Confucius v0.5.1-beta.1

Fixed

• Automatically recover legacy history when a Zotero reinstall or profile move
  leaves it without a task index. Keep verified backups, rebuild available
  conversations, notes and report links, and reuse saved titles and drafts without
  asking users to move files. Empty directory remnants no longer block startup.
• Resume interrupted recovery without duplicating tasks or overwriting newer
  local work. Damaged history records remain backed up; deleted history stays
  deleted, and recovered archives do not replay tools or restore execution grants.

Upgrade notes

• This is a Beta release. Enable “Include prereleases” in Confucius Settings →
  Update to receive it from 0.5.0 or an older Beta. Disabling Betas never downgrades
  the installed version, and automatic checking remains a separate preference.
• Recovery runs automatically at startup; no file moves, directory cleanup or
  model reconfiguration are required. Original files and verified recovery
  backups remain available. New work is saved in the local runtime directory;
  downgrading does not reverse that migration.
• Recovered tasks stay interrupted until the user explicitly continues them.
  Missing content without a local copy cannot be recreated. Use the update page's
  actual restart status after installation.

Validation and known limits

• All 1,179 tests, type checking, lint, workspace-version and skill-sync checks,
  and the production build passed. The Beta XPI manifest, update file, download
  URL, compatibility bounds and SHA-512 were verified together.
• In isolated macOS Zotero 10.0.4, the original 0.5.0 package reproduced the
  missing-task-index error. Installing the candidate passed 16 recovery checks:
  history, notes, reports, titles and Chinese drafts returned automatically;
  original files and backups retained their hashes; new messages and full
  restarts worked without duplicate imports or automatic model calls.
• Complete data upgrades from the original 0.5.0 and 0.5.0 Beta 6 packages passed
  20 and 26 checks, retaining all 28/40 original history files, research records
  and consumed execution budgets. Seven hot-update checks passed for workspace,
  draft, sidebar and settings restoration, including already-closed windows.
• Release CI
  passed on Node.js 22 and 24 and published both assets as a prerelease; Latest
  remains 0.5.0. Consecutive anonymous discovery and both XPI download paths passed
  in CI and local verification. All 25 extracted public-XPI files match the
  tested candidate; the original public ZIP checksum was verified separately.
• Unchanged 0.5.0 and Beta 6 clients upgraded through their own real GitHub
  updaters to the public package, passing 58 checks: 19 for a 0.5.0 installation
  already failing with the missing-index error, 27 for complete research-data
  preservation, and 12 for Beta-channel upgrading. Recovery required no manual
  file changes; saved drafts, installed hashes, restarts and no-downgrade behavior
  passed. All three online runs passed without retries.
• Windows and POSIX paths are covered by automated storage tests. Native Windows,
  Linux, other Zotero versions, paid model calls and external Codex/Kimi execution
  were not revalidated. Unreadable files or unverifiable backups remain protected
  and are retried; missing data with no surviving local copy cannot be rebuilt.
• See the prepublication Beta acceptance record
  and automatic recovery guide.

Full Changelog: v0.5.0...v0.5.1-beta.1

Confucius v0.5.0

Confucius 0.5.0 brings literature discovery and delegated research to the stable
channel, together with model configuration and conversation improvements from
the 0.5.0 Beta series.

Added

• Search OpenAlex from a research question with year, open-access and sorting
  controls. One persistent literature capsule shows selected candidates against
  the retrieved, deduplicated pool; its editor retains selections and reading
  position across searches and reopening.
• Confirm paper imports and fulltext acquisition, reuse existing Zotero items
  and valid PDFs, retry individual failures, or open a paper and attach its PDF.
  Continue using abstracts or currently available papers while remaining downloads
  finish. Missing abstracts can be looked up from Zotero, OpenAlex and Crossref
  with bounded waits, cached attempts and source labels.
• Delegate scoped work to up to three concurrent research subagents, with
  additional work queued and a shared parent budget. Full-width entries open one
  centered viewer with the main conversation's Markdown, progress, tool inputs
  and complete receipts; filters, expanded records and position survive switching.
• Use the optional models.dev catalog for reference model IDs and capabilities,
  and customize reasoning levels and request formats for each Native endpoint
  and model. Applying reference metadata preserves the gateway, key and model
  alias; saved configuration remains usable without the catalog.

