SourceLens 0.63.6

SourceLens 0.63.6

Released 2026/10/29.

Fixes

• Keep long-running Runs active while a LensNode is processing and cancel stale execution after the control plane marks a timeout.

SourceLens 0.63.5

SourceLens 0.63.5

Released 2026/10/28.

Improvements

• Chat run progress numbers each plan step and shows how long that step took. A running step keeps counting up, completed steps keep their duration after the conversation reloads, and restored history omits durations instead of guessing. Active and failed steps are now clearly distinguished by color.
• End retrieval runs early once further searching stops producing new evidence, and close the answer with what was confirmed, what could not be confirmed and why, plus what narrower query would settle it.
• Administrators: LensNode heartbeats and disconnect logs now report event-loop lag, send time and outbox depth, so control-channel drops can be traced to the code that stalled.

Fixes

• Fixed code-analysis tools returning empty results, and Runs that reach the graph step ceiling now finish with the evidence already gathered instead of failing.

SourceLens 0.63.4

SourceLens 0.63.4

Released 2026/10/22.

Improvements

• Chat runs collapse into a scannable summary with step counts and duration, answers suggest follow-up questions, and documents, tables, and shared Q&A pages use more compact responsive layouts.

SourceLens 0.63.3

SourceLens 0.63.3

Released 2026/10/22.

Improvements

• Administrators: Simplified the Assistant configuration page: deep thinking is now decided by the end user in each conversation, the unused LensNode picker is gone, and the run budget fields no longer fail to save when cleared.
• Shared Q&A links of a public Assistant can now be opened without signing in. Administrators choose the access policy globally, per Assistant, or let the publisher decide per link.

SourceLens 0.63.2

SourceLens 0.63.2

Released 2026/10/21.

Improvements

• Replace run-depth tiers with a Deep thinking toggle. Normal mode skips independent answer review for faster responses, with explicit reasoning requests, separate review status, and optional administrator execution budgets, while chat users can show or hide citations and Agent-kit responses omit platform citation markers.

SourceLens 0.63.1

SourceLens 0.63.1

Released 2026/10/20.

Improvements

• Compact internal Q&A feedback without losing original user input, discover authorized integration tools on demand, and use reasoning effort appropriate to planning, execution and review. Batch adjacent streamed text while preserving output order and reconnect delivery.

Fixes

• Keep report composition and HTML formatting in the selected Skill, while source Plugins provide facts and the runtime preserves evidence and delivery checks.
• Separate execution completion, source coverage and answer review. Limited coverage and unavailable reviews remain visible as warnings, while missing requirements and unsupported answers still produce partial results. Use one authoritative answer review when configured to avoid conflicting decisions and redundant model calls.

SourceLens 0.63.0

SourceLens 0.63.0

Released 2026/10/20.

Features

• Add a paginated Run attachment endpoint for Agent clients, with documented metadata and private downloads. Agent credentials issued by staff also require session ownership and cannot download unpublished files.
• Agent Integration offers revocable access tokens with 1, 3, or 6 months of validity, or no expiration. View active authorizations in settings and revoke individual tokens immediately. Previously issued JWTs remain valid until their original expiry.

Fixes

• Include published output files in Q&A run responses and allow agent credentials to download them through the existing private file endpoint. Session ownership and draft publication checks remain enforced.

SourceLens 0.62.2

SourceLens 0.62.2

Released 2026/10/20.

Improvements

• Q&A alerts show a chronological task time interval and place the question above diagnosis results. Long questions use a short excerpt, and details links replace repeated Run identifiers.

SourceLens 0.62.1

SourceLens 0.62.1

Released 2026/10/19.

Fixes

• Long runs recover from malformed summaries and share summary usage accounting. Activity pagination fills unread gaps and stops when complete. Models judge report wording without keyword blocking. Maximum execution keeps unlimited tokens and bypasses the default tool-call ceiling while honoring explicit limits. Original evidence survives compaction and checkpoints with retry order preserved. Review capacities agree across API and runtime. Reviewers check current attachment bytes against canonical source snapshots, replacing duplicate pages and stale draft reads. Repairs receive the same evidence and candidate, keeping drafts editable within bounded correction phases. Failed local diagnostics do not automatically mark sources incomplete. GitHub retries transient errors, short rate limits and incomplete search responses; unresolved source or review gaps still produce partial results.

SourceLens 0.62.0

SourceLens 0.62.0

Released 2026/10/19.

Features

• Administrators: Unified synchronization, processing and Q&A alerts with topic-based messages, personal email subscriptions, bounded diagnosis, multilingual notifications, native Feishu/WeCom/DingTalk targets, destination deduplication and retryable delivery records.

Improvements

• Administrators: Notification management focuses on channels, event subscriptions, and records. Compact subscription tables open detailed event and scope information from the subscription name.
• Administrators: Notification records show readable titles, event types, source users, channels, times, and concise statuses. A detail drawer provides context, delivery results, diagnosis, and navigation to the associated Run.
• Administrators: Event subscription editing groups events, receiving channels, scopes, and advanced settings, with consistent checkboxes, select-all controls, and preserved existing delivery settings.

Fixes

• Fix session lists getting stuck when archiving or restoring conversations, let document conversion continue after a parser timeout with accurate progress, and allow automatic answer review to handle larger evidence sets.
• Improve Q&A reliability with literal phrase search across converted documents, preserved source paths, and retained tool evidence when preparing final answers. Reduce model gateway cold-start failures and let disconnected answer streams close promptly.
• Fix activity reports stopping after invalid context summaries or reading only one work record per page. Allow pending report files a bounded opportunity to finish saving near the token limit, retain source coverage warnings, and explain why reports with unreviewed evidence or unsupported conclusions are withheld. Retry unparsed tool responses once with format correction, keep activity retrieval focused on canonical evidence pages, and deduplicate structured review material without dropping facts or increasing its budget.
• Automatically validate and attach final files generated in the output directory, including files produced by document tools. Remove the separate save step, retain evidence checks, and prevent rejected drafts or failed replacements from appearing as downloadable results.
• Administrators: Notifications include the assistant, initiating user, UTC timestamps, and question context, with readable Feishu cards and working audience-specific detail links. Notification statistics can recover from loading failures.
• Keep tool evidence available when reviewing a revised or budget-limited final answer. Unify answer and attachment completion checks, preserve source gaps, and retain available output as unverified when the final reviewer is unavailable. Failed review calls no longer default to a passing verdict. Preserve bounded source results before large tool output is moved to files. Honor the configured decision timeout throughout the review gateway. Skip malformed summaries immediately instead of repeating identical requests. Show explicitly when a failed summary preserved the original history. Keep large TypeSafe review evidence intact using extended context. Accept report statements about missing completion records without treating them as zero-work claims.