Work Intelligence 1.6.1

繁體中文更新說明.

Changed

• The architecture diagram reader lays out cards without authored positions along their connections: groups become blocks placed left to right by the links between them, blocks in the same column stack and center, and cards inside a group form sub-columns by their own links. Cycles turn back at the node declared first. Authored position values never move, and automatic cards are placed clear of them.
• Connections use orthogonal routes with rounded corners between the facing sides of two cards. Links that share a side get spread ports, routes go around cards in the way, parallel runs keep separate lanes, and among clear routes the one with the fewest crossings and bends wins.
• Edge labels are shown on the line in a pill (long labels are shortened; the full text stays in the tooltip). Each group gets one of six theme tones for its frame, its cards and its outgoing connections; hovering, focusing or selecting a card highlights its connections and dims the rest.
• Removed per-section refresh buttons whose data already updates on its own through the SQLite change stream and Agent request polling: Agent reads, open-item cleanup, metadata backfill status, AI report synthesis status, report presentation and the work report page. The header refresh, error retries, System status refresh and the backup list refresh (service and backup files live outside SQLite) stay.
• AI synthesis re-run buttons no longer read 重新整理: the small one says 重新產生／Regenerate and the main one 重新產生這份報告的整理. Knowledge page update/creation requests use the sparkles icon instead of the refresh icon.

The architecture reader is frontend only: the architecture JSON format, the database schema (30) and the MCP contracts are unchanged.

Work Intelligence 1.6.0

繁體中文更新說明.

Added

• pnpm check (and therefore the last step of pnpm bootstrap) reports the free pages inside the database file as reclaimable space. With at least 1 MiB free and either 5 MiB or 20% of the file, it warns and suggests stopping the server and Agents and running pnpm db:maintain. The check is read-only and never runs VACUUM itself.

Changed

• Backup retention is now 3 for manual backups, 3 for automatic daily backups and 3 for the Session-deletion snapshots (was 14 / 14 / 5). WORK_INTELLIGENCE_BACKUP_KEEP still overrides the manual count. Older backup files beyond the new limit are removed the next time a backup is written.
• Work-summary update history is pruned: for session_work_summary_updates rows older than 7 days, the stored previous/resulting work summary (only used to replay a retried idempotencyKey) is emptied. The request itself and the row stay, so the same key with different content still returns a conflict. The prune runs when the API server starts and then hourly with the backup check, needs no migration and never runs VACUUM; reclaim file space with pnpm db:maintain if wanted.
• Retrying an old work_update_session_work_summary key whose stored response was pruned returns work_summary_updated with duplicate: true and the Session's current work summary (without previousWorkSummary) instead of not_found.

Database schema remains 30; no migration is required.

Work Intelligence 1.5.3

繁體中文更新說明.

Added

• pnpm track <folder> [--name <name>] [--yes]: a terminal-only way for the user to add a project and set it to Tracked. It resolves relative paths from where the command was run, previews the name, path and status change (with a warning when roots overlap other projects), and asks [y/N] before writing; without a terminal and without --yes it only previews. Agents (MCP) cannot track a project.
• pnpm start now prints a short reminder when an installed Agent skill copy or the Codex save-reminder hook is out of date, and points to pnpm bootstrap and pnpm check. The check only compares local files, prints nothing when everything is current or not installed, swallows every error, and never delays or blocks startup.
• pnpm check as the main name for the read-only diagnostics; pnpm run doctor keeps working. pnpm bootstrap step 5 and the docs now use the new name.
• When pnpm db:maintain or pnpm db:redact --apply cannot lock the database, the CLI now says why: whether the Work Intelligence server is running (and how to stop it), how many Agent MCP connections are active, or that another program may have the file open. The checks are read-only and never replace the original error.
• A bootstrap section in the release checklist for the Windows interactive install and update flow.