Changed

• New tasks start without implicit paper bindings and default to time grouping.
  Article grouping follows the task's creation sources; research started without
  an article has a separate category. Acquired papers remain evidence without
  multiplying sidebar entries.
• After accepting research material, hide the prominent acquisition action;
  missing fulltext remains available through a secondary link. New candidate
  changes still require review, and abstracts are not presented as read fulltext.
• Keep the literature capsule and jump-to-latest action aligned with the composer,
  and share conversation controls, keyboard behavior and high-contrast colors
  with the child viewer. Long traces load in batches and unchanged content stays
  in place while polling.

Fixed

• End parent waiting and release capacity when a child fails to initialize;
  stopped or retried attempts cannot publish late results into their successor.
  Restored executions no longer leave unfinished tool calls looking active.
• Show current tool results after reopening a collapsed group. Ignore obsolete
  disclosure events and out-of-order archive pages, preserve selection and
  focus during updates, and defer filtering until input composition completes.
• Honor the user's decision to continue research despite repeated acquisition
  calls. Late previews, searches and downloads cannot undo that decision or
  restart stale waiting; completed confirmations cannot steal another control's
  focus after task switches or further interaction.
• Recover missing release-list attachments through the selected release's asset
  endpoint, retaining channel, timeout, size and checksum validation. Publication
  verifies the anonymous contract used by older updaters and withdraws an
  unverified release to draft while retaining its original tag and files.

Upgrade notes

• This is a stable release, higher than every 0.5.0 Beta. Stable and Beta users
  can receive it through Confucius Settings → Update; enabling Betas is not
  required. Explicit channel and automatic-check preferences remain independent
  and survive upgrade, and disabling Betas never downgrades an installation.
• Existing tasks, drafts, papers, annotations, reports, memories, candidate
  decisions and recorded child traces are retained. Creation sources are frozen
  when older tasks load, excluding identifiable acquired papers. Complete
  original grouping and missing historical trace events cannot be reconstructed
  when earlier versions did not record them.
• No model or endpoint reconfiguration is required. OpenAlex settings are
  independent of model credentials. Catalog entries are reference metadata;
  provider-specific IDs and reasoning formats can differ.
• Installation restores an open workspace and its draft when hot update is
  supported; use the update page's actual restart status. Interrupted downloads
  remain retryable, and upgrading does not automatically restart research.

Validation and known limits

• All 1,165 tests, type checking, lint, workspace-version checks, skill-sync
  checks and the production build passed for 0.5.0. The candidate XPI manifest,
  both update files, download URLs and SHA-512 were verified together.
• In isolated macOS Zotero 10.0.4, normal installations upgraded from the original
  0.4.6 and Beta 5 packages to the candidate and passed 20 and 26 migration checks.
  Tasks, Chinese drafts, notes, annotations, report revisions, memory, saved
  execution budgets and all 28/40 original history files survived full restarts;
  Beta research selections and child traces also survived. Seven hot-update
  checks passed for workspace restoration, sidebar state and closed windows.
• Native interaction checks passed for literature research (29), subagents (25),
  sidebar navigation (29; 615 clicks) and model settings (9), including narrow
  windows, keyboard input, themes and reordered responses. A synthetic 4,131-event
  child trace loaded, filtered and exposed its final receipt in about 333 ms;
  idle polling made no trace DOM changes. Anonymous OpenAlex search also passed.
• Release CI
  passed on Node.js 22 and 24 and published all three assets as the latest stable
  release. Consecutive anonymous discovery and both XPI download paths passed in
  CI and local verification. The public XPI's 25 extracted files are byte-identical
  to the locally tested candidate; its ZIP checksum was verified separately.
• Ten unchanged published clients (0.4.3–0.4.6 and 0.5.0 Beta 1–6) upgraded through
  their own real GitHub updater to the public 0.5.0 XPI, passing 166 checks without
  retries or manual installation. Stable and Beta channel selections, installed
  package hashes, saved drafts and restarts passed. The 0.4.6 and Beta 5 online
  migrations additionally preserved the complete research-data fixtures described
  above. An already-current installation was not offered again, and disabling
  Betas did not downgrade it.
• Windows, Linux, other Zotero versions, paid model calls, external Codex/Kimi
  end-to-end execution and authenticated fulltext retrieval were not revalidated
  for this stable release. Model and most metadata checks use controlled fixtures;
  abstract coverage and available PDFs still depend on their providers.
• See the 0.5.0 prepublication acceptance record
  and literature guide.

