AI Agent Hub
Back to plugins
🧠

dsh-trace-repeat

Memory Updated 2026.08.16

Run the following command in DeepSeek Harness:

dsh plugin install p2coder/dsh-trace-repeat

Paste the following prompt into your AI chat to install this plugin:

Install the plugin by running dsh plugin install p2coder/dsh-trace-repeat in the DeepSeek Harness terminal; the source is available at https://github.com/p2coder/dsh-trace-repeat .

About this plugin

In DSH Web, a session streams forward and is hard to walk back once it diverges. dsh-trace-repeat pins every completed reasoning turn and every finished tool call into an immutable version, carrying full provider, model, reasoning-effort, token-usage, and input/output metadata so you can always see exactly what happened.

The plugin centres on a record-play-resume loop. Atomic file writes (tmp plus rename) keep the trace index and version files crash-safe; each version is materialised as a git commit with a dedicated worktree for isolated, reproducible state. The /trace command family lets you browse the version timeline, inspect full details, and resume execution from any balanced version (a reasoning-completed point with no dangling tool calls). A new session is spawned in the corresponding worktree, inherits a balanced context, and continues to be traced. Pause and resume are tightly integrated with task-control: a safe pause waits until in-flight tool results land, and resume marks the version and continues recording seamlessly.

Ideal for engineers running long reasoning chains or debugging tool-call sequences in DSH Web, especially when you need to reproduce a runaway inference path, compare inputs and outputs across versions, or pick up execution from an intermediate state without starting over.

Use Cases

  • Walk back through a divergent reasoning chain to locate the problem version
  • Compare model inputs and outputs across different versions
  • Resume execution from a balanced midpoint without restarting

Best For

  • Engineers debugging long reasoning chains in DSH Web
  • Developers who need to reproduce tool-call execution paths
  • AI application leads managing version comparison and execution recovery