown receipt and never invents reported tokens (2383.3451ms) ✔ real Chat SSE carries UTF8 and split tool JSON; Runtime executes only after DONE, stores partial progress and original usage (3860.9375ms) ✔ Responses typed SSE validates original item identity, terminal data, raw usage and tool result pair (3903.2606ms) ✔ incomplete network stream retains unknown attempt and never automatically switches pool or dispatches fragment tools (2817.2902ms) ✔ stopping an active HTTP stream closes only original request and keeps partial text without executing tools (2325.5745ms) ✔ explicit streaming probe uses actual SSE despite unknown declaration and binds its own original attempt (1004.2546ms) ✔ coalesced initial SSE events publish the latest partial state even when the upstream stays idle (2837.7373ms) ✔ stream timeout closes the original HTTP response, preserves redacted capture and never retries (1287.9515ms) ✔ stream idle timeout preserves incomplete tool arguments, records unknown use and never fails over (1577.0492ms) ✔ SSE accepts CR-only final event boundaries and retains exact invalid UTF8 bytes without dispatch (1.2335ms) ▶ stream parser rejects truncation, finish length, malformed JSON, duplicate sequence, changed final item and oversized data ✔ reject 0 (0.2368ms) ✔ reject 1 (0.1866ms) ✔ reject 2 (0.1512ms) ✔ reject 3 (4.8353ms) ✔ reject 4 (0.1407ms) ✔ reject 5 (0.3746ms) ✔ stream parser rejects truncation, finish length, malformed JSON, duplicate sequence, changed final item and oversized data (6.422ms) ✔ compiled host retains interrupted HTTP 200 and explicitly resumes after the recorded cooldown (10648.7638ms) ✔ compiled host retains interrupted HTTP 503 and explicitly resumes after the recorded cooldown (10131.5539ms) ✔ owned IPv4 service works after a failed IPv6 candidate without a second model POST or global DNS change (25.6583ms) ✔ environment HTTP proxy tunnels the owned request and NO_PROXY sends a separate direct request (10.5225ms) ✔ original connect timeout has a bounded nested cause and does not masquerade as response generation timeout (0.7054ms) ✔ POST accepted then reset is unknown, is charged once and never automatically retried (3.4461ms) ✔ partial HTTP bodies retain original request id and bytes, and an unknown pool attempt never switches model (2871.5972ms) ✔ response wait timeout after an accepted POST is unknown; cancellation closes the owned socket and permits explicit reuse (148.1414ms) ✔ older configurations receive bounded wait defaults without changing the original object (1.2348ms) ✔ active stream survives sixty seconds; byte activity resets idle but cannot extend total deadline (0.7963ms) ✔ silent stream has its own accurate deadline; caller cancellation and cleanup preserve their cause (1.3956ms) ✔ waiting for headers is bounded independently of stream-idle time (0.2305ms) ✔ wait settings persist, pin new runs and leave model pool, credentials and old run snapshots intact (1442.2924ms) ✔ plain endpoint checks only TCP, records family coverage and never reads model credentials or writes routing/health/attempts (1404.7509ms) ✔ local DNS failure is separate from TCP/TLS and arbitrary input, stale revision and TLS overrides are rejected (1158.6971ms) ✔ refused TCP socket preserves its stage and cannot masquerade as a TLS or API response (1093.9866ms) ✔ TLS validation succeeds for a trusted owned certificate and rejects the same untrusted certificate (1865.5754ms) ✔ a TCP peer that never answers TLS is bounded, and cancellation waits for owned sockets to close (1933.8852ms) ✔ a stalled lookup is cancellable and the overall deadline bounds DNS without late socket work (1132.5663ms) ✔ dual-stack fallback and explicit family results remain separate and never change global DNS defaults (1216.3385ms) ✔ proxy settings are reported as scope only; diagnostics do not contact the proxy or leak its credentials (1156.7297ms) ✔ history is bounded, survives reopening, marks interrupted work and reports changed revisions without rewriting frozen targets (1653.0524ms) ✔ HTTP API, host activity, configuration edits, model start and upgrade honor the same active diagnostic and stop operation (1253.1827ms) ✔ provider shutdown cancels diagnostics and permits explicit reuse after a failed host stop (1229.1274ms) ✔ unchanged proposals cannot hide repeated original-log reads or identical session inspections (8295.7799ms) ✔ reusing a verified proof never resets identical-state detection or creates a new proof (7769.976ms) ✔ coverage merges adjacent pages, separates changed sources and never claims old Run coverage in a continuation (1.0514ms) ✔ paged inspection finds a corrupted old artifact outside the recent fifty and retains original evidence without model/tool execution (5040.3493ms) ✔ inspection rejects ambiguous/out-of-range pagination and reports unknown tool/task identity without resolving it (834.516ms) ✔ original source locks all 60 byte-identical files including modules outside the nine-file prompt experiment (950.4792ms) ✔ a new original executable or missing license cannot silently join the fixed trusted-source profile (921.5789ms) ✔ original mode has a separate source identity and does not accept retry or privilege fields (966.9943ms) ✔ actual original model input requires the complete postPrompt once in system context and cannot borrow a user or tool copy (0.304ms) ✔ Windows directory permissions normalize overlapping grants without allowing an unrelated sibling (836.5044ms) ✔ durability broker rejects foreign authentication, outside paths and wrong file identity before real fsync (937.2809ms) ✔ exact patches preserve BOM/CRLF, unicode and all unselected bytes and back up the original on apply (1038.6306ms) ✔ several patches are matched against one original snapshot in any input order and keep LF (904.4896ms) ✔ mixed line endings require exact anchors and preserve untouched formatting (893.8319ms) ✔ ambiguous, overlapping, missing and malformed patch anchors never create a proposal or write (816.891ms) ✔ stale original hashes and changes before apply preserve owner edits and mark conflicts (818.0299ms) ✔ no-op BOM patches/full content produce no change and replacements cannot split Unicode bytes (704.9442ms) ✔ real owned Vite starts only after approval, records OS ownership and live changes, and stops its tree (14060.433ms) ✔ occupied ports and approval-time conflicts never take over an unrelated listener (9125.5557ms) ✔ background preview final-log failure retains unknown evidence without an unhandled startup promise (12136.6281ms) ✔ stale source, tampered plans, wrong ownership and arbitrary scripts cannot dispatch a preview (8086.0124ms) ✔ restart retains original unknown preview, blocks mutation, reconciles absent identity without replay (7550.1986ms) ✔ an actual shell command keeps Node and UTF-8 output while omitting host-only credentials (874.9472ms) ✔ real npm tests and build do not inherit arbitrary host credentials or Node injection settings (3215.4354ms) ✔ real Python tests and compilation keep their sealed inputs without host credentials (2449.1507ms) ✔ a process exiting between the Windows snapshot and owner query is absent only after a fresh query (1877.9231ms) ✔ owner observation errors remain errors while a process is present, replaced, or cannot be rechecked (1572.797ms) ✔ cache controls require desktop authority, exact inputs and owned project identity while preserving task evidence (2211.9683ms) ✔ cache enablement and derived reuse survive a real host restart without becoming task recovery state (4840.3977ms) ✔ persistent parsing is project/SHA/config bound and produces identical semantic indexes after restart (1717.4271ms) 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 'main.ts', LF will be replaced by CRLF the next time Git touches it warning: in the working copy of 'value.ts', LF will be replaced by CRLF the next time Git touches it ✔ same-size restored-mtime changes, additions, renames and deletions cannot reuse stale entries or import targets (1217.238ms) Switched to a new branch 'other' ✔ branch and commit changes invalidate even byte-identical sources (3181.9833ms) ✔ corrupt, malformed and schema-invalid records fall back to fresh parsing (1580.9962ms) ✔ disable and clear persist without touching sources, sessions, configuration or original artifacts (1549.4183ms) ✔ cache capacity bounds disk and memory by bytes and LRU project count (1199.5377ms) ✔ incomplete scans do not load or save cached entries; unavailable files leave the current result scope (1364.1651ms) ✔ project and cache junctions cannot expose or delete an outside directory (818.3606ms) ✔ concurrent file changes refuse a snapshot and clear during pending work cannot repopulate it (999.5056ms) ✔ file access mode changes invalidate parsed material even when bytes are unchanged (833.0544ms) ✔ tracked directory replaced by a junction invalidates access scope and never uses its old cache (2608.9615ms) ✔ JS/TS/React AST indexes actual declarations, exports, imports and verified source lines (656.6544ms) ✔ default expressions, named default exports and declared functions retain their AST evidence (602.8001ms) ✔ incremental refresh invalidates changed/deleted symbols and recalculates reused import targets (1143.205ms) ✔ Python candidates exclude comments and multiline strings and declare limited evidence (726.2277ms) ✔ indexes isolate project identity and root and evict bounded caches without sharing results (899.5382ms) ✔ Git ignore, credential paths and links stay outside the index and no project files are executed (1376.2968ms) ✔ malformed, binary, non-UTF8 and bounded files expose partial coverage while good files remain searchable (897.2575ms) ✔ index query rejects invalid limits/types and does not imply reference or alias resolution (843.7805ms) ✔ read-only engineering inspection disables configured Git clean filters and preserves index bytes (2257.1141ms) ✔ Python multiline declarations retain candidate scope without interpreting expressions as declarations (732.0391ms) ✔ TypeScript module candidates support import-equals and MTS/CTS while opaque imports stay explicit (792.433ms) ✔ actual HTTP requests reuse only this project source-checked memory as redacted data and preserve sent evidence after archive/restart (2082.4495ms) ✔ actual HTTP request excludes archived, timed-out and foreign-project memories (4151.6651ms) ✔ actual next HTTP request removes expired memory after a tool observes a source change without replaying that tool (4378.6752ms) ✔ incomplete inventory and excluded-file SHA changes cannot enter actual HTTP memory context or inspect_session snapshots (3503.0035ms) ✔ corrupt original artifact expires its memory and cannot reach an HTTP model request (2641.502ms) ✔ memory JSON is bounded in escaped UTF-8 bytes and entry count, skips oversized entries and retains complete multilingual values (4013.0023ms) ✔ optional memory yields to the complete goal and original code evidence under the whole-request budget (1219.7407ms) ✔ changed sealed memory artifacts or binding fields fail verification while the original request remains immutable (1838.3875ms) ✔ a malformed project declaration omits optional memory while the repair Task still reaches the owned HTTP service (1720.6496ms) ✔ an expiration persistence failure prevents dispatch and is not hidden as an optional scan failure (1378.6287ms) ✔ project memories persist across restart, keep source provenance, and are visible across tasks (1181.8218ms) ✔ source changes expire current project memories and keep the original evidence (1141.2304ms) ✔ incomplete source scans mark current memories unverified without expiring or reusing them (1234.5631ms) ✔ explicit file provenance is rechecked even when the inventory excludes that path (1565.9326ms) ✔ execution snapshots do not expose project memories before their source scope is validated (1043.5011ms) ✔ project memory records stay isolated and archive requires the owning session (1157.4834ms) ✔ project memory HTTP