Changed

• scripts/run-pnpm.mjs no longer passes an argument array together with shell: true on Windows, which removes the Node.js DEP0190 warning when scripts are run directly with node; arguments are quoted for cmd.exe and unsafe characters are rejected.
• MCP tests get the same 20 s timeout as the storage and server tests (30 s for the recall-evaluation tests), so Windows coverage runs no longer time out while SQLite waits on a briefly locked file.

Database schema remains 30; no migration is required.

Work Intelligence 1.5.2

繁體中文更新說明.

Added

• pnpm bootstrap as the single install path: checks Node.js and pnpm, then installs, builds, connects Agents (pnpm setup:agents), optionally sets up start-at-login and runs the doctor. Global Agent and service settings keep their own preview and confirmation. Flags: --yes, --skip-agents, --service, --help. Re-running it after git pull updates the checkout; when the login service is already installed it reminds you to restart it with pnpm service:install instead of asking again.
• Doctor lists top-level *.sqlite files in the repository root (read-only) and reports a Windows drive-relative WORK_INTELLIGENCE_DB as an error.

Changed

• pnpm setup:agents skips the manual MCP, skill and hook registration for an Agent whose Work Intelligence plugin is enabled, so plugin users do not get duplicate tools and reminders.
• WORK_INTELLIGENCE_DB is resolved the same way by the server, MCP, doctor and login service: relative paths resolve from the repository root, and a drive-relative path such as C:Users... (backslashes eaten by a shell) is rejected instead of silently creating a database in the working directory.
• The minimum Node.js version is 22.13 everywhere (node:sqlite needs no flag from 22.13); engines.node is declared in the root package.json.
• The README quick start is git clone → pnpm bootstrap → pnpm start; plugin and manual registration are listed as alternatives.

Database schema remains 30; no migration is required.

Work Intelligence 1.5.1

繁體中文更新說明.

Fixed

• Save all eligible report synthesis sources and read their fixed IDs in pages, even after newer records push them outside the current-period 200-Session window. Report unavailable sources explicitly and scope evidence pagination to selected sources.
• Separate first-load, failure and empty synthesis states; guard creation while source status is unknown. Display saved citations separately from current-period totals and let readers expand long executive summaries.
• Preserve outstanding-item and cleanup-list reading positions during live updates by keeping stable item IDs.
• Bound pending report source pointers in general Agent context without discarding stored request sources.
• Keep the current report version before history when creation timestamps match, including bounded history pages.
• Omit reconnect recommendations when the MCP runtime is healthy.

Added

• Default MCP Agent report previews with a configurable 24,000-character response budget, explicit omission metadata and full-source lookup instructions. view: full preserves the complete legacy response; REST report fields remain full.
• On-demand English STE-inspired and plain Traditional Chinese writing guides in the checkout, MCP resources and distributable plugin. Synthesis contract v4 keeps coverage diagnostics in one section and preserves factual and verification distinctions.

Database schema remains 30; no migration or historical content rewrite is required.

Work Intelligence 1.5.0

繁體中文更新說明.

Changed

• Refresh both READMEs and the Traditional Chinese 1.5.0 release guide; retire 17 historical OpenSpec handoffs, preserving two unique audits under docs/experiments/.

• New diagrams require explicit architecture JSON v1. Historical Mermaid stays readable as escaped, copyable, downloadable source; remove Mermaid, KaTeX and their renderer dependencies without rewriting stored diagrams.

Fixed

• Use the full available content width on every page and reserve stable scrollbar space, preventing width jumps when navigating to or from the graph.

• Keep cleanup request creation available during background list refreshes after the current project data has loaded; retain first-load and active-request guards.

• Add 12px content padding to Session related-work hints and align titles, dates and shared paths, with wrapping at narrow widths.

• Scale all 53 activity weeks to the available container width on load and resize, with square cells and responsive month labels; retain keyboard navigation and daily report links.

• Keep committed Knowledge creation and decision promotion successful when SSE supersedes a background refresh.

• Fit the timeline to its measured viewport automatically on first load, period/project changes and returning from list view; keep manual zoom during resize. Full-year ranges fit tablet widths.

