
> tablaze@0.1.0 test
> npm run build && node --test --test-concurrency=1 tests/*.test.mjs


> tablaze@0.1.0 build
> tsc -p tsconfig.json

✔ workspace export keeps active selection consistent when the active tab closes during storage capture (1346.399208ms)
✔ PDF metadata stays bound to its captured tab through asynchronous artifact persistence (664.697958ms)
✔ drag revalidates observed targets after asynchronous geometry reads before mouse input (550.766042ms)
✔ upload chooser revalidates its visible trigger after asynchronous local file checks (572.347166ms)
✔ newly discovered scroll containers retain node identity guards when replaced (488.27775ms)
✔ scrollable containers expose ranges and accept guarded horizontal and vertical input (565.911667ms)
✔ dragging emits the browser drop event and delivers the real data payload (686.859ms)
✔ hidden file chooser transfers explicit local bytes without a hidden DOM reference (625.079709ms)
✔ PDF exports real private artifacts with verifiable digest and source (601.81125ms)
✔ workspace roundtrip restores isolated durable storage, owned tabs, and active selection (1783.503167ms)
✔ workspace validation precedes navigation and failed restore rolls back owned sessions (796.945584ms)
✔ per-decision catalogs hide execution metadata and context changes skip queued writes (13.191875ms)
✔ a context change during planning or write-ahead persistence never executes the stale handler (2.664292ms)
✔ trusted write effects retain ambiguity for thrown, malformed, error, and unknown outcomes (10.591834ms)
✔ finish rechecks its lease and cannot reuse verification from the old context (1.711584ms)
✔ a context change between closed and newly prepared leases invalidates old verification (1.454458ms)
✔ only a confirmed session close preserves current evidence across context changes and later writes still invalidate it (4.257084ms)
✔ bound checkpoints reject missing runtimes and changed registry or same-origin tenant identities (1.44375ms)
✔ version-one and version-two checkpoints explicitly migrate unbound and cannot gain a registry (1.55225ms)
✔ resume rechecks a pending call effect rather than trusting a tampered mutating flag (1.230083ms)
✔ initialization uses the bound dispatcher and each lease closes before terminal persistence (1.459709ms)
✔ a prepared lease that arrives after cancellation is still closed without planning (3.901917ms)
✔ catalog release failure or cancellation cannot turn accepted verification into success (2.268375ms)
✔ run CLI completes a scripted model task through real Chrome and prints a bounded report (886.079916ms)
✔ run CLI returns human input with exit 2 and an explicit failure with exit 1 (423.071959ms)
✔ run CLI does not accept an unverified model completion and enforces the step limit (229.046125ms)
✔ CLI rejects agent-only options outside run and validates required arguments and limits (2249.147416ms)
✔ CLI reports cleanup failure separately from an already verified successful agent result (697.30525ms)
✔ restoration abort handles a rejected asynchronous dispose without an unhandled rejection (324.095333ms)
✔ CLI publishes bounded planner diagnostics without exposing endpoint, key, or upstream error text (227.498333ms)
✔ HTTP planner failures expose only status and fixed classification without enabling default retries (19.43625ms)
✔ unknown exceptions cannot inject diagnostic codes and custom retry policy remains authoritative (3.657667ms)
✔ transport and malformed response diagnostics remain distinct and omit raw response data (3.586167ms)
✔ catalog failure is distinguishable before any planner or mutation runs (2.089125ms)
✔ application callbacks are classified without recursively notifying a broken event hook (5.201708ms)
✔ adapter usage callback failures retain their application phase and original retry policy input (1.091834ms)
✔ persistence failures prevent mutation or successful completion without changing resume schema (1.7405ms)
✔ a noncooperating checkpoint callback receives a bounded persistence-timeout diagnosis (5003.891333ms)
✔ ordinary MCP explicit unknown write persists uncertainty, skips the batch and cannot replay on resume (16.767416ms)
▶ explicit uncertainty overrides a superficially successful write in legacy and bound MCP execution
  ✔ ordinary MCP write annotation (1.590292ms)
  ✔ bound trusted write metadata (2.16725ms)
✔ explicit uncertainty overrides a superficially successful write in legacy and bound MCP execution (3.939584ms)
▶ ordinary pre-write failures and non-protocol markers remain recoverable by fresh observation
  ✔ absent (1.942625ms)
  ✔ false (1.685125ms)
  ✔ string true (1.25ms)
  ✔ nested payload (1.262709ms)
  ✔ text-only JSON (1.290208ms)
✔ ordinary pre-write failures and non-protocol markers remain recoverable by fresh observation (7.711709ms)
▶ read-only metadata and trusted not-started outcomes do not invent mutation ambiguity
  ✔ ordinary read (2.304917ms)
  ✔ bound read despite misleading write annotation (0.911167ms)
  ✔ not_started overrides a contradictory result marker (0.665375ms)
✔ read-only metadata and trusted not-started outcomes do not invent mutation ambiguity (4.062ms)
▶ failure after an explicit unknown result cannot erase the persisted reconciliation requirement
  ✔ after_tool persistence (1.127583ms)
  ✔ tool_result callback (0.942916ms)
✔ failure after an explicit unknown result cannot erase the persisted reconciliation requirement (2.154417ms)