API exposes read-only source status and explicit writes (3059.4812ms) ✔ two Purge source versions lock nine original inputs without executing JavaScript or postinstall and remain disabled by default (944.9651ms) ✔ incomplete, unsupported, relative and oversized sources cannot enable the experiment (914.318ms) ✔ invalid source fingerprints and invented privilege or retry fields fail before any parent Task is created (2310.4166ms) ✔ source edits between detection and plan are rejected without creating a parent Task (1833.3669ms) ✔ source edits after sealing cannot launch a child and do not reset its cumulative budget (8974.0795ms) ✔ literal experiment text is sealed into the parent, worker and each new Run while actual host approvals and saved results remain required (30396.0479ms) ✔ source changes after a model response stop before any native tool is accepted, retaining the actual usage once (11953.6655ms) ✔ source changes while approval is pending invalidate the old permission before an engineering check starts (16492.5991ms) ✔ altered text, interpolation, section order and policy flags cannot reuse a sealed experiment (2189.8219ms) ✔ the observed SDK system-message layout and the explicit system field both require exactly one effective section, never a user or tool copy (2160.4245ms) ✔ a missing or duplicated experimental section in the actual input is rejected before any model dispatch (21787.4063ms) ✔ experimental prompt overhead still participates in the original complete-input budget guard (11676.6851ms) ✔ source detection HTTP remains local-authenticated and read-only with no configured model requests (2293.6097ms) ✔ Python HTTP plans keep npm defaults, require owned approval and preserve actual results across restart (10675.3223ms) ✔ an isolated three-module Python Task repairs, checks and delivers a source-bound Git report without extra model calls (41157.2387ms) ✔ Python static inspection does not import project code or probe a project interpreter before approval (1395.9224ms) ✔ a real three-module Python failure, reviewed repairs, tests and compilation retain exact structured evidence (4623.0812ms) ✔ unittest counts repeated execution ordinals and subtest skips rather than deduplicating display names (3379.8579ms) ✔ unittest empty, skipped, expected failure and class initialization errors never verify a task (8438.6354ms) ✔ stdout claims cannot replace actual unittest results and an early exit cannot forge a successful collector (3090.7195ms) ✔ compilation reads actual Python source bytes without executing top-level project code (2809.4345ms) ✔ Node and Python mixed projects retain the existing npm default and require an explicit Python selection (2977.5793ms) ✔ changed project interpreters and forged Python plan bindings execute no check (1666.7259ms) ✔ malformed TOML and invalid Python test paths are reported without expanding npm check scope (1424.9374ms) ✔ stopping a real Python check terminates its child and keeps the completed result as cancelled (2075.4776ms) ✔ real pytest pass, failure, skip, collection errors and teardown errors use collected case outcomes (17272.1474ms) ✔ a real project virtual environment is detected statically and its actual version is recorded only during the approved check (8639.192ms) ✔ a missing interpreter and repeated pytest item objects cannot create a false verification pass (2495.702ms) ✔ a mixed turn with new bytes and three repeated modules advances, while repeat-only and independent no-op guards still pause (1.0053ms) ✔ runtime sends the next real HTTP request after a W-shaped mixed-progress turn instead of stopping on its trailing repeats (4852.3591ms) ✔ continuation reads advance through historical pages without a false pause while repeated current-Run pages still pause (1.2567ms) ✔ a new tool page replaces old latest-feedback space and still obeys its escaped protocol budget (0.3864ms) ✔ budgeted artifact pages count escaped protocol bytes and reconstruct original multilingual JSON without losing characters (9.0557ms) ✔ known artifact coverage detects semantic repeats despite changing limits, and advancing pages or direct evidence restore progress (0.3169ms) ✔ actual large artifact reads are present in the next bounded HTTP request and next_offset progresses without redoing tools (5444.124ms) ✔ a real repeated artifact loop pauses the same Task after recorded results, survives reload and only resumes explicitly (6788.9092ms) ✔ actual HTTP continuation after reload sends distinct historically covered pages without false no_progress or replay (10198.0262ms) ✔ continuations distinguish unread source, later original line ranges and already returned result wrappers (629.4558ms) ✔ pure-file batch returns unused small-page space while mixed and historical allocation stays conservative (0.3294ms) ✔ v3 pools mixed file/log reads, serves small sources first and leaves historical/mutating batches unchanged (0.2947ms) ✔ v3 W-shaped mixed reading returns complete small modules and substantially larger first test pages in the next bounded request (4928.229ms) ✔ observed upstream high input pauses before a new request, preserves reported usage and distinguishes the limiting ledger (981.0976ms) ✔ chat-completions: complete actual Unicode/CRLF pages and small source survive the next bounded wire request (5551.9692ms) ✔ responses: complete actual Unicode/CRLF pages and small source survive the next bounded wire request (4974.6677ms) ✔ anthropic-messages: complete actual Unicode/CRLF pages and small source survive the next bounded wire request (5299.0941ms) ✔ an installed app outside the source drive resolves registered SDK and Purge through actual input verification (1048.2117ms) ✔ standalone registration resolves the upstream metadata directory and missing upstream falls back to the other registered source (938.3799ms) ✔ registered hints keep explicit roots, environment and existing app-relative candidates in their original priority (1289.0962ms) ✔ unknown IDs, duplicate registered