Recover lost context. Trace forks and compactions. Debug how your agents actually worked.
Swob parses local histories through 12 native-format adapters and 1 Claude-compatible format. Where a source exposes the evidence, Swob adds lineage, incremental SQLite FTS5 search, execution inspection, provenance-aware audit, and optional AI Insights.
Website · Website source · Verified releases · Changelog
Important
v1.3.1 remains the current public installer; main may contain unreleased capabilities. The feature images below are English demo reconstructions based on the source UI; identifying text and sample data were localized and privacy-sanitized for publication. They show implemented layouts and workflows, not untouched production-data captures. Counts inside the images are illustrative and separate from the audited corpora below.
Current main · English demo reconstruction · privacy-sanitized illustrative data
AI coding sessions are not isolated chat files. A resume creates another file. A fork splits a task. A compaction replaces earlier context with a summary. Different agents store the same kind of work in incompatible locations. A normal viewer can show one transcript; it cannot explain where that transcript came from or why the agent lost the plot.
Swob treats session history as evidence:
- Trace lineage — navigate verified fork and continuation relationships in an interactive, force-directed Session Galaxy.
- Recover context — expand compacted Claude Code turns and preserve local backups after the source has disappeared.
- Debug execution — inspect tool/agent calls, token pressure, compaction boundaries, latency, framework overhead, errors, and anti-patterns.
- Search parsed histories — SQLite FTS5 indexes normalized messages incrementally. Formats without reliable body parsing are excluded, and source-specific search gaps stay visible.
- Resume safely — return to a supported CLI with its session ID and working directory, with source-aware validation.
| Audited result | What it means |
|---|---|
| 253 / 1,621 | 253 Claude Code source sessions in one audited library were already missing under the default 30-day retention policy; Swob still had local backups. |
| 93.58% | Verified resumability in the same 1,621-session, five-source audit corpus. This is an observed corpus result, not a universal success guarantee. |
| 1,704 sessions | Current local performance and UI corpus used to exercise the new index and dashboard. |
| 12+1 sources | Current main reads 12 native harness formats and 1 compatible format. Capability status remains source-specific and evidence-bound. |
The current graph is a real Canvas-based, force-directed view. It distinguishes verified lineage edges from softer project/source/time grouping and lets you pan, zoom, inspect, and open a session. A prior PixiJS prototype was deliberately reverted for further work; Swob does not claim WebGL rendering today.
Swob goes beyond transcript rendering:
- Execution Tree — reconstructs turns, tool calls, sub-agent spawns, errors, durations, and cumulative token use.
- Context Inspector — breaks context into user, assistant, tool input/output, system injection, thinking, image, and compact slices; marks compact boundaries and pressure warnings.
- Session Audit — a 12-dimension workflow audit covering research/edit balance, thinking evidence, latency, estimated cost, framework overhead, session type, model use, tool efficiency, interruptions, goal length, anti-patterns, and frustration signals.
- Provenance labels — metrics are labeled
reported,estimated, orunavailable; missing evidence is not presented as fact.
| Session Audit | Execution Tree + Context Inspector |
|---|---|
 |
 |
The local dashboard includes token and cost totals, a 365-day heatmap, source/model/project breakdowns, hourly and turn distributions, tool usage, code-change counts, and an audit report.
AI Insights is optional and off until configured. When explicitly invoked, it sends aggregate metrics plus a bounded sample of real user messages to the provider you configure. Read PRIVACY.md before enabling it.
| Source family | Status | Notes |
|---|---|---|
| Claude Code | Native | Transcript, search, usage, live watch, lineage, and terminal resume are available; desktop import is experimental. |
| Codex | Native | Transcript, search, usage, live watch, lineage, terminal resume, and native resume are available. |
| Cursor | Native | Transcript, live watch, and terminal resume are available; search is experimental; usage, lineage, and a native deep link are unavailable. |
| OpenCode | Native | Transcript, usage, archive, and terminal resume are available; search is experimental; live watch, lineage, and a native deep link are unavailable. |
| Zcode | Native | Transcript, usage, and archive are available; search and its workspace-opening deep link are experimental; live watch and terminal resume are unavailable. |
| Antigravity | Native | JSONL, Markdown, and known-schema SQLite transcripts, tools, search, and archive are available. Usage, relationships, subagents, and terminal resume remain experimental; encrypted protobuf is unavailable. |
| Grok / Factory | Native | Composite transcripts, tools, thinking, usage, relationships, search, archive, and compact/rewind boundaries are available. Terminal resume is unavailable pending installed-binary and post-launch verification. |
| Pi | Native | Transcript, tools, thinking, usage, relationships, search, and archive are available; live watch is unavailable and terminal resume remains experimental. |
| Kimi Code | Native | Native and migrated wire transcripts, tools, thinking, usage, relationships, subagents, search, and archive are available. Terminal resume remains experimental pending post-launch verification. |
| Hermes | Native | JSON snapshots and state.db transcripts, tools, thinking, search, and archive are available. Usage, relationships, subagents, and terminal resume remain evidence-scoped and experimental. |
| Qoder | Native | Compound JSONL transcript, tools, relationships, subagents, search, and archive are available. Persisted usage is retained without estimation in canonical v2, but product totals, thinking, format provenance, and terminal resume remain experimental pending first-party/local sampling. |
| Trae | Native (legacy layout) | Legacy plaintext state.vscdb transcript, search, and archive are experimental. Current encrypted ModularData, tools, thinking, usage, relationships, and Resume are unavailable. |
| Source family | Status | Notes |
|---|---|---|
| CC Mirror | Compatible | Claude-compatible transcript, search, and usage are available; live watch and archive are unavailable; terminal resume is experimental. |
Accuracy note: “Native-format adapter” means transcript parsing is implemented, not that every capability is available. Search, usage, lineage, live watch, archive, and resume vary by source. The canonical matrix is
src/shared/provider-capabilities.ts.
