v0.5.20

• Going back from a document that has not been read reads it first when requirements remain, so a pending
  action can use values from that page after leaving it.

• Answer checks retain a comparison's collection scope and completeness alongside its quoted records, so a
  whole-list winner keeps the coverage established during reading.

• Answer checks retain whether a quote covers a complete downloaded file, so totals can be verified from all
  its records. Partial excerpts and truncated downloads do not claim complete file coverage.

• Updated page evidence retires vanished context quotes as well as previous answers, so an earlier subtotal
  does not conflict with the subtotal after the basket changes. Quotes still shown remain available to checks.

• A confident request for recovery follows the pending recovery action when it is still valid, so a run can
  leave a page with no usable controls without exhausting recovery on the same state.

• Answer checks through the Vercel AI Gateway receive each subject's identifying quote, so supported answers
  finish as complete there as they do through OpenRouter. Unsupported answers stay rejected on both.

• Reply controls retain their own comment record as context when only one reply form is open.

• Provider attempt deadlines bound total response time, including responses that keep sending occasional bytes.

• Browser commands that remain unanswered for 120 seconds return unavailable even if health checks succeed.

• Closing the last owned browser tab returns unavailable with retained run evidence instead of a Python error.

• Styled upload labels expose their hidden file inputs as upload targets and receive supplied attachments.

• Page captures retain visible image DOM metadata, including source and alt text.

• Missing completion charges are recovered from matching provider generation receipts when available.
  Dollar caps still stop the run when the charge cannot be established.

• Optional --cursor feedback draws an agent overlay for visible local Chrome on Linux X11 with Cua Driver.
  It leaves the physical pointer and focus alone, hides when its page or window changes, and skips unverified
  window mappings. Cloud, headless, Wayland and remote Chrome connections run without it.

• Read captures retain empty form fields, disabled buttons, horizontal overflow and the absence of visible h1
  headings across accessible document and shadow roots. These DOM observations cite the source page without
  text fragments for synthetic labels. Field values supplied as URLs are checked before navigation.

• Recordings of attached windows leave the caller's page in place. Explicit submit buttons expose the form's
  destination to the authorization check, while focusing a field does not claim to submit it.

• Captured hover captions retain their DOM container and sibling position as source metadata.
  Answer checks can bind positional subjects without treating that context as a quoted value.

• Answer identity checks emit each literal identifying quote once and reference it across fields.
  Repeated names no longer consume output tokens for every requested value; citation and subject checks remain.

• Answer audits consider conflicting current quotes from the same source address and frame, including
  uncited notes. Conflicting values must be acknowledged rather than presented as one settled value;
  those other quotes cannot replace a claim's own citation support.

• Consecutive evidence notes share their exact source URL when it makes the rendered context smaller.
  Long referral addresses no longer repeat for every fact, leaving more room for required quotes and their basis.

• Repeated reads keep one copy of each identical prior-evidence packet in the novelty check. Distinct
  quotes, summaries, frames, headings and changes remain separate, and every stored fact and citation is retained.
  Duplicate context no longer crowds this check out of its input budget.

• The agent skill uses uncapped resource defaults, with limits supplied by the caller's task budget.
  It no longer imposes a 30-step limit or a $0.25 model-spend cap on each delegated task.

• Revealed hover targets keep their labels and sibling positions while their captions are visible.

• Final answer audits retain count scope, collection completeness and the quoted records behind each tally.

• Recovery checks whether a changed page satisfies the task before stopping for lack of further actions.
  The completion verifier still rejects unfinished work.

• Requests to expand or load content until a stopping condition keep that page change as an action
  requirement. A displayed total no longer replaces the requested work in the plan. The feed fixture
  also checks that every remaining batch was loaded, and advances to task version 6.

• A ranking shortcut requires a quoted statement of the active list order. Offered sort links no
  longer close a comparison; rejected shortcuts collect the current records before paging onward.
  Numbered links to the next page use the same paging path when their address preserves the list filters.

• Downloaded CSV files retain table structure, so row citations keep the column names needed to
  interpret their values. Bounded downloads still report omitted file content.

• Count claim checks retain the requested scope and whether the accumulated tally is complete.
  An ungrouped count no longer reaches the checker as an unlabeled number.

• Planning uses the default reading model and separates sign-in identities from requested data filters.
  A prerequisite account name no longer becomes an extra field in an exported report's answer checks.

• Answer verification rejects superseded facts from the same page, including calculated facts that
  remain in browsing history after their requirement evidence is replaced.

• A slow shortcut can finish reporting its model cost while browsing continues from the start page.
  Falling back no longer cancels a dispatched request and interrupts a run with a dollar limit.
  A late cost that exceeds the limit stops the run while retaining its already verified answer and data.

• Answer checks judge calculated results from their quoted operands. Reader lineage no longer blocks
  a calculation when its complete source records support it.