IDs, wrong versions and relative or inconsistent metadata are not lookup candidates (895.508ms) ✔ oversized, malformed, invalid UTF-8 and non-file registry inputs cannot silently enable a source (852.3122ms) ✔ a linked registry directory is rejected without reading or deleting its external source (1029.434ms) ✔ matching research declarations do not replace actual installed version and critical-source checks (895.6027ms) ✔ runtime guides map archived evidence while all original reports and successful/failed source bytes remain independently verifiable (1044.3827ms) ✔ missing required research reports fail packaging rather than creating a successful placeholder (904.6703ms) ✔ full verification reports only verified bytes, complete declared scope and actual directory inventory; changed/deleted files still fail (975.9704ms) ✔ preflight and activation continue one monotonic operation bound to installer hash; terminal replay and changed origin are rejected (8615.533ms) ✔ preflight failures and fixed-candidate mismatches retain the actual reason without publishing a pointer (5071.7297ms) ✔ cancel finishes only the original ready preparation, preserves data and pointer and refuses replay (7287.6464ms) ✔ the installation lock rejects concurrent same-id operations before they can overwrite original progress (4139.6844ms) ✔ original post-publication failure is linked to its rollback receipt and every original byte survives (4294.1131ms) ✔ a candidate mutation at the publication boundary still fails even after an earlier successful full validation (8689.9842ms) ✔ public progress marks a lost process unknown without editing its original record or claiming completion (4975.738ms) ✔ throwing console/UI progress observers cannot decide the installation result (3632.5836ms) ✔ a progress-only write failure after committed publication returns the authoritative receipt and a warning (4064.3176ms) ✔ browser and MCP approvals use actual tool categories; mobile exposes only category while keeping original fingerprint (573.3761ms) ✔ the compiled runtime version constant tracks package.json so a release bump cannot ship a stale served version (7.9918ms) ✔ context APIs retain exact request identity, paginate archived groups and reject foreign sessions/runs (4272.6733ms) ✔ delivery reports verify request body, sealed plan and original archived bytes, and show corruption as needing verification (2064.8178ms) ✔ native seed framing reserves capacity before optional history without changing the selected window (1042.3673ms) ✔ direct code evidence remains selected while recent historical-index browsing is archived within the complete request budget (1981.7135ms) ✔ one long user turn becomes bounded complete groups with exact UTF-8 history and recent JSON arguments (3424.7413ms) ✔ tool definitions, exact system bytes, output reserve and remaining Run/Task budget are accounted before requests (901.0127ms) ✔ oversized required user instructions pause intact and can be explicitly resumed after increasing the window (1476.7579ms) ✔ all task instructions, constraints and acceptance remain in the runtime system instead of an eight-item or byte prefix (1880.0525ms) ✔ a newest manageable multi-tool group stays exact even beyond the working target; oversized earlier groups are archived (1120.5505ms) ✔ invalid orphan result is rejected before archive or network preparation (635.5709ms) ✔ invalid missing result is rejected before archive or network preparation (680.8761ms) ✔ invalid wrong result is rejected before archive or network preparation (716.1276ms) ✔ invalid duplicate call is rejected before archive or network preparation (700.3659ms) ✔ a group larger than the artifact limit uses verified Unicode-safe chunks without replacing tool JSON (1428.2817ms) ✔ changed or missing archived evidence stops reuse and is never silently recreated (864.8688ms) ✔ transaction failure prevents publishing a plan and any corresponding model request (1013.9851ms) ✔ the Provider rejects body mutation after a sealed plan before calling HTTP (782.7565ms) ✔ HTTP failure and explicit same-Task resume reuse exact completed results and request hashes without repeating a command (8568.4816ms) ✔ a real process exit after plan persistence marks interruption and never issues an automatic model request (1396.465ms) ✔ local draft preserves exact original ranges and duplicate source instances without inferring outputs or accepting facts (986.2448ms) ✔ review keeps unknown blocking, references distinct from output and boundaries distinct from compliance evidence (1036.1705ms) ✔ draft adoption must account for every original range and does not fill paths from a reference (853.9637ms) ✔ source and requirement changes, other Tasks and tampered draft bytes are rejected without losing originals (2011.8871ms) ✔ partial explicit source review leaves the other source unreviewed while old file evidence remains valid (1823.9179ms) ✔ bounded draft does not claim full source coverage and cannot acknowledge an omitted remainder (624.628ms) ✔ draft HTTP preserves its original artifact across restart and rejects stale scope and source versions (4567.3854ms) ✔ requirements HTTP binds selection, enforces optimistic revisions, restores original proofs and leaves provider history untouched (5307.6803ms) ✔ research evidence binds actual original bytes; model completion is not user acceptance (1678.5247ms) ✔ same text with distinct user source instances stays distinct; an appended requirement only invalidates its own revision (2345.7678ms) ✔ foreign project, old Task and fabricated evidence strings are rejected with preserved failed proofs (3474.5465ms) ✔ source changes invalidate prior proofs and cannot certify a read that was already stale at binding (2668.2097ms) ✔ missing or tampered artifacts never become evidence and original failures remain after recovery (1902.2645ms) ✔ checks omitted by explicit agreement are reported as not executed and never passed (796.8691ms) ✔ user confirmation stays separate