✔ explicit initialization and popup following complete a real workflow without executing remaining origin calls (1243.840125ms)
✔ cancellation during terminal persistence cannot return succeeded with accepted evidence (5.182375ms)
✔ deadline expiry during terminal persistence is reported and returned elapsed time includes the wait (153.203667ms)
✔ returned checkpoint accounts for final persistence time before a later resume (44.490833ms)
✔ resume retains application restrictions unless the caller explicitly replaces them (3.382542ms)
✔ resume cannot silently remove a previously required application completion policy (3.365834ms)
✔ decision checkpoint resume preserves counters and IDs but requires fresh verification (15.809459ms)
✔ an interrupted maybe-applied mutation blocks every resumed planner until trusted reconciliation (2.916917ms)
✔ failed checkpoint persistence prevents dispatch and never retries a tool (2.8225ms)
✔ restored sessions reject old session IDs and refs until a fresh snapshot is obtained (2.322708ms)
✔ untrusted checkpoint shape, protocol pairing, IDs, and budgets are validated before execution (2.222833ms)
✔ stall detection uses repeated actual outcomes, provides one replan, then stops (2.308042ms)
✔ bounded retry and fallback only repeat planning; tools with uncertain effects are not retried (3.344375ms)
✔ opt-in compaction preserves complete tool groups and real verification evidence (7.882ms)
✔ HTTP adapter reports only actual provider usage and exposes transient status classification (10.34925ms)
✔ separate CLI processes restore private storage and fresh sessions without replaying completed writes (1765.574792ms)
✔ interrupted external side effect requires explicit reconciliation before CLI restores or plans (1647.576375ms)
✔ an exhausted saved CLI budget performs no page restoration or model request (1159.524541ms)
✔ explicit initialization traverses actual MCP before planning, with step-zero write-ahead and real evidence (12.651834ms)
✔ start URL is explicit and invalid targets fail before catalog, tools, or planning (0.834666ms)
✔ initialization obeys tool, history, deadline, and cancellation budgets without planning (33.781584ms)
✔ unknown initial effects require reconciliation and the initializer is never replayed after it (0.96925ms)
✔ an initializer that was never registered can start once when its saved prerequisite is supplied (0.384834ms)
✔ settled initialization survives resume and compaction with unique subsequent IDs (1.698916ms)
✔ version-one checkpoints migrate explicitly and initialization manifests reject inconsistent history (0.796833ms)
✔ write-ahead persistence failure never dispatches or replays the registered initializer (0.3805ms)
✔ replan_required skips later tool calls and still requires separate verification (0.312542ms)
✔ CLI start URL reaches real Chrome before its first model request and resumes without another initializer (1894.745666ms)
✔ CLI accepts a browser-wide popup policy but restricts start URL to run and validates both (1170.393541ms)
✔ agent drives real MCP browser tools and requires verified form state before finishing (692.695042ms)
✔ MCP schema rejects malformed arguments before side effects, then planner can recover (2.84825ms)
✔ stale-reference result stops the batch and is returned for replanning without retry (1.694791ms)
✔ failed checks and isError=true cannot satisfy completion even when passed=true (2.3055ms)
✔ completion rejects stale evidence, another session, invented IDs, and contradicted checks (5.233125ms)
✔ application task-specific acceptance can reject a technically passing unrelated check (0.950625ms)
✔ tool and planning budgets are bounded; invalid decisions never dispatch a tool (1.28025ms)
✔ deadline bounds a non-cooperating planner and abort propagates into an in-flight tool (22.211208ms)
✔ MCP cancellation reaches the actual server request signal (0.971167ms)
✔ HTTP planner preserves MCP image evidence and structured errors without real model calls (10.641959ms)
✔ HTTP planner rejects malformed model arguments, mixed finish calls, and oversized output (0.703834ms)
✔ an installed completion policy must explicitly return true (2.383875ms)
✔ history budget stops later mutations within the same planning decision (0.661375ms)
✔ cancellation before the dispatch microtask never starts a custom tool (0.278708ms)
✔ binding is frozen trusted metadata; observations and DOM edits do not consume it or hold the queue (501.982792ms)
✔ same-URL reload invalidates a document lease even if the cached navigation counter is delayed (575.708042ms)
✔ history routes and tab switch ABA invalidate bindings permanently (523.8385ms)
✔ a navigation during the asynchronous assertion cannot produce a valid stale result (565.658667ms)
✔ signal, explicit close, session close and engine disposal revoke leases without exposing browser handles (692.689166ms)
✔ cancelled acquisition releases a real late Document handle without affecting the session (481.237167ms)
✔ pre-cancelled acquisition creates no remote handle and opaque origins are refused (595.572292ms)
✔ engine disposal revokes an in-flight acquisition and releases its late handle (480.196958ms)
✔ independent CDP navigation invalidates a lease and lease cleanup leaves the external browser alive (667.529041ms)
✔ hotel form batch navigates and independent checks verify real URL and DOM (777.930875ms)
✔ snapshots obey budgets and do not include password or hidden values (334.590167ms)
