TraceScope is an MIT-licensed plugin for DeepSeek Harness with three capabilities: functional impact analysis, design diff analysis, and crash analysis. Version 0.3.6 adds Local crash capture, a crash source that doesn't wait for crash reporting.
GitHub: https://github.com/rebornace/dsh-tracescope | npm: @rebornace/dsh-tracescope
Crash analysis already works against APM services such as Firebase Crashlytics: it lists issues, opens the details, parses the stack, and writes the AI conclusion back.
Every one of those services only knows about crashes that were reported and aggregated. A crash that never got that far simply isn't in the list.
0.3.6 adds Local crash capture. Crashes and ANRs are pulled straight from a USB-connected device. Collection runs locally inside the plugin, with no APM dependency and no upload to an APM. Once a crash has been captured, it goes through the same analysis, write-back and defect-filing flow as the reported ones.
The sidebar's Data source section has a new Local crash capture entry. Choose Android or iOS, refresh the device, start recording. Crashes and ANRs raised during the recording are written locally and appear in the list; opening the source loads the history automatically, with the count shown next to the title.
Opening an entry shows the parsed stack, plus possible matches for the files in that stack against the configured code directory. A miss is only a soft hint: the full stack still reaches the analysis.
A few things worth knowing:
Captures go to ~/.tracescope/crash-inbox/shared/. TraceScope doesn't upload them to Crashlytics or any other APM. Whether the app's own APM SDK reports anything is decided by the app's configuration and has nothing to do with this capture.
Android capture goes through adb. If adb isn't installed, refreshing the device downloads platform-tools for you; no full SDK setup needed. USB debugging has to be on.
Captures stay on the machine that recorded them, and the list only shows entries in that machine's local directory. iOS capture needs macOS (libimobiledevice); Windows can't record iOS. Android capture works on Windows and macOS alike.
Capture on its own only fetches the crash. With an entry open, analysis offers two options:
| Entry | Output | Notes |
|---|---|---|
| Write reproduce steps (primary) | repro steps, environment info | symptom only, no root-cause guessing |
| Locate & fix | root cause, blast radius, minimal fix | meant for changing code |
Both quietly tidy the stack and drop the prompt into the session; the conclusion shows up in the sidebar once the agent calls the write-back tool. From there it can be copied, exported as Markdown, or filed as a defect on GitHub Issues, GitLab Issues, or a generic webhook.
DeepSeek Harness desktop: install @rebornace/dsh-tracescope from the plugin page, open the TraceScope sidebar, pick Local crash capture, connect a device and start recording.
MCP clients (Cursor, Claude Code, Codex, or any MCP-capable agent): add this to the config:
{
"mcpServers": {
"tracescope": {
"command": "npx",
"args": ["-y", "@rebornace/tracescope-mcp"]
}
}
}
Restart, then call tracescope_open_panel to get the same three-mode workbench the sidebar has. Local capture needs a connected device, so run it from the desktop sidebar or that workbench. Node.js 20 or newer.
All three backends sit behind extension points rather than being tied to one vendor:
| Surface | Available now | Can extend to (examples) |
|---|---|---|
| Crash sources | Local crash capture (USB), Firebase Crashlytics | Sentry, other APM services |
| Tracker / issue platforms | GitHub Issues, GitLab Issues, generic webhook | Jira, other issue trackers |
| Design sources | Figma | Other design tools on request |
New crash sources will be added over time; trackers and design sources come in on request. Opening an issue with the platform name, its auth method, whether it has a public API, and the use case is the fastest way to get one prioritized. Roadmap details live in the repo docs.
Crash collection now has two kinds of sources: online from APM services such as Firebase Crashlytics, and on-device over USB. Both feed the same parse / write-back / file-a-defect flow. What the plugin does hasn't changed: it automates the fetching, parsing and locating, and lowers the barrier to applying AI analysis to app quality.
Issues and PRs are welcome.
Links