Full Changelog: v0.4.6...v0.5.0

Confucius v0.5.0-beta.6

Download: confucius.xpi

Fixed

• Recover update discovery when GitHub's release list omits an already uploaded
  XPI by checking the selected release's asset endpoint. Keep checksum, size,
  channel and timeout safeguards, and preserve actionable errors on real failures.
• Upload and verify all release files in a draft before publishing. Check the
  anonymous update list and actual XPI download before completing the release job;
  prevent overwriting published releases and keep asset filenames visible.

Upgrade notes

• Old-version online updates have succeeded, but discovery remains
  intermittent. Use Confucius Settings → Update to check and install; a manual
  XPI installation is not a prerequisite. GitHub can still omit uploaded assets
  from its list, causing a retryable missing-package error. This remaining
  publication-compatibility defect has not passed stability acceptance.
• This is a prerelease. Include prereleases in Confucius Settings → Update
  controls future Beta discovery. Stable-only installations remain on the stable
  channel; disabling Betas does not downgrade an installation.
• Beta 5's published package is unchanged. This replacement release includes the
  update-discovery fix and all Beta 5 research improvements. No data migration or
  model reconfiguration is required; existing tasks and research data are retained.
• Source-code archives are not installable Zotero extensions. Installing the XPI
  preserves existing tasks, papers, settings and research history.

Validation and known limits

• All 1,155 automated tests, typecheck, lint, version checks, skill synchronization
  and build pass. Tests cover omitted assets, channel selection, invalid files,
  retryable network errors and a deadline covering both discovery requests.
• In isolated macOS Zotero 10.0.3, the installed candidate passed three live
  discovery checks against GitHub, 26 upgrade checks from public Beta 5 including
  full restarts, and seven window/sidebar hot-update checks. The discovery probe
  temporarily presented an older current version to reproduce Beta 5's missing
  list assets and used the real anonymous asset endpoint to recover them.
• This repair targets update discovery and publication. Paid model execution,
  authenticated fulltext retrieval, Windows, Linux and other Zotero versions
  were not revalidated.
• Publication check on September 23: the public XPI passed direct installation
  upgrades from Beta 4, Beta 5 and stable 0.4.6 (26, 26 and 20 checks), including
  full restarts and preserved data. The installed public Beta 6 code passed real
  asset-endpoint discovery, download, checksum verification and installation;
  the probe temporarily reported an older current version to exercise that path.
• Node.js 22/24 verification and draft-package checks passed, but the release job's
  legacy-client discovery gate failed because the anonymous list still omits
  assets. Initial old-version updater checks failed; these earlier failures are
  retained separately from the later successful checks below.
• Subsequent unchanged Beta 4, Beta 5 and stable 0.4.6 installations completed real
  online upgrades to this public package (33, 33 and 27 checks), preserving data
  through full restarts. Original Beta 1, Beta 3 and stable 0.4.3–0.4.5 also passed
  11 checks each. Beta 2 discovered the package but its download hit a network error.
• A continuous check in one unchanged Beta 5 installation returned five available
  updates and three missing-package errors across eight repeat checks. Successful
  upgrades therefore do not establish that discovery is consistently repaired.
• See the release acceptance record.

Full Changelog: v0.5.0-beta.5...v0.5.0-beta.6

Confucius v0.5.0-beta.5

Download: confucius.xpi

Changed

• Let research continue with abstracts before importing papers, or accept current
  results while remaining fulltext downloads finish in the background. The
  literature panel shows separate fulltext and abstract coverage among candidates.
• Fill missing abstracts on demand from local Zotero, OpenAlex and Crossref, with
  bounded waits, cached attempts and source labels. Abstract-only evidence remains
  separate from actually read fulltext.

Fixed

• Let users release research from fulltext waiting even when downloads fail or no
  PDFs are available. Repeated acquisition calls honor the same candidate decision;
  cancelled waits and late downloads cannot leave or restart a stale wait.
• Keep Continue responsive while remote searches are pending, preserve newer
  candidate decisions when search pages arrive, and discard confirmation previews
  that finish after continuing. Failed confirmations remain reviewable and retryable.

Upgrade notes

• This is a prerelease. Enable Include prereleases in Confucius Settings →
  Update, or install this release's confucius.xpi. Stable-only installations
  stay on the stable channel; disabling Betas does not downgrade an installation.
• No data migration or model reconfiguration is required for Beta 4 users.
  Existing papers, candidate decisions, child results, task history, drafts and
  saved settings remain available. Background downloads interrupted by a restart
  remain retryable; research does not restart automatically.