Public README claims checked on 2026-07-21. ✅ = explicitly documented; ◐ = adjacent or partial capability; — = not documented as a current feature. This is a capability map, not a quality score.
| Capability | Swob current main |
Claude Code History Viewer | Agent Sessions | SessionView |
|---|---|---|---|---|
| Local multi-harness history | ✅ 12 native + 1 compatible | ✅ 9 providers | ✅ 9+ agents | ✅ 9 tools |
| Visual session lineage graph | ✅ verified + grouping edges | ◐ Session Board, not lineage | — | ◐ child-session normalization, no lineage graph documented |
| Compact-history recovery | ✅ Claude Code | — | — | — |
| Execution tree / agent-call anatomy | ✅ | ◐ tool rendering | ◐ tool/output navigation | ◐ tool-call mix and child sessions |
| Context pressure inspector | ✅ per-turn categories + compact boundaries | ◐ token analytics | ◐ quota/session runway | ✅ session context/cache analytics |
| Provenance-aware health audit | ✅ | — | ◐ honest quota states | — |
| Local full-text search | ✅ SQLite FTS5 for parsed sources; source status varies | ✅ | ✅ local index | ✅ SQLite FTS5 |
| Resume in source CLI | ✅ where supported | ◐ open/focus by session | ✅ supported CLIs | ✅ where supported |
| Headless / browser mode | — | ✅ | — | ✅ |
GitHub Releases is the truth source for the current version, supported architectures, signatures, and immutable asset names. This README deliberately does not synthesize or permanently fall back to an installer URL.
As of 2026-07-24, the public baseline is the Developer ID signed and notarized v1.3.1 release for Apple Silicon and Intel Mac.
System requirement: macOS on Apple Silicon or Intel. Starting with v1.4.0, Releases are also expected to include an unsigned Windows x64 Beta installer named swob-<version>-windows-beta-x64.exe; Windows ARM64, auto-update, CLI/Skill auto-install, WSL, and sources other than Claude Code/Codex are not supported in this Beta. See the Windows Beta guide before installing. The current public v1.3.1 release remains macOS-only.
Important
One manual migration is required for v1.2.0 users. Because v1.2.0 predates the Developer ID trust root, it cannot safely auto-update into the signed release line. Download the matching v1.3.1 DMG from this repository and overwrite Swob once. Existing v1.3.0 users can update through the isolated, E2E-gated signed channel.
git clone https://github.com/IvyYang1999/swob.git
cd swob
npm ci
npm run devTo build local DMGs:
npm run build:macswob search "auth regression" # local cross-session search
swob list --source codex # filter by source
swob where <session-id> # locate package/transcript/backup/source
swob transcript status <id> # inspect four mtimes, lag, and blockers
swob transcript rebuild <id> # rebuild exactly one session package
swob doctor library # inspect writer health and stale count
swob resume <session-id> # print the source-aware resume command
swob insights # aggregate local usage
swob active # inspect running sessions
swob install # install CLI + Agent SkillCLI commands return JSON so other agents can query Swob without scraping the UI. See the CLI location and diagnostics contract for stable errors, privacy boundaries, and recovery behavior.
- Core browsing, indexing, lineage, audit, and Quick Report computation run locally.
- Swob does not include product analytics or session-upload telemetry in the reviewed current source.
- Startup update checks contact the release service but do not send session content.
- Optional AI Insights sends bounded real session samples only after an explicit confirmation.
- API credentials are stored in the local Swob Library configuration and are not encrypted by Swob in current
main. - SSH resume and terminal resume contact only destinations/actions the user explicitly configures.
Read the complete boundaries in PRIVACY.md. Report vulnerabilities through SECURITY.md, never through a public issue containing transcripts or credentials.
| Channel | What it contains |
|---|---|
| Stable v1.3.1 | Developer ID signed and notarized macOS release with trustworthy session detail fallbacks, logical conversation history, reliable group folding, avatar/provider-icon overrides, automatic CLI installation, and the v1.3.0 feature set. |
Current main |
Development source after v1.3.1. It may contain changes not yet represented by a stable installer; GitHub Releases remains the download truth source. |
Electron 40 · React 19 · TypeScript · Zustand · Tailwind CSS 4 · SQLite FTS5 · Recharts · electron-vite
Issues and pull requests are welcome. Contributions require a DCO Signed-off-by trailer; see CONTRIBUTING.md. Before sharing a fixture or screenshot, remove transcript content, absolute paths, credentials, cookies, and device identifiers. Security reports follow SECURITY.md.
Starting with v1.3.0, Swob is licensed under the Apache License 2.0. Releases v1.2.0 and earlier remain AGPL-3.0-only; this change is not retroactive. See NOTICE for copyright and trademark notices.