from engineering evidence, is bounded to current inputs and cannot be silently converted (1522.0988ms) ✔ existing multi-file results keep provenance and invalidate only evidence whose files changed (1867.1021ms) ✔ unrelated additions preserve the checked file facts but remain visible as unreviewed user sources (1464.7048ms) ✔ actual Git preview and publication recheck changed requirement bindings without touching HEAD or the owner index (8277.8285ms) ✔ no-test file application and manual completion keep pending user requirements for explicit review (2111.8582ms) ✔ revision races, forged source ranges, evidence fields and deletion cannot overwrite sealed requirements (1467.5399ms) ✔ requirement binding prevents completion and Git acceptance after a review changes; legacy tasks stay explicit (4530.6977ms) ✔ deep Windows data directories keep REA output within the external path allowance and retain owned evidence (2529.5984ms) ✔ user research registration creates an owned Task without model requests and seals the exact artifact identity (1167.9364ms) ✔ complete research binds original Evidence and Bundle, pages exact fields and reuses verified results after reload (2812.0471ms) ✔ artifact changes before analysis refuse dispatch and changes during analysis preserve unverified original responses (2905.2626ms) ✔ owned cancellation closes the connection, concurrent and automatic retries are refused, explicit new analysis preserves history (2535.7166ms) ✔ a dispatched transport timeout stays unknown and reload never launches or kills a saved PID (1554.9672ms) ✔ corrupt Evidence parts and changed binding records are refused instead of silently re-analysed (2277.0362ms) ✔ research documents over the single-artifact limit preserve exact Unicode bytes with owned verifiable parts (1421.5135ms) ✔ research tools are only sent for registered Task artifacts and actual HTTP uses the sealed tool contract without replay (6298.1516ms) ✔ file identity rejects directory and invalid formats without running a target (1074.0308ms) ✔ a stop from the event loop interrupts immediate model replies before later tools (2138.9655ms) ✔ a stop after the first tool preserves a complete interrupted multi-call protocol group (2954.2971ms) ✔ the compatible HTTP loop preserves call IDs, validates errors and waits before executing commands (9991.1531ms) ✔ denial prevents command execution and produces a paired denied result (3525.3804ms) ✔ stopping during approval interrupts pending calls without executing or replaying them (2586.9024ms) ✔ stopping a running fixture command terminates its own process tree and retains uncertain state (3552.8483ms) ✔ command output decodes split UTF-8 and excludes the workbench key from child environment (1614.9624ms) ✔ model-list testing uses a tentative key without changing the saved connection and redacts service errors (5.9629ms) ✔ truncated or repeated model calls fail before execution; tool input validation rejects unknown capabilities (4667.8737ms) ✔ request projection keeps exact tool pairs and complete recent user instructions while preserving disk history (1780.9205ms) ✔ local skill revisions require explicit enable, retain original text, survive restart and stay within the selected project (1607.2723ms) ✔ registration rejects non-SKILL paths, oversized/non-UTF8 files, duplicate sources and wrong field shapes without running project code (1012.9703ms) ✔ actual HTTP models receive only enabled current-project metadata, then a redacted owned skill page with complete tool pairing (4251.643ms) ✔ source edits after a catalog request prevent the read and appear as changed on the next request without replay (4805.3545ms) ✔ disabled and foreign skills cannot be read by guessing their IDs from another project (4636.7028ms) ✔ skill catalog takes remaining escaped JSON budget and cannot evict the user goal or current code (1711.0588ms) ✔ multilingual skill body pages fit actual tool JSON bytes and every page uses the same immutable artifact (6314.4247ms) ✔ registry failures roll back enable and refresh state; corrupted snapshots block requests before network (1674.587ms) ✔ saved skill catalogs retain SHA-bound historical evidence and reject corrupt catalog blobs (3348.0021ms) ✔ HTTP skill administration enforces project scope, revision checks and idle mutation without model calls (4082.1626ms) ✔ JSON state keeps an atomic backup and preserves corrupt primary files (915.7505ms) ✔ tool-pair recovery fills missing results and drops orphan or duplicate results (0.3398ms) ✔ restart pauses unfinished work, retains proposals and never replays a command (1432.1388ms) ✔ public snapshots carry increasing revisions without the private provider transcript or config key (1061.0854ms) ✔ a checkpoint seals all task requirements and original results without a model request and explicit continuation keeps the same Task without replay (5625.4336ms) ✔ checkpoint sources, requirements, plan and original tool changes block continuation and leave original checkpoint bytes intact (4673.1897ms) ✔ unknown tool effects and pending proposals stay blocked in recovery packages rather than being changed into completed work (4514.1458ms) ✔ checkpoint and referenced result corruption are rejected before dispatch and historical blobs are never rebuilt (4267.4671ms) ✔ checkpoint ownership, exact confirmation and concurrent operations reject foreign or stale resumes (4819.6543ms) ✔ Git identity changes and altered plan proofs cannot be accepted as the sealed recovery state (4577.4067ms) ✔ checkpoint records survive SQLite reopen and do not trigger automatic model or tool replay (4411.4392ms) ✔ incomplete scans and in-progress project activity refuse checkpoint creation while private credentials are redacted from saved requirements (3972.8285ms) ✔ oversized recovery evidence is refused before reading or publishing the checkpoint (4530.6346ms) ✔ compiled checkpoint HTTP APIs bind current Task/Run, isolate sessions and only resume on an explicit matching confirmation (10397.97ms) ✔ actual process exit before checkpoint commit retains only durable records without replay (4524.9791ms) ✔ actual process exit after checkpoint commit retains only durable records without replay (4534.2833ms) ✔ a real two-reply failure replay reaches request three at 16K with complete goal, original sources and sealed tool evidence (11547.7787ms) ✔ request deduplication retains long original intent, extra requirements, constraints and acceptance without changing the stored contract (2710.507ms) ✔ only the routine context avoids plan echoes; diagnostic API and requests without an included plan keep original observations (1131.8509ms) ✔ operational questions are diagnostic while explicit repair requests remain execution instructions (1.0368ms) ✔ an explicit prohibition on running engineering checks is not interpreted as a requirement to run them (1.6781ms) ✔ file delivery requires a positive instruction, not saved evidence or a prohibited write (0.2611ms) ✔ an unsubstantiated text claim cannot complete a requested file delivery (2039.5324ms) ✔ HTML without tests ends at reviewed file delivery and never offers automatic rechecking (3540.7136ms) ✔ diagnostic answers retain the unfinished task, limits and check result, and enforce read-only tools (5814.1655ms) ✔ review continuation is a system action and adds no testing demand or fake user instruction (3973.6182ms) ✔ a provider cannot execute tests against an explicit no-test task even if it ignores advertised tools (2949.6547ms) ✔ a budget pause becomes visible only after its input snapshot can be sealed with the original Run (2747.9847ms) ✔ alternating duplicate file reads and listing files still pauses without treating new pages as duplicates (23690.8479ms) ✔ task-plan HTTP mutations bind current Task/Run and retain original evidence through real host restart (6121.3304ms) ✔ concurrent stale HTTP plan revisions cannot replace conditions or attach proofs to a different plan (4152.9052ms) ✔ a real escaped long-file tool message keeps its proof identity within a tight request budget (1426.7151ms) ✔ task plans reject fabricated completion, invalid scopes and cyclic or missing dependencies (0.5254ms) ✔ sealed plans retain goal, contract hash and original steps through SQLite reopen without executing tools (1879.2673ms) ✔ proofs require original owned reads, honor dependencies and reject altered invocation arguments and bytes (2078.5511ms) ✔ new user requirements invalidate current proofs and reseal the same original operations in a new revision (2472.3348ms) ✔ applied change-set evidence binds reviewed current files and invalidates dependent steps after outside edits (2559.3904ms) ✔ actual npm checks are verified by original counted logs and current source, with failure history retained (8865.6206ms) ✔ real desktop and mobile browser proofs bind required actions and PNGs, rejecting source changes, a later failed plan and changed builds (12889.8019ms) ✔ Git delivery blocks incomplete plans and changed preview bindings, preserving HEAD and owner index before a fresh verified commit (14751.1616ms) ✔ the runtime supplies owned read IDs and completes a plan only after original step evidence verification (4254.0917ms) ✔ premature model completion pauses the original Task and resume preserves the existing plan without replay (3856.7464ms) ✔ plan operations participate in project activity and orderly shutdown (738.3887ms) ✔ actual process exit before plan save restores only committed plan evidence without replay (1710.3265ms) ✔ actual process exit after plan save restores only committed plan evidence without replay (1371.2075ms) ✔ actual process exit before proof save restores only committed plan evidence without replay (1382.3406ms) ✔ actual process exit after proof save restores only committed plan evidence without replay (1542.4539ms) ✔ the latest oversized session result reaches the next request together with already read source modules (3389.002ms) ✔ every call in a large latest group keeps its actual outcome, including denied and unknown results (2051.1476ms) ✔ a small complete latest tool group is sent once and needs no duplicate feedback view (1887.4283ms) ✔ real HTTP repeated state inspection pauses after three identical engineering outcomes and preserves tool pairing (4558.1706ms) ✔ fixed normal exploration, pagination, SHA changes and repeated ranges are external observations, not a score (1.3812ms) ✔ read-only research is valid work; model summaries and unchanged timestamped listings do not become engineering evidence (0.3829ms) ✔ new permissions, network and broken tools are failure evidence; repeated invalid inputs remain candidates (0.4749ms) ✔ waiting for original long checks, review and approvals stays waiting across repeated observations (0.7345ms) ✔ baseline facts, real repairs, current-source check results and requirement evidence have distinct growth (0.2238ms) ✔ bounded memory keeps unknown instead of claiming an unretained range is repeated; Task/Run instances cannot mix (11.7147ms) ✔ actual runtime records growing pages then a repeated page without changing model rounds, stop rules or original tool pairs (5375.5069ms) ✔ actual invalid tool calls repeat under the original step budget; observation never adds an early stop (4451.6604ms) ✔ actual command approval is observed as waiting and denial stays a permission failure (2655.8862ms) ✔ network errors remain model failures with shadow failure evidence, never a fabricated no-progress cause (4364.4586ms) ✔ observer receipt failure cannot change tool results, model budget or completion (2083.4906ms) ✔ a later receipt recovers after an observation write failure without misaligning retained source receipts (3091.3696ms) ✔ real applied files, counted checks and current requirement proofs grow after an existing-facts baseline (5019.1907ms) ✔ receipt retention stays aligned at the cap and a foreign step cannot enter the Run (2639.544ms) ✔ an actual approved long check records its owned running job