Validation and known limits

• All 1,144 automated tests, typecheck, lint, version checks, skill synchronization
  and build pass. The local Beta 5 candidate passed 24 literature checks, 17 subagent
  checks, nine model-settings checks, seven window/sidebar hot-update checks and
  26 upgrade checks from public Beta 4 in isolated macOS Zotero 10.0.3 installations.
• Chinese and English, light and dark themes, standard/narrow windows and full
  restarts were checked. Slow preview, search and abstract responses cannot undo
  the user's decision to continue. The model picker passed a complete isolated
  rerun after an initial check encountered a collapsed list.
• Model responses and most metadata use controlled fixtures. Anonymous OpenAlex
  returned 100 live papers and the models.dev catalog loaded online. Crossref
  fallback and deadlines were tested with injected responses; real abstract
  coverage, paid models, external Codex/Kimi end-to-end execution, authenticated
  fulltext retrieval, Windows, Linux, other Zotero versions and forced-colors
  were not revalidated for this Beta.
• See the release acceptance record
  and literature guide.
• Publication check on September 23: the public XPI passed direct installation
  upgrades from Beta 4 and stable 0.4.6, but GitHub's anonymous release-list response
  omitted this release's uploaded assets. In-app update discovery could not pass
  acceptance. This remains a publication-compatibility defect for older installed
  clients; a manual XPI installation does not establish that online upgrading is
  fixed. Later original-client upgrades to Beta 6 succeeded, but discovery still
  intermittently omits assets; Beta 7 was withdrawn after failed acceptance.

Full Changelog: v0.5.0-beta.4...v0.5.0-beta.5

Confucius v0.5.0-beta.4

Changed

• Present delegated research as full-width chat entries with live activity and tool
  counts. One shared, fixed-center bubble shows a searchable public trace, grouped tool
  inputs and results, model activity, conclusions and full evidence receipts.
  Switch children from their entries or the bubble's previous/next controls;
  filters, expanded records, loaded receipts and reading positions are retained
  per child while the bubble stays in place.
• Run up to three research subagents concurrently and queue additional work.
  Parent waiting status shows how many children have finished, and diagnostic
  exports include each available child's public process and evidence archive.

Fixed

• Release the concurrency slot and end parent waiting when a subagent fails to
  initialize. Ignore late callbacks from a stopped attempt after a retry, and
  retain external subagent tool results alongside their inputs in the trace.

Upgrade notes

• This is a prerelease. Enable Include prereleases in Confucius Settings →
  Update, or install this release's confucius.xpi. Stable-only installations
  remain on the stable channel; disabling Betas does not downgrade an installation.
• No data migration or model reconfiguration is required for Beta 3 users.
  Existing child results and recorded public events remain readable alongside
  task history, literature pools and candidate decisions. Missing events from
  older executions cannot be reconstructed.

Validation and known limits

• All 1,130 automated tests, typecheck, lint, version checks, skill synchronization
  and build pass. The local Beta 4 candidate passed 17 subagent interaction checks,
  17 literature checks, seven window/sidebar hot-update checks and 26 upgrade
  checks from public Beta 3 in isolated macOS Zotero 10.0.3 installations.
• Chinese and English, light and dark themes, 1,120/420/280 px windows and a short
  window were checked. Switching children preserves the same centered bubble
  and each child's reading state; late responses cannot replace another child's
  trace. Old child results, public events and research data survive full restarts.
• Model outputs and most searches use deterministic fixtures; anonymous OpenAlex
  retrieval fetched 100 live records. Paid online models, external Codex/Kimi
  end-to-end execution, authenticated full-text retrieval, Windows, Linux, other
  Zotero versions and forced-colors were not revalidated for this Beta.
• See the release acceptance record
  and subagent guide.

Full Changelog: v0.5.0-beta.3...v0.5.0-beta.4

Confucius v0.5.0-beta.3

Changed

• Collect each task's literature searches in one persistent capsule above the
  composer instead of repeating result cards in the conversation. The capsule
  shows candidate papers / retrieved and deduplicated papers, such as 9 / 100;
  the denominator is the task's actual pool, not the API's total search hits.
• Open the shared paper list above the capsule while keeping the conversation
  in place. Closing and reopening preserves candidate selections, filters,
  expanded abstracts and list position. Zero-result searches still offer a way
  to refine the query, and switching tasks clears the previous task's editor.

Fixed

• Keep the literature capsule and jump-to-latest button in separate spaces on
  the same row, aligned with the composer. Reuse workspace button feedback and
  fixed icon slots so hovering, pressing and expanding do not move the controls;
  narrow windows keep both actions accessible without overlap.