✔ a newer snapshot rejects the older revision before input (164.95125ms)
✔ a replaced DOM node with an identical label is rejected (200.068666ms)
✔ changed target semantics are rejected before a click (190.203042ms)
✔ overlay blocks input and retry requires fresh observation (1278.927875ms)
✔ semantic change during actionability waiting prevents the delayed click (425.670833ms)
✔ a base URL change invalidates the observed resolved link destination (230.354542ms)
✔ value verification never leaks a password substituted after a field-type read (268.942292ms)
✔ batch preserves completed work and reports failed and skipped steps (230.708625ms)
✔ incremental snapshots expose added and removed controls (294.774458ms)
✔ open shadow controls and explicitly selected frame controls execute (365.930667ms)
✔ sessions isolate storage and reject cross-session references (401.501583ms)
✔ concurrent operations serialize within one session (216.005583ms)
✔ capture contains a real JPEG and closing removes a session (203.937916ms)
✔ CDP disposal closes only owned pages and preserves the independent browser and existing state (837.921ms)
✔ stream and item error diagnostics may precede a successful completed turn (72.072333ms)
✔ only allowlisted structured diagnostic fields survive without deriving a retry policy (28.154542ms)
✔ a failed turn stays failed even if termination flushes completion usage and exits zero (27.074084ms)
✔ valid output and exit zero cannot substitute for an actual completed event (26.909166ms)
✔ a message-only error followed by a hanging process ends at the original deadline (259.723417ms)
✔ a nonzero exit and a completed response with invalid schema both remain failures (91.238667ms)
✔ gateway counts one successful dispatch after diagnostics and retains unknown usage on incomplete failure (110.675875ms)
✔ run failure classification uses terminal AgentFailure and never erases a returned failure (0.22425ms)
✔ Codex process decisions complete the Agent write/verify/evidence loop with isolated transient files (109.4405ms)
✔ Codex only inherits allowlisted environment and uses bounded deduplicated inline images (29.321167ms)
✔ Codex generic notifications do not abort successful terminal events or leak provider text (27.962084ms)
✔ Codex incomplete cannot dispatch or imply Agent success (27.928375ms)
✔ Codex nonzero cannot dispatch or imply Agent success (26.163667ms)
✔ Codex failed-terminal cannot dispatch or imply Agent success (24.574875ms)
✔ Codex bad-event cannot dispatch or imply Agent success (24.220459ms)
✔ Codex external cannot dispatch or imply Agent success (24.789416ms)
✔ Codex stderr-limit cannot dispatch or imply Agent success (26.639792ms)
✔ Codex large cannot dispatch or imply Agent success (25.491917ms)
✔ Codex bad-json cannot dispatch or imply Agent success (25.985083ms)
✔ Codex symlink cannot dispatch or imply Agent success (26.719667ms)
✔ Codex validates original tool arguments and the final schema before dispatch (133.693541ms)
✔ Codex timeout drains final reported usage, deletes files and keeps fixed failure diagnostics (514.514292ms)
✔ Codex one-request cancellation is isolated and close drains every outstanding process (59.208625ms)
✔ Codex missing usage counters remain absent and usage hook errors still produce process diagnostics (32.759667ms)
✔ Codex file cleanup failure remains visible to Agent and close without dropping measured usage (33.052833ms)
✔ Codex invalid configuration fails before child launch (0.141666ms)
✔ successful Codex leader cannot leave an owned stdio-independent descendant running (68.882334ms)
✔ cancelled Codex process retains usage emitted while handling SIGTERM (542.284792ms)
✔ gateway close waits for cancellation usage before exposing final metrics (106.625166ms)
✔ named tool_choice rejects a different schema-valid tool (45.455792ms)
✔ tool_choice none permits a text response with tools present (30.540416ms)
✔ tool_choice auto with no selected tools is a stopped text completion (30.369666ms)
✔ timeout terminates a SIGTERM-resistant descendant after its parent exits (1716.00225ms)
✔ Codex transport adapts function decisions, disables external tools, and preserves actual CLI usage (44.521792ms)
✔ Browser Use structured schema and inline images share the same Codex inference transport (31.294208ms)
✔ Codex transport rejects schema-invalid function arguments before the browser runtime sees them (31.785708ms)
✔ automatic tool choice can return ordinary text without inventing an executed call (29.666459ms)
✔ unexpected Codex tool activity invalidates the attempt and stops its own process (27.615708ms)
✔ Codex inference deadline and cancellation stop owned child processes without treating them as model answers (212.442708ms)
✔ shared model gateway bridges both framework formats through the same fake Codex process settings (97.423542ms)
✔ Browser Use retains done, true counts, portable history and judge phase when post-task judge times out (221.279ms)
✔ explicit judge false is preserved and ordinary full return stays distinct from cleanup (66.485292ms)
✔ judge usage contains only reported public response values, and judge exceptions retain their original phase (164.672334ms)
✔ pre-done timeout, post-done exception and non-judge finalization timeout preserve distinct observed phases (489.674583ms)
✔ owned cleanup has a bounded outcome and cannot silently count as a completed run (102.205667ms)