• Scalar reads retain cited headings that identify their values, so answer verification can bind a price
  or count to its named subject. Heading evidence stays within its frame.

• Answer verification requires a response for each requested check. Structured model responses cannot
  omit identity bindings or field judgments through empty maps.

• Completion checks preserve each requested item's fields even when a draft plan lists generic field names.
  Quoted identities keep one item's value from answering another item's field, including compound claims.
  An identity can come from the quoted page body or its captured title.

• Cancelled Jev requests retain estimated input costs instead of disappearing from the spend ledger.
  Cancelled LLM requests keep reported earlier charges and mark missing billing as unknown.

• Older quotes retain their original page title when unchanged text is captured under a new title.

• Quoted excerpts can include known source-block labels copied by the reader. Those labels are removed
  before matching real page text, while unknown labels and ambiguous quotes remain rejected.

• Navigation headings stay within their menu instead of identifying later page facts. Navigation notes
  show an explicit prefix for long source addresses, leaving room for previously collected facts.

• Answer repairs reuse quoted notes when selected citations omit available evidence. Each repair must
  improve source coverage, or reduce assertion failures without worsening source coverage. Identical
  audit requests are reused within a run; changed evidence, entities, claims or scope are checked again.
  Component values cannot stand in for totals, and rewritten conclusions do not restore the read budget.

• Completion checks each requested answer output against its cited sources. Missing or ambiguous outputs
  return to recovery, so a partially answered requirement cannot finish from its other cited fields.

• Identical captured spans on different addresses retain separate citations and counted records.

• Headerless table rows retain the real leading cells identifying their columns and keep separate citations for counting.

• Removing a doubted answer claim checks whether its requested output is missing, even when other fields
  still cite the same requirement. Optional details can still be removed.

• Prose reads use larger source chunks to keep navigation and following details in one read. Quote checks
  and the smaller choice and counting budgets remain in place.

• Navigation links keep their stale-attempt history when card labels or context change, so a redraw cannot
  restore a repeatedly failing destination. Separate in-page actions retain their own retry history.

• run_task and connect_cdp take allowed_origins, a list of exact http(s)://host[:port] origins that limits
  the documents a run may inspect and control. Navigations, redirects, frames (in or out of process) and popups
  outside the list are refused before they are sent, and a window or frame already open outside it is never read,
  captured or screenshotted. It scopes documents, not the network: images, scripts and requests a granted page makes
  to other hosts still load. Service workers are bypassed so every navigation meets the check, a scoped run delivers
  no live frames or screenshots, and record is refused with it. Opaque child documents are excluded from text
  and controls. Without it nothing changes.

• run_task and connect_cdp take check_access, an async callback awaited before browser startup and every
  observation, capture, screenshot, address, navigation and action. If it raises, the run stops with a BrowserError that keeps only the
  exception's type.

• Evaluation evidence can carry token-based API-equivalent cost ranges separately from recorded spend.

• Runs no longer stop at implicit step, Jev-call, LLM-call, dollar or time ceilings. Callers can still set explicit limits when they need them.

• Complete benchmark suites can be published independently. Existing published suites stay
  present, and each suite retains its coverage, evidence and regression checks.

• Navigation notes write each source URL once and reference it from the remaining claims, preserving full
  addresses while leaving more room for product details and progress.

• Identical control relevance questions share one Jev score within each pass. Controls remain separate
  action targets, with fresh selection and authorization.

• Rewording an already collected source quote no longer counts as read progress, so timer redraws and
  paraphrased claims cannot keep restoring a barren page's read budget. Newly evidenced requirements still count.

• Links that repeatedly fail their freshness check leave the choices for that document, so intervening reads
  cannot keep retrying the same stale link. Recovery can choose another route without weakening input guards.

• Prose reads can cite a verified excerpt within a large source block, so later reads carry the supporting
  passage instead of unrelated page text. Code copies the original span and rejects missing or ambiguous matches.

• Independent answer claim checks run in parallel batches. Each omission check keeps its own quoted-evidence
  budget instead of sharing it with unrelated claim questions. Missing or failed checks cannot verify an answer.

• Action decisions keep recent progress claims when a source quote exceeds their notes budget, so collecting
  a large quote does not hide which items were already checked. Reads and completion checks retain full quotes.

• Page redraws retain completed Jev relevance scores for unchanged controls in the same document and task
  context. Unfinished batches still run in parallel, and every action is chosen from the fresh page.

• Collected evidence includes a verbatim source quote once when the fact text is identical, reducing repeated
  input in later reads and checks while preserving source quotes and citation links.

• The internal benchmark catalog includes the opt-in Codex and Cua Driver comparison arm, run by Parallax.

• Final screenshots allow the fresh page check to finish its loading wait before capture. Loading pages
  retain their ending frame, with visible secrets still withheld.

• A directly quoted total can answer a count after an earlier record tally was incomplete. Derived counts
  still require complete records, and page scope and answer claims remain verified.