Upgrade notes

• This is a prerelease. Enable Include prereleases in Confucius Settings →
  Update, or install this release's confucius.xpi. Stable-only installations
  stay on the stable channel; disabling Betas does not downgrade an installation.
• No data migration or model reconfiguration is required for Beta 2 users.
  Existing task literature pools and candidate selections remain available from
  the capsule; previous search activity remains in the conversation.

Validation and known limits

• All 1,120 automated tests, typecheck, lint, version checks, skill synchronization
  and build pass. The local Beta 3 candidate passed 17 literature checks, seven
  window/sidebar hot-update checks and 23 upgrade checks from public Beta 2 in
  isolated macOS Zotero 10.0.3 installations, including full restarts.
• Chinese light and English dark windows were checked at 1,100, 420 and 280 px,
  including pointer feedback, aligned controls, keyboard use, empty results and
  retained selection/filter/scroll state. Existing Beta 2 literature pools,
  queries and candidate decisions survive upgrade and another restart.
• Model outputs and most search results use deterministic fixtures; anonymous
  OpenAlex retrieval fetched 100 live records. Paid online models, authenticated
  cached-fulltext retrieval, institutional login, Windows, Linux, other Zotero
  versions and forced-colors were not revalidated for this Beta.
• See the release acceptance record
  and literature guide.

Full Changelog: v0.5.0-beta.2...v0.5.0-beta.3

Confucius v0.5.0-beta.2

Added

• Look up reference model IDs, context windows, output limits and reasoning
  options from the open-source models.dev catalog in Native model settings.
  Applying an entry preserves the configured gateway address, key and model alias;
  saved settings and model requests remain usable without the catalog.
• Customize Native reasoning levels and their request format per endpoint and
  model, including provider-specific values such as ultra. Settings survive
  restarts and feed the composer and actual requests; switching models does not
  reuse another model's overrides. Catalog token-budget controls are not imported.

Fixed

• Filter catalog entries as you type in a styled, scrollable list inside Settings.
  Keyboard selection and input methods are supported; catalog and reasoning-format
  choices avoid native popup menus that can overflow Zotero windows.
• Clear the prior catalog selection when editing the filter, keep late responses
  from replacing newer results, and allow failed catalog downloads to be retried.

Upgrade notes

• This is a prerelease. Enable Include prereleases in Confucius Settings →
  Update, or install this release's confucius.xpi. Stable-only installations
  stay on the stable channel; disabling Betas does not downgrade an installation.
• No library migration or endpoint reconfiguration is required for Beta 1 users.
  Existing tasks, drafts, sources, credentials, notes, reports and history are
  retained. Custom reasoning settings are saved separately for each endpoint
  and exact model ID; Codex and Kimi keep their own model capability lists.
• Catalog entries are reference metadata, not a universal model ID standard.
  Check your provider's supported IDs and reasoning format before applying them.

Validation and known limits

• All 1,120 automated tests, typecheck, lint, version checks and build pass.
  Isolated macOS Zotero 10.0.3 testing passed nine model-settings checks, seven
  window/sidebar hot-update checks, and 20 upgrade checks from the public
  0.5.0-beta.1 package, including task preservation and full restarts.
• The catalog was fetched from models.dev. Custom reasoning_effort forwarding
  used a deterministic local endpoint; real paid model calls were not exercised.
  Metadata may differ from a gateway, and token-budget controls are not imported.
  Windows, Linux, other Zotero versions and forced-colors were not revalidated.
• See the release acceptance record
  and model settings guide.

Full Changelog: v0.5.0-beta.1...v0.5.0-beta.2

Confucius v0.5.0-beta.1

Confucius becomes a research assistant that starts with a question: discover
literature through OpenAlex, select candidates, obtain available full text,
and delegate focused research before synthesizing the evidence.

Added

• Search OpenAlex from the conversation with year and open-access filters,
  relevance/date/citation sorting, and cursor paging. Each page fetches up to
  100 papers; the deduplicated pool and API total hits stay distinct.
• Review recommended candidates in an inline literature card. When it leaves
  the viewport, a capsule above the composer opens the same editor without
  moving the conversation or losing selections, filters, or reading position.
• Configure, replace, clear, and test an OpenAlex key in Settings. Retrieval
  uses Zotero's built-in HTTP client and keeps the key out of model context.
• Confirm an exact candidate batch before import, binding, and download.
  Reuse existing items and valid PDFs, then try open-access and cached PDFs.
  Retry failures individually or open the paper in a browser and drop its PDF
  onto the corresponding item. Login pages and invalid PDFs are rejected.
• Delegate scoped research to isolated subagents across Native, Codex, and Kimi,
  with at most two running concurrently, a shared parent budget, durable results,
  and a read-only progress popup. Closing a popup keeps the work running.