✔ Browser Use retains observed download paths when public kill resets browser metadata (59.776459ms)
✔ runner reports independent business success, observed agent done, and full return as separate axes after adapter failure (526.451458ms)
✔ summary retains legacy completedAndPassed while adding explicit axes (0.081041ms)
✔ Tablaze flags reach the official SDK, and proposed finish stays distinct from accepted completion (161.639166ms)
✔ comparison runner stops the owned adapter process group after output overflow (645.756ms)
✔ Browser Use source pin accepts only exact git commits or hashed official archives (59.462875ms)
✔ fixture judges actual records, rejects claimed success, and resets isolated attempts (19.286083ms)
✔ download judge requires the actual matching file, not a download request or a claimed path (4.728583ms)
✔ authorization-return judge requires one provider approval and one app submission (3.661667ms)
✔ model gateway enforces equal wire settings and measures returned usage without a paid provider (5.208083ms)
✔ missing model/dependencies stay not_run and never become a competitor failure or a win (29.439958ms)
✔ scripted Tablaze smoke completes diverse real-browser tasks with independent judges (15337.568125ms)
✔ typed backend write, actual browser receipt, and Agent verification complete one reservation (664.907875ms)
✔ real navigation while planning removes the restricted tool before its stale decision dispatches (583.517167ms)
✔ handler rechecks a real browser binding after waiting and rejects a stale external write (506.800084ms)
✔ same-origin tenant change cannot resume an uncertain write; matching identity reconciles without replay (1584.81125ms)
✔ CLI rejects a bound checkpoint before launching a browser or restoring its URLs (201.858458ms)
✔ input validation rejects extra fields and attempted context injection before entering a handler (3.72025ms)
✔ global tools work without a browser and receive trusted context, stable callId and a signal (0.97775ms)
✔ caller-selected output may deliberately include a context value (0.609125ms)
✔ published output schema validates the actual MCP envelope and relocated reusable refs (15.895792ms)
✔ write handler exceptions and invalid outputs remain unknown without exposing raw errors (2.519958ms)
✔ trusted read effects remain reads when handlers report failures (0.303666ms)
✔ duplicate registrations, forged handles and reserved names are rejected (0.434292ms)
✔ origins reject wildcard and non-origin restrictions and require a browser binding (0.257875ms)
✔ site filtering includes the exact port and metadata still covers invisible tools (6.142833ms)
✔ closed catalogs release a captured guard once and cannot dispatch handlers (0.253292ms)
✔ catalog context keys stay stable across leases and change when the captured document changes (0.440875ms)
✔ binding changes after catalog preparation refuse dispatch before entering a handler (3.077667ms)
✔ context identity is rechecked after asynchronous validation before the handler starts (0.651875ms)
✔ handler assertCurrent rejects a binding change across await and a started write remains unknown (0.4365ms)
✔ handler assertCurrent also rejects its cancelled dispatch after a noncooperating await (0.2885ms)
✔ a retained execution context is revoked when a successful dispatch has returned (0.295125ms)
✔ independent dispatch cancellation bounds a hanging getContext check before_handler (0.433ms)
✔ independent dispatch cancellation bounds a hanging guard check before_handler (0.304542ms)
✔ independent dispatch cancellation bounds a hanging getContext check after_handler (0.413791ms)
✔ independent dispatch cancellation bounds a hanging guard check after_handler (0.391833ms)
✔ stable contract hashes exclude opaque context values while including schemas and all registry tools (0.329542ms)
✔ an already established registry rejects base-contract and tenant identity changes (0.173958ms)
✔ context resolver failures expose only a fixed registry error (0.219458ms)
✔ runAgent uses catalog dispatch, preserves callId, and records unknown custom writes for reconciliation (2.447416ms)
✔ resume rejects schema, version, effect, origin, base-contract and same-origin tenant changes before planning (3.992291ms)
✔ reconciliation of an unknown write cannot rebind its checkpoint to another tenant on the same origin (0.899167ms)
✔ runAgent discards a decision when the captured browser binding changes during planning (0.536875ms)
✔ viewport observation reaches the 501st button after scrolling a truncated page (670.744041ms)
✔ scoped observation exposes the selected subtree and resets incompatible diff baselines (178.255333ms)
✔ viewport text clips individual offscreen lines within a single text node (144.178625ms)
✔ viewport observation excludes controls and text clipped by scroll-container ancestry (137.735458ms)
✔ snapshots, extraction, verification and guarded actions share open shadow and slot content (230.338167ms)
✔ hidden hosts, hidden slots and raw form values do not leak through composed readers (273.890041ms)
✔ text matching searches beyond bounded evidence and across composed text-node boundaries (179.601958ms)
✔ explicit hidden accessibility labels still name guarded controls without leaking form values (224.27925ms)
✔ link and table extraction report item and character truncation accurately (176.95775ms)
✔ scan-budget exhaustion is explicit in both snapshots and extraction (273.977791ms)