as waiting before the original counted result (7652.6896ms) ✔ a live SQLite host with the actual 1.9 artifact object upgrades, keeps its original blobs and verifies real browser PNG bytes (4668.7742ms) ✔ failed upgrade restores the original artifact object, provider and request handler (2207.8733ms) ✔ claiming the same runtime owner twice preserves its original metadata and every database row (852.885ms) ✔ sessions without controlled-engine history retain their prior export shape and never expose another session record (1105.1664ms) ✔ live workers, controlled drivers, or MCP operations refuse an upgrade before changing handlers, managers or database rows (814.8977ms) ✔ an idle upgrade loads the newly introduced engine record map, resets obsolete managers and preserves original rows, sessions and provider objects (1081.0285ms) ✔ an installation failure restores old managers, original record maps, handlers and all existing database rows (1910.5675ms) ✔ the owned controlled driver remains an active session and project even after outer task activity has ended (849.5869ms) ✔ release and upgrade help or invalid arguments have no file or network setup side effects in an independent directory (857.9308ms) ✔ hot upgrade adopts a retained virtual data root without reopening SQLite or changing rows, Provider or routing (1815.5231ms) ✔ failed adoption restores the original virtual data-root name and original open database (877.0526ms) ✔ hot upgrade refreshes retained model and importer methods while preserving credential and routing objects (1095.1274ms) ✔ hot upgrade refuses active probes and restores retained model methods if installation fails (757.4883ms) ✔ the upgrade compatibility gate includes the served current version and the previous formal patch (0.4276ms) ✔ a directly started compiled host upgrades twice with the same PID, sessions, applied bytes and volatile provider (4200.486ms) ✔ the running 0.1.0 process upgrades in place and retains its project, session, files and volatile model connection (8805.4425ms) ✔ HTTP worker identities require a current sealed plan, reject overrides and close unused plans without touching source (23135.3183ms) ✔ HTTP unknown recovery only accepts the original evidence SHA and preserves changed source for explicit retention (22933.4304ms) ✔ verified isolated outputs form a new parent batch, rerun real checks and commit only reviewed source (68350.5512ms) ✔ overlapping worker paths are explicit conflicts even if each independent task checks successfully (38684.5495ms) ✔ source edits after isolated work preserve the owner file and invalidate collection without proposing a merge (44902.9384ms) ✔ changed child bytes after parent preview block the entire batch before any parent write (63421.6721ms) ✔ corrupt original check logs cannot become collected outputs or a parent review (48859.4073ms) ✔ stopping a parent accounts late responses and executes none of their proposed tools (18346.5372ms) ✔ unreported interrupted requests reserve estimated usage and the total deadline cannot reset on resume (17901.6273ms) ✔ a deadline before worker request dispatch records no invented usage and cannot reset on resume (16177.8811ms) ✔ a response exceeding the cumulative token allocation stops before registering or executing its tools (17902.6513ms) ✔ cumulative tool allocation rejects the entire returned batch before any tool side effect (19564.8542ms) ✔ single response 16 tool ceiling rejects the entire returned batch before any tool side effect (21311.3904ms) ✔ consumed request steps remain consumed across explicit worker resumes (24274.9088ms) ✔ a saved completed response with missing usage is conservatively accounted exactly once during recovery (21534.7613ms) ✔ an unexecuted second workspace plan resumes its original identity without recreating the first checkout (22457.741ms) ✔ a verified interrupted parent proposal resumes sealing without duplicating or replacing its original change (68639.4292ms) ✔ recovery attaches a sealed parent batch saved before its group reference without creating another batch (55813.2637ms) ✔ explicitly retaining an unexecuted parent merge cancels its batch and frees the next parent goal (54665.9561ms) ✔ unknown collaboration blocks ordinary proposals but permits only original batch recovery and rollback (58411.8219ms) ✔ after a restart the batch application lazily reinstates the worker result guard before any parent write (59875.2762ms) ✔ unknown original tool inputs must match before acknowledgement, and reviewed checkpoint IDs preserve the unknown fact (18100.5361ms) ✔ a parent stop observes and waits for a manually approved child check and its real descendant process (17975.473ms) ✔ a real host crash preserves unknown checks, refuses a still living owned PID and never replays a model or tool (32192.1522ms) ✔ a persisted child initialization resumes the original control run after a failure before its worker record (20233.7553ms) ✔ failed request intent persistence never dispatches a worker request or changes source (21124.62ms) ✔ failed usage persistence rolls back the uncommitted usage and conservatively reserves one unknown request without tools (16620.065ms) ✔ a child file change after observation prevents acknowledgement even when the tool inputs did not change (15854.2755ms) ✔ worker scopes and allocated budgets reject aliases, extra permissions and aggregate overflow (713.3608ms) ✔ two sequential Git workspaces freeze a new parent goal, separate owned tasks and preserve all source bytes (15791.3862ms) ✔ real isolated changes, approvals and npm checks produce original owned outputs, not model completion claims (41502.2487ms) ✔ forged worker tools cannot execute terminal or Git even when absent from the advertised tool list (16641.8873ms) ✔ a completion-only model pauses each worker without fabricated changes or checks (24935.0353ms) ✔ restart retains frozen scope and marks interrupted parent and workers unknown without sending requests (14102.9297ms) ✔ all actually