v0.47.0

New features

• Review daily problems in the Friction Log, including corrections, tool errors, and recurring patterns linked to exact messages. Browse dates, download Markdown, or read findings through the CLI and MCP. Enable it with [friction] enabled = true.
• Turn Friction Log patterns into Kata issues, link existing issues, or remove links from the page or CLI. Enable hourly automatic filing with [friction.kata] auto_file = true.
• Generate an optional model-written Friction Log summary with your configured insights provider. Summaries run only when requested and stay attached to the reviewed digest.
• Find the recorded session and prompt behind code with agentsview why file.go:120-140 or agentsview why --commit <sha>. List a session's recorded commits with agentsview session commits <id>.
• Turn on search by meaning from Settings or the search palette without editing configuration or restarting. Local installations use an available Ollama server or download the built-in embedding server with progress.
• Find sessions covering two to five concepts through MCP or the content search API. Results include a supporting passage for each concept.
• Browse saved Claude and Codex history from Docbank archives or backup snapshots alongside local sessions. Imported history remains searchable when the archive is disconnected.
• Find matching Docbank file versions beneath retained transcript images and attachments. The optional integration also exposes matches through the API and MCP.
• Import Open Interpreter sessions with their own label, usage totals, and resume commands. Command Code v3 sessions also import with thinking, tool calls, and token usage.
• Browse Kiro CLI subagent runs beneath their launching sessions, including nested subagents. The launching transcript shows each delegation's prompt and result.
• Choose message colors and a Color-blind safe chart palette in Settings > Appearance. Optional message markers, chart patterns, and dashed trend lines distinguish content beyond color alone.
• Diagnose an expensive session with agentsview session diagnose <id>. The report combines usage, context growth, compactions, large tool results, retry runs, and child-session links.
• Inspect per-session tokens and costs through MCP get_session_usage, including model totals and request breakdowns. Batch CLI lookups use agentsview session usage-batch <id>....
• Import xAI billing CSV files with agentsview usage grok <csv>. Reimports deduplicate billing records, but overlapping local sessions or Cursor billing imports may double count usage.
• Filter hosted sessions, searches, analytics, and usage by contributing source devices. Sessions supplied by several selected devices count once.
• Run PostgreSQL serving and watched local ingestion together with agentsview replica pg serve --push-watch. Add --embed to build and publish embeddings in the same process.
• Export visible dialogue with --text-mode visible, omitting structured tool and reasoning content. Reporting exports support hourly snapshots by originating machine with --schema-version 5 --by-machine.
• Discover commands from a task description with agentsview cli search. Inspect supported agents, session folders, and captured details with agentsview agents, without a daemon or archive.

Improvements

• Usage loads faster on large archives. In separate 30,000-session benchmarks, repeat visits drop from about 24 seconds to under a second on PostgreSQL and DuckDB.
• Activity, Signals, Trends, and session statistics load faster through reduced reads and reuse of unchanged results. ClickHouse session navigation also stays responsive while dashboard queries run.
• Session lists and long transcripts open and scroll with fewer stalls. Sidebar rows appear before their titles finish loading, and code highlighting blocks the page less.
• Sidebar and content searches return sooner, including substring, regex, and multi-term searches. The first local startup after upgrading builds a new search index before serving requests.
• Imports, archive rebuilds, and live Claude and Codex updates finish sooner. PostgreSQL pushes preserve unchanged transcript sections, DuckDB rebuilds write faster, and S3 sync reuses shared Codex parent downloads.
• Usage charts refresh automatically while visible and catch up when you return. Hidden tabs cancel pending Usage queries instead of holding browser connections.
• Usage charts omit zero-value tooltip rows and group positive series below 0.5% into Other. Token Attribution adds Linear, Sqrt, and Log choices for treemap label sizes.
• Activity timelines merge narrow bars into readable time intervals on smaller screens while preserving totals and selection accuracy.
• Read Claude hook-command output in collapsible cards controlled by the Hooks filter. Hook text is searchable and appears in HTML and Markdown exports.
• See recorded reasoning effort on Pi-family and Posit Assistant messages.
• Use grouped CLI commands under source, report, replica, and index. Previous spellings remain available for at least one release and print their replacements. Failures include specific recovery hints and structured JSON diagnostics where supported.
• Reduce background work while idle or on battery. Local sync pauses after ten idle minutes and resumes on archive access; scheduled embedding builds, Recall model passes, and archive audits wait for AC power. Set sync_interval to adjust watcher-triggered sync frequency.
• Keep using AgentsView while Claude.ai Sync runs by closing its dialog; only Stop cancels the run within the page. Closing or reloading the page still ends Sync. An organization allowlist also limits future imports after a restart.
• Recover useful search results when search by meaning is unavailable. Hybrid CLI and MCP searches return keyword matches with an explanation, and empty searches report matches hidden by session filters.
• Check stalled PostgreSQL and ClickHouse push watchers with status --max-stale. Status reports failed, overdue, or unfinished pushes with a nonzero exit code.

Bug fixes

• Usage totals include blank Claude thinking responses, unfinished Copilot CLI sessions, and Hermes or Augure Desktop background model calls.
• Codex Fast-tier sessions use Fast rate estimates. ClickHouse and DuckDB use the latest Copilot-reported session cost.
• VS Code Copilot and Positron replies keep updating as they grow, preserve pins and notes through rebuilds and pushes, and render without duplicate-tool-call failures.
• DeepSeek Harness sessions import when logs contain job lifecycle events or compressed frames without checksums.
• Claude channel and peer messages remain visible and searchable. Slash commands containing < in their arguments no longer disappear or leave answers beneath the wrong prompt.
• Codex forks link to their source sessions. Pi subagent runs and OMP parent links also receive the correct relationships.
• Archive rebuilds complete despite unavailable Codex fork parents while preserving saved child history. The Sync button performs required rebuilds and reports failures.
• Remote and temporary worktree sessions reuse established repository project names when recorded repository evidence supports the match. Explicit assignments and mapping rules retain priority.
• All-history Usage opens during startup price refreshes without internal errors. Concurrent Usage requests in a new time zone avoid database-lock failures.
• Analytics panels remain usable when another panel fails. Failed panels show their own error and Retry control, and old filter results no longer appear current.
• Activity's Sessions table keeps its selected sort order and column widths after syncs.
• Desktop windows reconnect after daemon exits or endpoint changes, accept localhost and IPv6 loopback endpoints, and avoid reloads caused by archive contention. The Linux AppImage also starts on affected Wayland systems.
• Explicitly started daemons stay running until stopped. Daemon restart also completes in Linux containers whose exited children remain unreaped.
• Claude and Codex Cloud transcripts show generated images inline with full-size viewing. Codex code-mode wait calls show their returned output, and Codex Cloud tool calls no longer appear twice.
• Hosted search and session browsing remain consistent during incoming uploads and publication. Empty source files upload successfully, and permanent upload rejections stop retrying indefinitely.
• Background skill discovery avoids protected macOS folders by default. Live watches recover after macOS session folders are recreated, and large Linux archives retain watches beyond 8,192 folders.
• Transcript-only archives stop reporting retries based solely on repeated tool names. Session links using an unambiguous raw UUID also open correctly.