Changed

• New tasks start without implicit paper bindings. Switching PDFs does not
  rebind the task. Explicit source edits update its current article grouping.
  Paper presets appear only for compatible sources and prepare
  the next request in the current conversation.
• New installations organize tasks by time; explicit existing choices remain.
• Preserve pools, candidate decisions, acquisition records, and completed
  subagent references across restarts and reply branches. Full pools and child
  archives stay outside the task index and are read through paginated tools.
• Present Confucius as a research assistant in both READMEs, with an OpenAlex
  walkthrough and a separate Beta download entry.

Upgrade notes

• This is a prerelease. Enable Include prereleases in Confucius Settings →
  Update, or install this release's confucius.xpi. Stable-only installations
  remain on the stable channel; disabling Betas does not downgrade an install.
• Existing tasks, sources, notes, annotations, reports, and history are retained.
  Literature pools and subagent records are added only when used. Back up both
  the Zotero library and Confucius runtime data before moving between versions;
  older releases do not expose the new research records.
• Configure OpenAlex separately from your model provider. No additional SDK,
  browser extension, or download service is needed for literature retrieval.
  Browser-assisted acquisition uses the access you already have; collecting
  papers alone does not automatically start reading them.

Validation and known limits

• Automated tests, typecheck, lint, and build pass. Isolated macOS Zotero testing
  covers prompt-driven search, card/capsule continuity, keyboard and task
  switching, candidate confirmation, real PDF import/rejection, targeted drops,
  delegated research, and restart persistence. Anonymous OpenAlex search fetched
  100 live records. Upgrade from the public 0.4.6 package and window/sidebar
  hot-update checks also pass, preserving existing tasks, sources, and content.
• Real authenticated cached-fulltext retrieval and online Native/Codex/Kimi
  model calls were not exercised: authorized test credentials were unavailable.
  Model outputs and provider failures use deterministic fixtures. Windows,
  Linux, other Zotero versions, and real institutional browser login were not
  revalidated for this Beta.
• See the release acceptance record
  and literature acceptance scope.

Full Changelog: v0.4.6...v0.5.0-beta.1

Confucius v0.4.6

Confucius 0.4.6 adds selection questions across readers and conversations,
refines paper reports, and fixes task, memory and tool reliability.

Added

• Select text in PDFs, reports or conversations to ask a question in a shared
  single-line input. Streaming answers and follow-ups persist in an independent
  btw history. PDF questions retrieve relevant completed conversations for the
  same article; report and conversation questions stay within their owning task.
  Read-only Native, Codex and Kimi runs leave the main task and memory untouched.
  Closing the popup keeps generation running; drafts and saved partial answers
  survive reopening. Network search is not included.

• Report preferences now guide source selection, explanations and organization
  for researchers entering an unfamiliar field. Readable reports receive one
  best-effort direct revision with the task's model and reasoning settings in
  an independent context. Failed improvements retain the saved report.

• Deep-read tasks offer independent report layout, writing voice and reading
  focus choices, with a live excerpt preview. Choices persist with the task
  across restarts, engine changes and branches, and apply to later revisions.
  Choosing later retains the draft without starting research.

• Reading reports support source passages alongside explanations and expandable
  language or technical help. Narrow readers stack paired passages; Zotero
  note exports expand the help so its content stays visible.

Fixed

• Collection and saved-search tools expose complete counts and continuation
  offsets. Research presets allow all resolved members, including papers beyond
  the first 100, while keeping their initial source preview bounded.

• Ordinary follow-up questions after a completed preset report can finish
  without generating another report. Continuing with attached files or task
  references now includes the newly submitted material.

• Response branches copy the report revision at the selected response into
  independent task-owned files. Older shared branch reports are separated when
  their historical revisions are still available.

• Repeated protection changes and retries after a rejected memory proposal
  remain actionable; retries of the same identified operation stay idempotent.

• Knowledge-base and report-to-knowledge writes retain their intended contents
  and stable creation IDs, allowing lost filesystem receipts to be reconciled
  after restart without blocking subsequent writes in other tasks.

• Reading a memory reloads its Markdown source and persists usage separately,
  preserving external edits and avoiding recreation of externally deleted files.

