Trace Compare & Live Maze — see how your agent actually explored (main path, detours, backtracks, subagent branches)

[lamost423]

Every agent session is a maze run: the model commits to a path, hits dead ends, backtracks, retries. The transcript shows you what it said — not where it wandered. I built a plugin that draws the wandering.

Two surfaces:

• Trace Compare — upload 1 session log for a single-run maze, or 2 logs for a same-axis comparison (e.g. flash vs pro on the same task): per-turn alignment lines, manual anchors, and a per-turn detour inventory that tallies which run wasted more time where.
• Live Maze — a per-session tab where the same maze grows in real time; since v0.4.0, subagent child sessions fold in as branch nodes on the parent's timeline (spawn point, the child's own tool calls as sub-bars, live growth while running).

Honesty rules I care about: verdicts are deterministic and explainable (no LLM judging) — error flags → failure signatures (head/tail windows only, so error text quoted in a git log doesn't count as a failure) → per-tool-class rules, plus behavioral blind-retry detection borrowed from AgentLens-style waste analysis. Every verdict shows its rationale on hover. Tokens are read from assistant/message usage, not stream-chunk counts. Idle gaps fold; everything else stays wall-clock.

New in v0.5.0: the whole UI is bilingual (en/中文) and live-follows your dsh language setting — verdicts are stored as language-neutral structured keys, so switching languages re-renders instantly without re-parsing.

Already listed in awesome-deepseek-harness (Visualization) and Awesome DSH Plugins.

Install (rc.6–rc.8 verified, tracked per rc):

dsh plugin --profile web add dsh-trace-compare

Repo (docs in EN/中文): https://github.com/lamost423/dsh-trace-compare

Feedback very welcome — especially session logs where the verdict rules get something wrong; the thresholds live in one VERDICT_RULES constant and I calibrate against real corpora.

一句话中文版：把智能体真实的探索过程画出来——主干、失败支路、折返点、子代理支路，全部落在同一根墙钟时间轴上；支持双会话同轴对比与会话内实时生长。仓库 README 即中文文档（English

[ylwl1997]

Drawing the maze from the session log (including subagent branches) without a second LLM judge is the useful part — transcripts hide the backtracks. Compare-two-logs on the same axis is a real debugging loop, not a screenshot gallery.

dsh plugin --profile web add dsh-trace-compare

(#3582 is the duplicate; this thread is the one to keep.)

If you want it in the community catalog: https://github.com/ylwl1997/dshbase/issues/new?template=plugin-submission.yml
https://dshbase.com/plugins/directory/

[lamost423]

Thanks — that's exactly the design bet: the session log already contains the ground truth of where the agent wandered, so a second LLM judge would only add cost and unexplainable verdicts on top of it. Every rule is in one VERDICT_RULES constant, calibrated against real corpora (871 steps so far) — if the maze ever flags something wrong, a session log in an issue is the most useful thing you can send me.

Submitted to the dshbase directory as suggested: ylwl1997/dshbase#67 🙏

[lamost423]

Thanks — you nailed the two design bets: verdicts are deterministic and drawn from the session log alone (no second model, so every ✗ is reproducible and explainable on hover), and same-axis compare is exactly how I debug (two providers, same task). And thanks for the quick verification over at dshbase — impressively fast pipeline. Appreciate the dedupe note on #3582 too.