What changed,
and where CodeSesh is going

Find Chinese history with fewer reads

Chinese searches confirm matching text only for the sessions needed to fill the results, preserving exact matches and ordering while avoiding unnecessary reads across large histories.

Do less work in the background

Workers scan sources after changes instead of polling every Agent. Pricing refreshes skip work already completed and avoid rewriting unchanged costs, reducing unnecessary work during Hub sync.

Retain less data after activity stops

Idle parser state, expired page snapshots, and outdated statistics caches are released. Slow live updates are merged, and failed detail refreshes no longer let inactive session caches keep growing.

Keep DeepChat references in replay

Messages that reference another DeepChat session now include its saved title and session ID as text, so replay preserves that context alongside the original message.

Understand your usage at a glance

Compare tokens, API value, and sessions with the previous period. Inspect daily token composition, model and Agent shares, and active hours, with estimates labeled clearly.

Find and browse the right session

Preview search results on wide screens as you move through them with the keyboard. Browse project sessions by day, see child sessions in context, and open tools in place during replay.

Learn and navigate with less friction

Search, projects, bookmarks, and preferences share one sidebar. The redesigned product site adds interactive previews, clearer guides, copyable commands, and a browsable release timeline.

Search history from 19 Agents

Find and replay Qoder, Qoder CN, Gemini CLI, Droid, and oh-my-pi sessions alongside your existing history, including tools and subagents. Droid usage is recorded per session; oh-my-pi currently supports the default profile.

See more of what happened

Claude Code workflow subagents and renamed sessions appear correctly. Codex timelines include built-in web search, tool search, and image-generation calls, while OpenCode history stays available during V2 migration.

Understand usage and estimated cost

Antigravity CLI usage and Kimi-Cli subagent usage are included, and identifiable replayed Codex fork usage is no longer counted twice. Estimates account for supported long-context and service tiers, one-hour cache writes, and corrected model prices.

Keep history collection reliable

Damaged text lines no longer fail an entire JSONL source. Idle transcripts stop retaining parser state, and temporary git timeouts no longer trigger unnecessary project-wide rereads. The first refresh after upgrading reparses cached history and may take longer.

Keep Worker sync moving

Uploads with nullable fields no longer get stuck on repeated checksum failures. Sessions whose incremental updates cannot be restored are shown as missing until a full copy arrives. Upgrade the Hub before its Workers.

Read replies, failures, and usage accurately

Kimi text replies and usage, OpenCode V1 cache and reasoning tokens, and Claude Code tool failures are read correctly. Codex reasoning tokens are no longer counted twice.

Find and inspect the right history

Search repairs a missing index and preserves parent session references. Agent links work without recent sessions, and deeply nested tool data can be expanded.

Keep daily statistics in your time zone

Dashboard daily totals, sparklines, and comparison periods follow the requested time zone, including when the server runs elsewhere.

Skip unchanged database sessions

Unrelated database writes in DeepChat, Cherry Studio, MiniMax Code, OpenCode, and ZCode no longer cause every session to be parsed again. Changes to OpenCode and ZCode child sessions still update their parent history.

Trace Worker interruptions

Timestamped structured logs record connections, pauses, recovery, and request or scan failures, with Hub error codes to help identify what happened.

Use less disk for the same history

Message data is stored compressed. On a real 11 GB cache, the size dropped below 8 GB. The first start after upgrading converts the existing cache in place, which can take several minutes.

Keep active sessions and dashboards responsive

Active Claude Code and Codex sessions resume parsing where they left off, and Dashboard and project statistics read compact usage summaries. A hidden browser tab holds updates until you come back.

Sync Workers with a fraction of the traffic

Workers send only new messages to the Hub instead of whole sessions, and idle Workers do much less disk and CPU work. Upgrade the Hub before its Workers.

Get correct results from filters and JSON output

File and kind filters in search match only activity from the session's own machine and respond much faster. codesesh --json completes even while an Agent is writing.

Search large histories with less work

Search indexes each message instead of whole sessions and skips embedded images. Chinese searches use the index, and long active Codex sessions refresh without re-reading the whole file.

Keep the local cache smaller

The session cache and Worker queue reclaim free space on open. The first start after upgrading rebuilds the search index, which can take minutes on very large caches.

See which machine each source comes from

Workers show host name, OS, and per-Agent session counts and last activity. Ignore an expected missing source, and filter projects by source node.

Know what each Agent supports

A capability matrix shows whether each Agent provides usage, recorded cost, tool results, reasoning, and session trees.

Understand collection status on each machine

See per-Agent discovery, collection progress, completion, missing sources, and errors from the local status panel or Worker details.

Read archived and compressed Codex sessions

Find history stored in archived folders or Zstandard-compressed rollouts. Compression and resume preserve the same session instead of creating duplicates.