✔ ordinary draggable and aria-labeled divs deliver actual native drop payload (695.122709ms)
✔ mouse-driven custom drag supports the observed source covering its drop target (955.799875ms)
✔ nested scaled iframe drag maps main-page coordinates and preserves native payload (1395.252ms)
✔ open Shadow DOM drag target hit testing reaches the actual source and destination (699.844291ms)
✔ an overlay appearing after the initial mouse move cannot intercept mousedown (1281.556916ms)
✔ an overlay appearing during movement cancels the drag before drop and releases the mouse (884.367125ms)
✔ mouseup RPC rejection is a failed action even if the release was applied; cleanup releases held input (1392.415833ms)
✔ an unreleasable mouse closes only the owned session and reports input cleanup failure (820.55275ms)
✔ schema extraction binds every populated leaf to exact source text and authoritative URLs (15.722417ms)
✔ multiple sources can produce nested data and a checked aggregate without a provider dependency (6.873ms)
✔ invented sources, URLs, approximate quotes and missing or non-leaf citations are rejected (22.538458ms)
✔ JSON pointers escape literal keys and require evidence for null and empty containers (2.609917ms)
✔ schema types, required fields, formats and local references are enforced without coercion (10.991625ms)
✔ literal schema const data is not interpreted as schema references (3.348208ms)
✔ source and candidate limits fail before accepting output (14.792541ms)
✔ quote presence is distinct from truth and the application can reject unsupported claims (17.124917ms)
✔ cancellation and deadline stop waiting on non-cooperating extractors without retries (17.24925ms)
✔ deterministic fields create typed values and preserve their exact DOM provenance (2.289916ms)
✔ field-map cardinality, optional omissions, strict conversions and observation binding are enforced (21.061625ms)
✔ browser inspector uses composed text, current form state and real page provenance (599.507833ms)
✔ find traverses actual virtualized rows and returns a fresh actionable viewport ref (4876.87325ms)
✔ MCP find supports a selected iframe and preserves current snapshot semantics (831.308542ms)
✔ closing a CDP session cancels its unfinished download without closing the external browser (786.240084ms)
✔ a tab created after disposal begins is closed instead of escaping session ownership (715.179458ms)
✔ coordinate clicks never move to a sibling tab when the observed tab closes (740.7665ms)
✔ allowed native form submission preserves one POST with the actual encoded business values (743.364958ms)
✔ allowed 307 redirect preserves the original POST method and body without duplicate dispatch (715.843625ms)
✔ same-origin 302 processes Set-Cookie and a matching workspace restores that cookie and policy (1038.172625ms)
✔ explicit open/goto/new-tab deny the other port before a request or a new tab exists (737.123167ms)
✔ multi-hop HTTP 302 is intercepted before the denied hop during open, goto and new-tab (951.996208ms)
✔ iframe redirect is blocked before its denied server receives a request and counts remain per context (1014.934334ms)
✔ one context blocking a document does not interrupt a different context pending business action (1215.283292ms)
✔ a new popup first-document redirect is blocked even before an ordinary popup snapshot exists (836.145459ms)
✔ page script and meta refresh cannot issue an HTTP document request to a denied origin (1104.505334ms)
✔ an allowed localhost child of the 127.0.0.1 page retains interception through its cross-site redirect (980.094792ms)
✔ non-document fetch and image requests reach the other origin: navigation policy is not a network firewall (560.159792ms)
✔ workspace policy changes, missing hashes, policy removal and newly added policy reject before browser launch (0.712041ms)
✔ CDP with configured navigation policy is rejected by the constructor without a connection attempt (52.554208ms)
✔ absent navigation policy retains ordinary browsing and does not attach a policy hash or diagnostics (449.957667ms)
✔ doctor validates and summarizes canonical policy without launching or printing origins (782.401458ms)
✔ policy file parsing rejects unsafe shapes, origin syntax, unknown fields and raw error disclosure (2874.772167ms)
✔ policy reader enforces a 64 KiB file bound and refuses directories and missing files (760.801625ms)
✔ policy reader refuses a FIFO without waiting for a writer (194.702667ms)
✔ policy and external CDP or setup are rejected before reading the policy path (759.216208ms)
✔ malformed private checkpoints do not disclose JSON excerpts before policy restoration (579.754542ms)
✔ MCP startup accepts a valid policy and exposes tools without launching a browser (203.739583ms)
✔ empty CLI checkpoints retain policy identity and reject changed, removed or legacy policy on resume (1278.845875ms)
✔ dispose closes its already-owned browser while policy bootstrap is still awaiting a CDP response (2.652417ms)
✔ an in-flight real navigation reports a safe policy fault instead of the transport URL/error (604.647167ms)
✔ policy failure after a real write retains partial action evidence and skips the next mutation (679.938708ms)
✔ undefined disables policy while configured empty policy permits only HTTP(S) (0.75525ms)
✔ explicit empty allowlist denies all targets and differs from an absent allowlist (0.357625ms)
✔ exact origin matching preserves scheme, port and host boundaries with blocked precedence (0.258166ms)
