Observability & Trajectory

Goal

Use DeepSeek Harness session logs so you can explain *what the model saw* and reproduce failures.

Append-only session logs

Official design notes emphasize that prompts, chain-of-thought (when present), tool calls and results, sub-agent scheduling, and context injections are written to an append-only stream. That stream is the source of truth for debugging.

Trajectory view

In the Trajectory UI you can inspect events by source, then resume, fork, retrieve, or replay from the same event stream. Treat it like a flight recorder for agent runs.

Practical audit checklist

• Did the model receive the files you intended?
• Were tool arguments within the workspace?
• Did a plugin inject unexpected context?
• Can you replay the failing step after a config change?

Why it matters for teams

Observability turns “the agent went weird” into a reviewable artifact — essential for security reviews and for teaching juniors how harnesses differ from chatbots.

Wrap-up of the tutorial path

You now have: concept → install → first task → architecture → modes → surfaces → plugins → observability. Return to Install when onboarding someone new, or open FAQ for alias and official-status questions.

FAQ

• Chat export vs Trajectory? Why append-only logs matter: From session logs to Trajectory.
• UI up but tools look wrong? Troubleshoot and the npx / 3080 ladder.

Keep verifying upstream

Developer preview APIs evolve. Pin versions in production experiments and re-read the GitHub README when something breaks after an upgrade.