• Move the dashboard activity heatmap directly below the page header into a full-width section, ahead of stats, weekly insights and attention lists.

• Reminder display preferences support hide, snooze for seven days and restore with source/preference CAS. Unresolved totals stay authoritative, changed sources resurface, and migration 30 includes portable transfer and permanent deletion.

Added

• Agent Session reads support optional completion/handoff views and a fixed field whitelist with policy-gated, snapshot-consistent SQL loading. Keep full legacy reads unchanged, retain source safety flags and expose missing historical fields; projected structured metadata avoids duplicating source content.

• Preview and copy basic or current report-presentation Markdown with source gates, revision checks, escaped text and a download/manual-copy fallback. Clipboard success is shown only after a successful write.

• Dashboard weekly insights use full report totals with verification priority and scoped links.

• Report synthesis versions now have independent section pin/hide/restore controls and manual title/detail edits. The original summary and source references stay intact, edits are audited and portable, and concurrent changes keep the draft for an explicit comparison before applying.

• Suggest up to five same-project Sessions sharing confirmed files in Session detail. A covering file-posting index bounds hot-path reads, excludes existing links/voided or unavailable sources and reports partial coverage; suggestions never create links.

• Aggregate eight existing reminder sources in the Overview with project/type filters, URL pagination, bounded scrolling, explicit failed/partial coverage and policy-gated domain links. Knowledge links open the exact source history, and metadata backfill retains its project/request scope. Dashboard week dates come from the server clock; source status is changed only in its owning workflow.

Documentation

• Document the staged Mermaid removal plan and the M1/M2 implementation completed for this release, covering native-format writes, safe historical-source reading, dependencies, data lifecycle and verification. Historical stored diagrams were not batch-rewritten.

• Document the five-feature integration study for attention aggregation, related-work suggestions, report section controls, dashboard insights and report copying, with UX wireframes, data/security boundaries, rollout criteria and a synthetic candidate-query probe. The study records staged research; completed production implementations for this release are listed above.

Work Intelligence 1.4.0

Added

• Session verification adds in_progress (進行中), with an audited edit option, Agent schema/contracts, report and dashboard counts, timeline visuals and portable export/import. It stays distinct from passed, failed, not run and historical missing verification; it does not change the finalized-record lifecycle. Report risk classification collects source ids in one pass.

• Session diagrams accept validated architecture JSON v1 alongside Mermaid, with a bounded node/group/edge/path schema, versioned storage migration, portable import/export, structured secret masking and an MCP schema resource. The Web renders architecture snapshots as grouped cards with sources, direct relations and author-defined paths, with zoom/pan and escaped-source fallback.

• A reproducible, opt-in Archify architecture probe compares a pinned upstream renderer with existing Mermaid across production CSP, themes, locales, three browsers and 6/10/100/500-node synthetic diagrams. The findings retain Mermaid and choose validated JSON plus a project-owned viewer for structured architecture diagrams.

• Session diagrams have an expanded reader with a resizable wide panel, zoom buttons, actual size, fit to view, mouse/touch drag to pan and keyboard navigation. The selected diagram is kept in the URL so reload reopens it. Mermaid remains under the existing strict CSP; invalid diagrams retain their source and offer retry.

Fixed

• Cached synthetic performance databases now rebind their project roots to the current run’s temporary directories; deleted roots from an earlier run no longer cause policy-gated fixture setup to skip. The cache file itself is unchanged.

• Pin Mermaid’s transitive KaTeX to the patched 0.18.2 release for GHSA-238p-pmpm-9mq7 (inherited renderer settings could bypass trust restrictions when another component had already polluted Object.prototype). Keep strict Mermaid security/CSP and test normal math plus rejected untrusted links in an isolated process.

• Mermaid's root-level htmlLabels: false now keeps labels in SVG; the diagram-specific setting is deprecated and was ignored by the current renderer, causing excessive spacing in larger flowcharts.