✔ WHATWG normalization covers case, default ports, Unicode hostnames, IPv4 and IPv6 (0.114875ms)
✔ canonical identity is versioned, stable across equivalent ordering and does not collapse deny-all (0.220333ms)
✔ compiled policy and copied arrays cannot be mutated and detached methods retain their policy (0.227708ms)
✔ only origins are accepted as rules; URL parser repairs and wildcard guesses are rejected (0.883625ms)
✔ strict bounded configuration rejects unknown fields, wrong types, sparse arrays and accessors (0.693125ms)
✔ combined allow/block input is limited to 200 entries before normalization or deduplication (0.691542ms)
✔ invalid configuration, malformed navigation and denied URLs never surface private input (0.387333ms)
✔ pre-cancelled open does not start a browser or create a session (3.558125ms)
✔ cancellation during shared browser startup returns promptly and leaves it available (574.46075ms)
✔ a context that arrives after cancellation is closed without touching a sibling session (491.512583ms)
✔ page creation cancellation closes the owned context and any late page (667.374541ms)
✔ navigation cancellation removes only the opening attempt and drains its resources (681.695292ms)
✔ snapshot cancellation cannot register a completed late observation as a live session (609.314583ms)
✔ MCP tab_open propagates cancellation to the engine while sibling sessions survive (695.092292ms)
✔ CDP cancellation cleans a late owned page and leaves independent pages and browser alive (1016.815875ms)
✔ dispose is bounded while an isolated context acquisition has not returned (2438.329292ms)
✔ dispose is bounded while an isolated page acquisition has not returned (2498.555458ms)
✔ dispose reports context close failures rather than claiming complete cleanup (623.060958ms)
✔ dispose retains an existing session cleanup after close removes the session from its map (2350.84375ms)
✔ CDP disposal reports an unidentified pending page and closes it after late arrival (2790.386583ms)
✔ planner removes only exact plain duplicates while preserving errors and distinct or annotated text (10.354875ms)
✔ planner retains text-only fallback and near-duplicate serialization (0.256959ms)
✔ projection preserves image order and complete assistant/tool groups with and without image support (0.507208ms)
✔ projection leaves successful evidence, full history, events, and saved checkpoints intact (3.605ms)
✔ default stay keeps the opener active and completes remaining actions (740.27525ms)
✔ follow-single returns popup observation and truthfully skips remaining old-ref actions (797.218959ms)
✔ multiple candidates are retained without choosing the first or stealing focus (857.887292ms)
✔ a late popup is owned but never retroactively changes the active tab (1567.134958ms)
✔ a different owned background opener cannot enter the foreground action window (945.62125ms)
✔ iframe-triggered unique popup is associated with the observed owning page (970.122333ms)
✔ about:blank popups return bounded observations without waiting for eventual navigation (812.838959ms)
✔ popup collection consumes the existing batch deadline and never retries the input (697.572083ms)
✔ workspace preserves policy and old version-one records restore the stay default (2067.159125ms)
✔ failed workspace restoration rolls back policy as well as newly owned sessions (546.34975ms)
✔ stdio handshake exposes the bounded browser tools and structured session errors (211.52125ms)
✔ stdio tools drive isolated Chromium and return browser evidence (909.377834ms)
✔ SDK cancellation stops the browser batch before its later side effect (815.962083ms)
✔ CDP connection errors never return endpoint credentials (219.2905ms)
✔ a total batch budget ends a longer individual wait and reports skipped work (999.651708ms)
✔ CLI help, version and doctor are usable without launching a browser (774.609792ms)
✔ provider selection preserves required explicit model and rejects mismatched options before execution (2850.7035ms)
✔ new provider-only options cannot start the default MCP server (770.951042ms)
✔ Anthropic CLI forwards its explicit model, native tool schema, key and output cap; retains real usage (211.051292ms)
✔ Ollama CLI uses native requests, optional num_predict and no invented token counters (415.530459ms)
✔ explicit compatible provider keeps the previous endpoint and key contract (207.664417ms)
✔ Codex CLI uses the chosen executable and explicit model/effort without a model endpoint (650.963042ms)
✔ Codex executable defaults to codex on PATH without inventing a reasoning setting (645.049834ms)
✔ Codex failures expose fixed diagnostics without stderr, credentials or provider error text (644.660708ms)
✔ Codex provider integrates start-url, actual browser verification and normal cleanup (1250.676625ms)
✔ Codex checkpoint resume preserves prior budgets and still requires provider/model on each invocation (880.838208ms)
✔ bound checkpoints are rejected before any Codex process or browser restoration (198.345917ms)
✔ SIGTERM cancels an in-flight Codex planner and terminates its child (642.889709ms)
✔ anthropic: native wire request completes a verified real MCP tool loop (33.52075ms)
✔ anthropic: images and errors preserve paired repeated-name tool results without mutating history (4.449916ms)
✔ anthropic: text-only selection omits image bytes and native defaults remain explicit (2.097042ms)
✔ anthropic: malformed native decisions, mixed controls and unknown control fields fail safely (5.765208ms)