• Report writebacks preserve independent Zotero note and knowledge-base
  associations across target switches and report revisions.

• Optional MCP discovery runs in the background with a cancellable ten-second
  deadline per server. Unresponsive servers no longer delay workspace startup.

• Deep-read delivery no longer waits for a prescribed post-draft reading sequence
  or a quality verdict. Ready means deliverable; citation and atomic revision
  checks still apply. Explicitly unfinished drafts remain drafts.

• The report-style control and instructions apply only to paper-reading tasks.
  Clearing or changing the preset hides the control without leaving space in
  the composer; saved choices return when paper reading is selected again.

• Removed the PDF reader's annotation-batch toolbar filter and its repeated
  reader refreshes. Native annotation tags remain available; new annotations
  carry one batch-start timestamp tag. A background migration consolidates the
  two old batch tags on verified, editable annotations while preserving content,
  timestamps and unrelated tags.

• Artifact readers no longer use decorative dividers in the toolbar, citation
  section, Markdown, literature lists or annotation lists. Header buttons share
  the workspace styles and spacing in both regular and compact layouts; narrow
  windows wrap actions instead of shrinking their labels. An empty revision
  shows a placeholder instead of a blank title plus a stray rule.

• Focusing the artifact reader in Zotero no longer draws a full-width line
  below the header. Keyboard scrolling and control focus indicators remain
  available in embedded readers and separate windows.

• Legacy reading-companion reports render again: reports saved with an empty
  Markdown body plus a readingGuide fall back to the flattened guide text
  in the reader, artifact_read and model context, and stay editable via
  artifact_patch. New reports remain plain Markdown without companion data.

• Artifact saves advertised to models use the current task's body shape and
  allow extra wire fields. The host still drops placeholders before the strict
  save check, so a nested oneOf dump no longer blocks the first draft.

• Duplicate in-progress tool updates are not stored. The working clock follows
  the turn start, instead of resetting on every streamed fragment.

• An external engine that ends a turn without tools while required work remains
  is stalled instead of starting another empty prompt.

• Long attached paper titles in the workspace window ellipsize instead of
  shrinking the New task button.

Removed

• The reading companion (陪读), private passage questions, PDF-following guide
  view and on-demand report mode have been removed. Deep-read tasks again
  deliver a single research report directly. Formatted Zotero note saving and
  the unrelated bug fixes remain.

Upgrade notes

• This is a stable update from 0.4.5. Existing tasks, annotations, notes and
  report revisions remain available; no model reconfiguration is required.
• Deep-read tasks now deliver one research report. The old reading-companion,
  PDF-following and on-demand-report views are removed. Existing guide-only
  reports are shown as plain text. The new btw conversations are separate
  from old private passage discussions and do not import their history.
• Existing verified annotation batch tags are consolidated in the background.
  Report branches gain independent artifact copies when their historical
  revisions are available. Back up runtime data before downgrading; older
  versions cannot expose the new btw histories or report preferences.

Validation and known limits

• All 1,080 automated tests, type checking, lint, build and skill synchronization
  passed locally. Selection questions passed 12 isolated Zotero checks,
  including restart recovery; report-style controls passed 44 checks on
  macOS with Zotero 10.0.1. See the selection acceptance record.
• Native acceptance uses controlled model responses; Codex and Kimi are covered
  by adapter and isolation tests, without live service runs in this release.
  Other operating systems and Zotero versions were not retested. Selection
  questions do not include network search.
• Upgrading the published 0.4.5 package to the candidate passed 20 persistence
  and recovery checks, including unchanged history bodies and no duplicate note
  writes. Seven hot-update checks covered restored windows, sidebars, drafts,
  settings and closed-workspace behavior. See the release acceptance record.

Full Changelog: v0.4.5...v0.4.6

Confucius v0.4.5

Confucius 0.4.5 adds an article-centered reading companion, private passage
questions and formatted Zotero notes, and restores the workspace after hot updates.