Browse Antigravity CLI conversations

Read local SQLite conversations and tool calls with titles and workspace context. Support is partial: IDE history is excluded, tool outcomes and usage are unknown, and message times use the session update time.

Load activity statistics with less work

Dashboard activity-hour queries read the requested date range through an existing index, avoiding unrelated history and message content while keeping your source and project filters.

Resume synchronization after a Hub restart

Workers no longer wait on stale connections from the previous Hub process or a fixed one-minute retry. Existing pairing, confirmed progress, and queued history are preserved.

Locate slow startup and Dashboard steps

More detailed timing logs distinguish startup stages and Dashboard calculations, helping investigate delays when opening a large local or shared history library.

Find Agent history without a local installation

Hub projects and dashboards include archived and remote Agent sessions even when that Agent is absent from the Hub machine. Explicit Agent filters still apply.

Keep recent date ranges moving

Default 7-, 14-, 30-, and 90-day views follow the current local date, including after the server has been running for days. Ranges with an explicit end date remain fixed.

Follow practical guides in three languages

English, Chinese, and Japanese guides cover installation, session search, LAN access, Hub/Worker setup, background services, and token and cost interpretation.

Explore the product before installing

The redesigned site brings product previews, supported Agents, guides, and installation options into a clearer tour, with light and dark themes and mobile navigation.

Keep history collection and uploads moving

Workers continue scanning unfinished history and send the next queued upload as soon as the previous one is confirmed. Failed requests still wait before retrying, and pending content stays saved locally.

Understand why a rescan started

Task details distinguish a manual rescan from Hub recovery or Worker replacement, so an automatic recovery is easier to understand.

See scanning and upload confirmation separately

Available progress shows scanned source items for a pending Agent. Older or offline reports are not presented as current counts, and a task finishes only after its uploads are confirmed.

Recover from stale synchronization state

Workers refresh their recovery state when the Hub changes again, instead of repeatedly retrying an outdated recovery and remaining offline in the Hub.

Keep queued history available for upload

Recovery preserves queued session content even if the source log is no longer available. It also handles a lost recovery acknowledgment followed by another Hub change.

Search Chinese text and open the matching message

Find short Chinese terms within longer text and jump from a message result to its first match. Search explains the active scope, time range, and 50-session display limit so you can refine the results.

Read long conversations without loading everything first

Messages load in pages as you approach the bottom. Retry a failed page or choose Load all messages, while keeping your reading position. Markdown copying and receipts still use the complete conversation.

Know which machine a session belongs to

Search results identify the source machine. Session details show its sync status and explain where to run a copied resume command, helping you return to the right environment.

Reduce repeated work as history grows

CodeSesh reuses unchanged messages and list pages, limits statistics reads to the selected sessions, and reduces temporary memory used by Hub/Worker synchronization.

Review Codex threads and recorded answers

Read user messages and final replies from thread lookups, expand more turns when needed, and see asynchronous questions alongside the user's recorded answers.

Inspect Code Mode output and source

DeepChat, Codex, and Pi Code Mode calls show text and images in their original order. Expand and copy the source, inspect failures, and review available Pi execution details.

Read tables and trust recent activity

Agent replies render Markdown tables with scrolling contained inside the message. Codex settings and other metadata no longer move old chats into recent activity; the next scan corrects saved timestamps.

Read less when opening saved history

Dashboard queries use indexes for session summaries, usage, and activity. Bookmarks load their saved sessions without reading the full archive.

Diagnose the complete page load

Loading logs now cover projects, the dashboard, and bookmarks alongside sessions. Slow query logs show where time is spent, helping diagnose delays on a local server or Hub.

Sync session counts without blocking uploads

The Hub now accepts count changes in metadata-only updates, including when Codex records new environment context or developer instructions without adding replayable messages.

Keep existing conversation content intact

The Hub saves the updated count without rewriting stored messages or search content, preserving the conversation available for replay and search.

Bring multiple machines into one view

Pair Workers with your Hub and filter sessions by source. Standalone mode remains available; Workers send session content and metadata to the Hub you choose.

See collection problems and request rescans

Inspect queue status and Agent errors, queue rescans for selected Agents, and review task history. Offline Workers can collect changes for later upload.

Replace a Worker without losing its identity

Guided pairing covers local, LAN, and public HTTPS connections. Replace a Worker while retaining its archived history, bookmarks, aliases, and project groups.

Install and manage local data more easily

New native installation channels include curl, Homebrew, and Scoop. Databases, pricing cache, and logs move into ~/.codesesh/ with a guided migration from older locations.

See where session usage comes from

Review usage and available costs by model and provider, including uncached input, output, cache reads, and cache writes. Estimated costs are labeled, and unavailable details remain unspecified.