✔ anthropic: HTTP diagnostics discard private error bodies and do not retry (1.780125ms)
✔ anthropic: native response size is bounded before dropping unknown fields (1.401417ms)
✔ anthropic: cancellation interrupts a pending native response body (1.641833ms)
✔ anthropic: actual missing or invalid usage counters stay absent (1.117375ms)
✔ anthropic: malformed JSON and interrupted bodies retain distinct safe response failures (54.033875ms)
✔ anthropic: unpaired history is rejected before any native request (0.4925ms)
✔ ollama: native wire request completes a verified real MCP tool loop (4.805875ms)
✔ ollama: images and errors preserve paired repeated-name tool results without mutating history (1.481791ms)
✔ ollama: text-only selection omits image bytes and native defaults remain explicit (0.911834ms)
✔ ollama: malformed native decisions, mixed controls and unknown control fields fail safely (3.053916ms)
✔ ollama: HTTP diagnostics discard private error bodies and do not retry (0.832667ms)
✔ ollama: native response size is bounded before dropping unknown fields (0.812708ms)
✔ ollama: cancellation interrupts a pending native response body (0.873666ms)
✔ ollama: actual missing or invalid usage counters stay absent (0.659625ms)
✔ ollama: malformed JSON and interrupted bodies retain distinct safe response failures (52.729416ms)
✔ ollama: unpaired history is rejected before any native request (0.26775ms)
✔ native adapters validate configuration and use only their configured/default native endpoint (0.532959ms)
✔ native factory header validation never includes private names, values or keys in exceptions (0.501583ms)
✔ spawned native-provider CLI rejects an invalid environment key without echoing it to stderr (397.00775ms)
✔ real MCP form fill/save verifies current refs and visible status without another snapshot (716.800625ms)
✔ read-only ref checks preserve actionability and can read a consumed current revision (252.7ms)
✔ ref checks support textarea, select, empty values and open shadow controls (294.242833ms)
✔ MCP and engine reject missing/ambiguous targets and a ref without snapshot_id (182.187791ms)
✔ new observations and other sessions reject old snapshot identities; unknown refs fail (345.682791ms)
✔ replaced, renamed, hidden and navigated targets are rejected without selector fallback (927.341417ms)
✔ parallel status waits cannot preserve ref evidence from a replaced node or old document (821.880875ms)
✔ a later status cannot keep a previously matching value after the same node changes (404.779ms)
✔ iframe refs retain their frame identity and reject detached frame evidence (373.912541ms)
✔ sensitive and non-form targets fail without leaking values; evidence remains bounded (342.450125ms)
✔ queued verification waits for the action batch and shares the original readonly revision (431.700709ms)
✔ scoped credential completes a real login and never appears in normal evidence or artifacts (976.035166ms)
✔ iframe credentials require both the frame and top-level origin to match (757.787416ms)
✔ MCP uses only the credential alias and reports login evidence without disclosing the value (2698.600666ms)
✔ doctor reports only secret metadata and never launches the configured browser (395.638958ms)
✔ malformed secret configuration fails before doctor or run with one fixed message (1143.727208ms)
✔ CDP and setup reject secrets before reading the configuration path (764.492ms)
✔ stdio accepts configured aliases and exposes the public action without resolving or launching (205.931583ms)
✔ CLI run saves secret policy identity and requires the same configuration on resume (824.677458ms)
✔ configuration does not read env values until an allowed-origin resolution and supports rotation (4.60025ms)
✔ returned configuration and scopes are frozen, with only explicit artifact opt-in (1.187792ms)
✔ resolvers reject missing, inherited, empty and oversized values without disclosing their source (1.114625ms)
✔ aborted resolution does not read env and an abort during a getter cannot return the value (0.633083ms)
✔ strict JSON shape and shared SDK metadata limits reject malformed private configurations (4.039ms)
✔ reader enforces the actual 64 KiB boundary, UTF-8 and regular files with fixed errors (1.124959ms)
✔ reader refuses a FIFO without blocking on a writer (2.797208ms)
✔ pre-page bridge exposes only bounded metadata and fills the original control without focusing a decoy (455.807083ms)
✔ old handles fail after identical DOM replacement without filling the replacement (174.787041ms)
✔ opaque HTTP sandbox remains denied despite replacing window.origin in page script (176.422625ms)
✔ captured native primitives withstand later prototype and global replacements (154.754875ms)
✔ bridge property and operation references cannot be replaced by page code (142.460792ms)
✔ invalid bindings, values and unsupported or disabled controls reject before writing (158.337917ms)
✔ an input listener detaching the filled node reports wrote=true and does not dispatch change or fill its replacement (142.229958ms)
✔ same-URL navigation generates a fresh document identity and rejects the prior binding (166.929167ms)
✔ text projection uses lookahead before clipping and never promotes its trailing content (1.131792ms)
✔ raw boundaries remain exact when a secret resembles the replacement marker or overlaps another secret (0.805708ms)