Added

• Deep-read artifacts offer a reading companion alongside the research report.
  Ordered signposts and checkpoints pair original-language excerpts with content
  and writing explanations, source navigation, optional worked steps and self-checks.
  The continuous article uses concise rounded summary blocks, clickable excerpts,
  and a single bottom composer with selection quoting and answer bubbles.
  Artifacts fill the workspace by default, with Back and Separate window controls
  that preserve the current view and drafts. View tabs have roomier selected states.
  Summary blocks float without covering source links; the current PDF section stays
  lit with a warm background and soft glow. PDF page changes follow the matching
  passage, double-clicks focus it, and summaries link back to their starting page.
  Reading position, expanded sections, quotations and drafts persist locally.
• New deep reads deliver a reviewed companion first. Readers can generate a
  research report on demand in the same artifact, with version-bound writeback
  for either view. Existing reports and older task workflows remain compatible.
• Checkpoint questions run in private reading branches with independent model
  sessions, events, recovery and read-only paper tools. They are excluded from
  main-task history, global retrieval, memory extraction and report generation.
  Source revisions retain earlier discussions; adding a report keeps the branch.

Fixed

• Collapsing a passage conversation keeps its bubbles closed while focusing the
  composer. Questions retain their submitted text, and text typed while a request
  is pending remains in the draft. Successful answers no longer display their
  completion reason as an error.
• Artifact actions now say “Save Zotero note”. The save dialog uses clear view
  buttons and a full formatted note preview, with existing content folded away.
  Saved notes preserve headings, quotations, tables, equations and source links
  instead of displaying raw Markdown; confirmation remains bound to its version
  and selected view, excluding private reading discussions.
• New annotation archives identify their source PDF without treating each mark
  as a separate paper, so source-scoped tasks can reread their saved annotations
  while access to other papers remains restricted.
• Deep-read review tracks source content actually delivered to the model,
  including large annotation results restored from history. All annotation
  pages from the same snapshot must be read before the artifact can become ready;
  incomplete reviews identify the next unread offset and survive task recovery.
• Rejected artifact patches remain available as unapplied review candidates,
  preserving all proposed corrections when evidence review begins. Successful
  patch receipts report the number of applied text edits.
• Cancelling or timing out an artifact edit before its atomic write starts no
  longer saves another revision in the background. Writes already dispatched
  retain their actual outcome for recovery.
• Retry status distinguishes gateway timeouts, model service failures, rate
  limits and transport interruptions. Native event and checkpoint timestamps
  follow elapsed wall time, and trace export uses durable event sequence order.
• Hot updates automatically reopen the workspace on the new plugin, restoring
  the selected conversation, unsent text, conversation-list visibility and open
  settings tab. Both the separate window and sidebar support the handoff; a
  closed workspace stays closed. Older untracked windows are recovered and
  replaced instead of retaining a stale host and version label.
• Update status distinguishes a completed hot update from an installation staged
  for restart. A completed install waits for add-on startup and reports the new
  version without asking for a Zotero restart. Further update checks and channel
  changes remain available after a hot update.

Upgrade notes

• This is a stable release; prereleases do not need to be enabled. Successful
  hot updates restore open workspaces automatically. Restart only when the
  installer explicitly reports that the update is staged for restart.
• Existing reports, artifact revisions, task histories, drafts, annotation
  permissions and explicit update settings are retained. Older deep-read tasks
  keep their original completion requirements; upgrades do not automatically
  generate companions or rewrite existing content. New reading branches use
  separate local storage and do not migrate the main task history.
• See the reading companion guide
  for PDF following, private questions and saving either view as a Zotero note.

Validation and known limits

• The candidate passed 1,025 tests with no failures or skips, typecheck, lint,
  build, workspace version checks and skill synchronization.
• On macOS with Zotero 10.0.1, the installed candidate passed 20 companion/note
  checks, 28 navigation checks with 615 counted clicks, and seven hot-update
  checks. Upgrade and restart checks passed from public 0.4.4 (20) and
  0.4.3-beta.2 (21), preserving original history, drafts, sources, notes and receipts.
• Native model isolation and Codex/Kimi session contracts are covered by controlled
  executors. Real Codex tests delivered a reviewed companion after restart and
  answered in a private session; the final candidate also retained that session
  through an update and answered a follow-up. On-demand report generation
  completed in the same artifact, preserving the guide and private branch and
  excluding private questions. Kimi authenticated, read the paper
  and saved 14 highlights, but did not finish the companion within a 12-minute
  initial wait and a 15-minute resumed wait. Its full question/report flow remains
  unverified. A real Native service was not tested because no API key was configured.
• PDF following uses physical pages; same-page checkpoints retain the current
  selection. The companion must be open for PDF double-click navigation. No PDF
  margin overlay is included. Live coverage is macOS with Zotero 10.0.1;
  Windows/Linux and other declared Zotero versions have not been revalidated.
• The release acceptance record
  separates candidate installation checks from earlier
  companion engine and content checks.

Full Changelog: v0.4.4...v0.4.5