• Runs that stop on a blocked page, an error or a limit retain a final screenshot when step frames are enabled.
  The capture checks the current page for secrets and preserves the run status.

• Jev chooses repeated short facts as one value, retaining every source quote and context. Duplicate locations
  no longer split its confidence and send an otherwise clear lookup to the prose reader. Headerless tables
  describe values by their row rather than inventing column headers.

• A run whose form submission the server answered with an error no longer ends complete because a later page
  loaded cleanly. A message form that answered 503 with a page still saying "your message has been sent" was
  reloaded, the reload answered 200, and the run reported the message delivered. The last submission's status now
  counts: an error there ends the run the way an error on the final page does, unless a later submission succeeded.

• Icon-only buttons use recognized SVG glyph names when other labels are missing. Button-styled links expose
  one clickable control. Empty plans retain the requested outcome, and required CI checks reject failed jobs.

• Evaluation pages report grades, completion, retries, unknown costs and missing coverage separately.

• Headed sections preserve styled pairs as separate records, so counts can distinguish groups from their
  items. Covered-target recovery excludes controls belonging to the target's own fixed container.

• Task-supplied sign-in addresses open directly. Blank opening pages get one reload before any interaction,
  so a cold app can recover without repeating a submission. Shortcuts leave unspecified filters to observed
  controls, and recovery can return toward the caller's start page when a guessed address makes no progress.

• Form controls retain their input name and autocomplete purpose after validation changes their labels.
  Recovery corrections are bound to the observed field and reuse only supplied, non-secret values.

• Tasks inspecting access restrictions can read a sign-in wall or HTTP access denial as evidence. Protected
  retrieval still requires login, and bot challenges still block the run.

• Record counts use quoted child records instead of a scalar group total. Covered clicks report the visible
  obstruction, and requested transactions have their outcome read before another interaction.
  Malformed tally field ranges get one repair attempt on the saved capture before later counts are marked incomplete.

• External capsule endpoints preserve cookie security attributes while mapping their declared hostnames.
  Strict state supplements use independently sampled sequences and leave missing evidence ungraded.

• Every release requires a successful Evals run on its exact commit. A release with no published comparison
  once skipped the fixture gate and could ship agent code the Evals suites had not passed.

• The LLM can use Vercel AI Gateway when no OpenRouter key is configured. The CLI, MCP server and fixture
  evals share that routing, and scheduled evals check their credentials before starting a task.

• fastbrowse serve, which the JavaScript SDK starts, accepts AI_GATEWAY_API_KEY alone too. Its startup
  check still asked for OPENROUTER_API_KEY and refused a run the LLM could have served through the gateway.

• External eval corpora can run pinned tasks with reproducible selection, attempt records and independent
  grades. Ungraded attempts retain an unknown pass state instead of contributing to a benchmark score.

• Fixture evals leave a page's scripted cookie banner to the agent. The browser layer's own consent refusal
  had removed the mock support portal's banner before the agent could accept it, so that task's requirement
  could not be satisfied.

• Published eval comparisons require a clean build, complete task coverage, at least three measured repeats
  and a ledger retaining every attempt and its spend. CI refuses missing evidence and matched correctness,
  latency or cost regressions. A release needs successful fixture evals and measurements from the same code.

• Interrupted live comparisons retain in-flight attempts and wait for active browsers to stop before closing
  the ledger. Spend that cannot be recovered stays unknown instead of disappearing from the report.

• External corpus comparisons reset each site before every attempt and report completion separately from
  independent correctness. WindTunnel tasks marked excluded by upstream are left out, and authenticated
  fixtures use named credentials scoped to their site.

• External comparison figures require three paired repeats for each included task. Scattered partial repeats
  cannot combine into a score, and stopped draws retain diagnostics without a comparison headline. A completed
  draw also refuses a figure when a task every requested arm could attempt has an ungraded or missing repeat,
  instead of dropping that task after its grades are known.

• Structured extraction reads scalar values from every page the run quoted, not only the one it ended on. A
  sorted listing's later page holds the pricier products, so a price quoted on an earlier page had no candidate
  and the data came back empty while the answer was right. A quote the final page corrected is not offered as a
  current value, and a candidate pool wider than Jev's option ceiling is grouped rather than dropped.

• On a previously read page, Jev checks for new evidence alongside short-fact choices. A confident
  duplicate assessment skips the prose read; changed values and uncertain assessments still reach a reader.
  Runs that hit a money or time limit return collected facts with citations and keep budget_exceeded status.
  Provider and browser errors also preserve collected evidence with their failure status.

• A portable fastbrowse skill lets Codex and Claude Code delegate website tasks through MCP or the CLI.
  Coding agents prefer an existing local Chrome connection, then local Chrome, with cloud as a fallback.
  The skill guide covers installation, configuration, citations, and authorization.