✔ snapshot projection covers fields, options, diffs and metadata without changing reference identity (0.546375ms)
✔ link extraction preserves href-first raw allocation when redaction shrinks the last field (0.15375ms)
✔ link text and table cells redact before their field and aggregate boundaries (0.13475ms)
✔ short secrets cannot expand links or table cells past the aggregate public budget (1.721625ms)
✔ expansion followed by shrinking never reallocates the original reader budget to later fields (0.561542ms)
✔ text extraction restores the original 20000 character boundary (0.090041ms)
✔ selected metadata is copied and recursively redacted with bounded rejection (0.281667ms)
✔ metadata bounds replacement expansion as well as raw input size (43.779042ms)
✔ published and runtime MCP schemas accept only a bounded alias plus observed ref for fill_secret (28.654042ms)
✔ secret failures retain their fixed public code in structured and text MCP results (1.32925ms)
✔ SDK server rejects secret configuration with external CDP before any resolver is called (0.233167ms)
✔ undefined disables secrets; compiled facade is frozen, detached methods work, and values are private (2.111625ms)
✔ strict options reject unknown keys, accessors, prototypes and wrong optional types without invoking getters (0.224625ms)
✔ strict definitions reject invalid names, versions, duplicate aliases, unknown fields and inherited resolvers (0.255542ms)
✔ arrays reject holes, extra properties and getters; alias and configured-origin limits are bounded (0.926834ms)
✔ origin lists require exact nonempty HTTP(S) origins; wildcard, credentials, path and opaque origins are rejected (0.241917ms)
✔ frame and top origins are both required, including port and scheme; scope failure never invokes resolver (0.384667ms)
✔ hash normalizes ordering, duplicates and explicit default top scope; excludes callbacks and resolved values (0.517ms)
✔ caller mutations cannot change cloned scopes, metadata or captured resolver (0.129417ms)
✔ context is checked before resolving and again before delivering; failed delivery still redacts the resolved value (0.205041ms)
✔ resolver and context exceptions are fixed and never expose raw thrown objects or messages (0.570291ms)
✔ pre-aborted calls invoke neither context nor resolver and never expose the abort reason (0.127959ms)
✔ cancellation bounds a hanging pre-context check and does not enter the resolver later (2.188084ms)
✔ cancellation bounds resolver wait, passes an aborted signal, and registers its late value while store remains alive (0.335958ms)
✔ late resolver rejection after cancellation is observed without unhandled rejection (0.269291ms)
✔ cancellation bounds a hanging post-resolution context check while preserving redaction (0.299042ms)
✔ all rotated and cross-alias values remain redacted; duplicate values do not consume retention slots (1.904583ms)
✔ resolved values must be nonempty bounded well-formed strings and errors contain no returned value (0.300208ms)
✔ 128 distinct-value limit fails before delivery without dropping earlier redactions (2.247708ms)
✔ 64 KiB retention limit counts UTF-8 bytes and rejects overflow atomically (9.568583ms)
✔ raw, whitespace-normalized, URI and JSON-escaped variants redact literally and contains follows the same set (0.242584ms)
✔ overlapping secrets use longest simultaneous replacement and never reprocess the replacement marker (0.145583ms)
✔ crossing and chained matches cover every sensitive character in original coordinates (0.330375ms)
✔ public aliases equal to known values are hidden without changing configured resolution or unrelated names (0.165ms)
✔ a late cancelled value and a newer rotated value are both retained until close (1.298917ms)
✔ padding covers maximum URI and JSON forms; redaction before truncation does not reveal a secret prefix (2.969541ms)
✔ close revokes active resolution, aborts the resolver signal and prevents silent redaction bypass afterward (0.457375ms)
✔ MCP schema extraction returns typed shadow/form data and keeps observed references usable (546.688791ms)
✔ MCP preserves structured validation error codes and pointer issues without raw page values (188.97ms)
✔ MCP rejects sensitive, hidden, raw-form and truncated fields without partial extracted data (222.713792ms)
✔ MCP extraction follows the explicitly observed frame and cites its own URL (243.192583ms)
✔ MCP bounds field counts and rejects unsupported field options before extraction (174.330166ms)
✔ real actions show a non-intercepting pointer on owned pages and rejected refs do not fake a click (3376.513625ms)
✔ operators can disable the visual pointer (592.849083ms)
✔ sessions do not add a pointer by default, even when headed (650.345292ms)
✔ popup tabs stay owned, switching invalidates references and last close ends session (557.464167ms)
✔ navigation and explicit tabs preserve session state and handle browser history (309.884459ms)
✔ upload reads an explicit local fixture and download produces the actual file (333.8855ms)
✔ native confirm and prompt use one-shot policies and disclose actual results (442.150958ms)
✔ storage state restores cookies and localStorage across independent engine lifetimes (976.282416ms)
✔ hover and double-click execute guarded references, coordinates reject offscreen input (308.474125ms)
✔ CSS-pixel screenshots support actual canvas coordinate input (210.85375ms)
ℹ tests 438
ℹ suites 0
ℹ pass 438
ℹ fail 0
ℹ cancelled 0
ℹ skipped 0
ℹ todo 0
ℹ duration_ms 175331.289125