returned same-version fragments reach the bounded request, including overlap and Unicode, with independent original provenance (4581.9155ms) ✔ unread gaps and different range origins remain separate; partial sending names the precise missing already-observed range (1056.1026ms) ✔ same-SHA conflicting overlaps and invalid coordinates fail closed, while a new SHA never inherits old fragments (1116.1633ms) ✔ versioned old views keep the original single-page semantics and verification (631.1409ms) ✔ small code and actual failed command stay visible when their original multi-tool group cannot fit (1179.0739ms) ✔ mixed group retains six distinct small file observations with exact multilingual text and bounded request bytes (1873.1226ms) ✔ working observations are bound to the current Task and stop treating a pre-application file revision as current evidence (1797.4167ms) ✔ a sealed observation rejects changed or missing original result artifacts instead of recreating evidence (1414.125ms) ✔ a real passed-check observation is excluded after a source edit invalidates the check, while its historical sent evidence remains valid (5348.426ms) ✔ budgeted file pages reconstruct UTF-8 text including long single lines and surrogate boundaries (4.5427ms) ✔ actual HTTP multi-tool reads survive an oversized runtime snapshot and the sealed evidence survives reopening (11654.7316ms) ✔ actual HTTP file pagination advances over a long file with each returned page visible in the next request (12120.8951ms) ✔ later mixed history pages cannot displace the actual module sources and complete requirements (8007.2891ms) ✔ a later narrow page of the same file revision cannot evict the wider actual observation (3791.6762ms) ✔ six small source files stay complete across metadata turns even when mandatory evidence exceeds the soft target (3204.0569ms) ✔ historical read-artifact navigation points to the original artifact and next page, never its result wrapper (1191.9649ms) ✔ a 1.6.1 full-provenance sent view remains verifiable after introducing the compact view (1313.1548ms) ✔ actual HTTP requests keep four observed modules after a later large history read and preserve all source bytes (12711.5343ms) ✔ workspace HTTP ownership, reviewed creation and reopen survive restart without changing source files (16307.4106ms) ✔ isolated model Task applies a batch, executes real checks and commits at the Git review boundary with a source-bound report (45980.1766ms) warning: in the working copy of 'src/空 格.mjs', LF will be replaced by CRLF the next time Git touches it ✔ worktree checkout disables hooks, fsmonitor and all filter command variants while honoring built-in EOL conversion (7943.2784ms) ✔ an ignored external file at checkout is observed and never accepted as pristine creation (5839.3775ms) ✔ outside content hidden by assume-unchanged cannot pass workspace verification and blocks further source execution (4973.2586ms) ✔ workspace preparation refuses known shared-repository activity and preserves archived child sessions (6597.7928ms) ✔ compact checkout identities create in deep Unicode source directories while retaining the human branch name (6946.666ms) ✔ Windows Git metadata path limits reject creation before publishing a branch or changing source bytes (2603.142ms) ✔ a pending source snapshot participates in shutdown and completes before storage can close (2707.38ms) ✔ source configuration values stay private and a configuration change invalidates a saved review (2987.019ms) ✔ workspace review rejects excluded dependency before any checkout (3033.4995ms) ✔ workspace review rejects large committed blob before any checkout (2709.1063ms) ✔ workspace review rejects case collision before any checkout (3080.2027ms) ✔ workspace review rejects committed symbolic link before any checkout (3002.0711ms) ✔ workspace review rejects external config include before any checkout (2098.162ms) ✔ actual process exit at workspace intent preserves source and reconciles without creation replay (5408.897ms) ✔ actual process exit at workspace git_started preserves source and reconciles without creation replay (7115.7305ms) ✔ actual process exit at workspace git_finished preserves source and reconciles without creation replay (8482.6086ms) ✔ actual process exit at workspace verifying preserves source and reconciles without creation replay (9421.356ms) ✔ actual process exit at workspace registering preserves source and reconciles without creation replay (11213.881ms) ✔ actual process exit at workspace project_registered preserves source and reconciles without creation replay (11371.8855ms) ✔ actual process exit at workspace session_registered preserves source and reconciles without creation replay (9459.6471ms) ✔ actual process exit at workspace ready preserves source and reconciles without creation replay (6708.4742ms) ✔ outside changes at project registration remain unknown and preserve the already registered child identity (8343.5077ms) ✔ managed worktree preserves source index and dirty files, opens a distinct project and binds Run identity (6317.1055ms) ✔ stale review, branch/path collisions and forged ownership do not run worktree creation (3430.7569ms) ✔ workspace project verification is read-only and rejects physical Git identity changes (4718.525ms) ℹ tests 899 ℹ suites 0 ℹ pass 898 ℹ fail 1 ℹ cancelled 0 ℹ skipped 0 ℹ todo 0 ℹ duration_ms 2361965.779 ✖ failing tests: test at tests\api.test.mjs:96:1 ✖ a real bundle restart preserves proposals, interrupts old approvals and requests fresh consent on resume (21946.3358ms) Error: Timed out waiting for test state. at until (file:///D:/Ai-project/REA%E9%80%86%E5%90%91/Agent-Workbench/tests/helpers.mjs:39:38) at process.processTicksAndRejections (node:internal/process/task_queues:104:5) at async TestContext. (file:///D:/Ai-project/REA%E9%80%86%E5%90%91/Agent-Workbench/tests/api.test.mjs:101:3) at async Test.run (node:internal/test_runner/test:1409:7) at async Test.processPendingSubtests (node:internal/test_runner/test:974:7)