Every Run, Fully Traceable
DeepSeek Harness (harness.vin) Trajectory records a complete append-only event stream for every session: system prompts, chains of thought, tool calls and results, subagent scheduling, and every context injection. Resume, fork, search, and replay all share the same event stream — keeping agent execution transparent and traceable.
How Trajectory Records a Run
User Input
The agent receives the user request and starts the task. The event stream records the request content and context injections.
LLM Reasoning
The large language model analyzes the request and produces an execution plan. The event stream records assembled model requests and chains of thought.
Tool Calls
The agent invokes tools such as code execution or file operations. The event stream records tool names, inputs, and results.
Subagent Scheduling
The agent may delegate work to subagents. The event stream records every subagent scheduling and context injection.
Final Output
The agent completes the task and returns results. The whole session can be resumed, forked, searched, and replayed.
Core Capabilities
Complete Run Recording
Every session records a complete append-only event stream: system prompts, chains of thought, tool calls and results, subagent scheduling, and every context injection.
Run Replay
Replay agent runs. Resume, fork, search, and replay all operate on the same event stream, making issue localization direct.
Inspect by Source
Every record in the event stream can be inspected by source: model requests, tool calls, subagent scheduling, and context injections — each categorized for quick lookup.
Fork & Resume
Resume or fork any run into a new session: continue with the existing context, or explore a different execution path.
JSONL on Disk
Session logs persist as uncompressed JSONL in the session directory, including assembled model requests and tool calls — ready for offline analysis and archiving.
Single Event Stream
Recording, resuming, forking, searching, and replaying all build on the same event stream, so every view stays consistent.
Trajectory Records
Session logs persist as uncompressed JSONL in the session directory — one line per record in the event stream:
# Session logs are saved in the session directory (uncompressed JSONL)
# Each line holds an assembled model request or tool call
$ cat sessions/example-001/*.jsonl
{"type":"system_prompt","content":"You are a helpful software engineer assistant."}
{"type":"llm_request","model":"deepseek-v4-flash","messages":[...]}
{"type":"tool_call","tool":"bash","input":"pytest -q"}
{"type":"tool_result","tool":"bash","output":"3 passed in 1.2s"}
{"type":"context_injection","source":"plugin","content":"..."}Related Docs
FAQ
What is DeepSeek Harness Trajectory?
Trajectory records a complete append-only event stream for every session: system prompts, chains of thought, tool calls and results, subagent scheduling, and every context injection.
How do I view an agent's run trajectory?
Inspect the event stream in the Web UI's trajectory view, or read the JSONL logs in the session directory directly — they include assembled model requests and tool calls.
Can trajectories be replayed?
Yes. Resume, fork, search, and replay all operate on the same event stream.
Where is trajectory data stored?
Session logs persist as uncompressed JSONL in the session directory (session_root), ready for offline analysis and archiving.