Save the complete receipt

Export a CodeSesh-branded PNG that includes the entire receipt, even when its details extend beyond the visible drawer.

Load less data when reopening history

Session lists restore from compact display metadata, avoiding a read of pricing details that the list does not need.

Choose a standalone executable or npm

Use a native executable on supported macOS, Linux, and Windows systems, or keep launching with npx codesesh. The npm launcher requires Node.js 22 or later.

Keep browsing while prices refresh

Model prices refresh in the background. Updated estimates use cached usage instead of rereading your conversation files.

Reuse work when reopening your history

Cached session metadata, incremental scans, and reduced snapshot decoding work limit what CodeSesh needs to process again at startup.

Replay OpenCode V2 conversations

Search and replay saved messages, reasoning, tool results, child sessions, and compaction records. Copy a resume command to continue a session in OpenCode.

Usage that respects forks and subtasks

Session usage and costs follow OpenCode's recorded totals, so copied fork history and child-session usage are not counted again. Explicit zero costs stay zero.

History follows your selected database

Choose a custom OpenCode database and keep saved message changes in sync. Switching database paths refreshes cached sessions even when the selected database is older than the last scan.

Stop a background refresh without an error stack

Closing CodeSesh while sessions are being scanned or indexed no longer prints a session refresh failure caused by the shutdown itself.

Cancel history backfills with the same clear feedback

Stopping while older sessions are being loaded follows the same cancellation handling, keeping expected shutdowns out of failure reports.

Follow MiniMax Code sessions and subtasks

Browse parent and child conversations, review reasoning, tool details, and usage, and copy a resume command to continue a session in MiniMax Code.

Keep saved history current

Automatic refresh picks up new messages, edits, and deletions in MiniMax Code, including writes that directory watchers can miss.

More room for content on mobile

The product site's opening artwork is shorter on mobile screens, leaving more space for the page content.

DeepChat conversations alongside your coding sessions

Browse native and ACP sessions from DeepChat’s current unencrypted local database, replay messages, and review usage as records update.

Cherry Studio Agent sessions and assistant chats

Read Cherry Studio 2.x Agent sessions with their workspaces and follow the selected branch of assistant chats, with usage tied to that branch.

Clearer dates, messages, and small costs

Date separators and layouts that adapt to narrower screens make long conversations easier to follow. Positive costs below one cent are shown as <$0.01 instead of $0.00.

Guides to finding your local history

English, Chinese, and Japanese guides explain where Claude Code and Codex records live, how to search and replay them, and what to check when a session is missing.

Model token shares

Compare model token totals and percentages in overview and project dashboards, with details available by pointer or keyboard.

Usage within your selected dates

Model totals follow the dashboard date range, so activity outside that range no longer enters the comparison.

Consistent navigation and lighter loading

Navigation gains shared hover feedback, agent introduction pages load on demand, and product site assets use caching and homepage artwork preloading.

Active hours by time of day

See when you collaborate with AI agents throughout the day, grouped by morning, afternoon, and evening periods with intensity cues.

Resilient live synchronization

Reconnection and transient stream failures recover automatically, keeping session lists and dashboard aggregates in sync without manual reloads.

Retryable sessions and durable aliases

Temporarily unavailable session details can be retried directly, and custom session aliases remain intact during concurrent live updates.

Native trilingual interface

Switch seamlessly between English, Simplified Chinese, and Japanese with preferences saved across sessions, while session transcripts and code remain in their original form.

Live-updated model pricing

Automatically caches up-to-date pricing from models.dev before scanning, keeping token cost tracking accurate as model catalogs expand.

Faster full-history analytics

Database-indexed usage queries and streamlined tree traversal deliver snappy dashboard metrics and smoother browsing across large archives.

Copy sessions as Markdown

Move a complete conversation into documentation, an issue, or another tool without rebuilding its structure by hand.

Refresh only what changed

Incremental scans update changed search documents instead of rebuilding the entire index, reducing work on large archives.

Keep interaction responsive

Cold session details load outside the main thread, and live project and session state stays consistent through reloads.

A complete Japanese product site

Japanese-speaking visitors can now understand CodeSesh, its local data boundary, and its core workflow in their own language.

Stronger local access controls

Loopback API requests are authenticated, and trusted-proxy deployment now requires a safer host and HTTPS configuration.

Last known-good state survives failures

Interrupted scans, unavailable agents, and failed cache migrations no longer replace usable history with partial state.

Usage follows message time

Sessions spanning several days no longer assign all tokens and cost to a single day in the dashboard.

Failures stay explicit

Agent, scan, cache, and Web failures no longer silently appear as empty or successfully published state.

Less repeated work

Cached metadata, reused queries, and localized rendering reduce startup, search, and live-update overhead.