What's really happening inside your DSH sessions?

[xyrrrrr-r]

What's really happening inside your DSH sessions?

Your context window fills up and you don't know why. Token consumption stays high but you can't tell which message is eating the budget. Sub-agents hand off tasks and you have no idea what was actually said. A tool fails and the original error log is gone. Sound familiar?

I built dsh-eval-control — a session trajectory analysis plugin for DSH. In short: it turns the session black box into a visual dashboard. https://github.com/xyrrrrr-r/dsh-eval-control

What can it do?

See how your context grows

Track context window usage by step, turn, or timeline. Distinguish between summary compression, history trimming, injections, and spikes. Filter messages by token estimate, role, and sorting — spot the expensive ones instantly.

See how your agents collaborate

What instructions did the main agent send to sub-agents? Are sub-agents talking to each other directly? Who picked up that task on the board? The collaboration graph and interaction timeline lay it all out — blue solid lines for messages, orange dashed lines for task assignments. Click any segment to see the original text and evidence.

Tool errors leave evidence

Tool call distribution, failures highlighted in red, raw error text, and rule-based root cause analysis. Click any alert or metric to open the evidence drawer on the right: conclusion, rule, actual computation inputs, evidence grade, and linked raw events — all the way down to sessionId + seq-level persistent events.

Smart alerts, no manual monitoring

Context overload, frequent compression, high error rate, token waste, long intervals — five rules enabled by default with adjustable thresholds. Context overload and frequent compression alerts can also read and apply recommended configuration directly.

Compare sessions, generate diff reports

Pick 2–4 sessions, generate token consumption comparison, context timeline charts, radar charts, and diff reports. Export as PDF / Markdown / HTML, or generate read-only share links with expiration and optional password protection.

How does it compare to other context plugins?

	Other plugins	dsh-eval-control
Context window composition	✅	✅
Message token estimation	✅	✅
Compression event visualization	✅	✅
Distinguish summary compression vs history trimming		✅
Tool call analysis & error evidence		✅
Multi-agent team collaboration graph		✅
Smart alerts + config application		✅
Evidence chain (from conclusion to raw event)		✅
Session comparison + diff reports		✅
Export PDF/MD/HTML + read-only sharing		✅

The short version: what they see, I see too. What they don't see — I see that as well.

Try it

dsh plugin --profile web add dsh-eval-control

Replace web with your actual profile. Reload the profile, open a session, and select Trajectory Analysis at the top.

All data stays local — nothing is uploaded to external services. Analysis is based on event projections; it doesn't intercept model calls or affect performance.

More info

• GitHub repo: https://github.com/xyrrrrr-r/dsh-eval-control
• Panel GIF demos and full documentation in the repo README
• Open an issue if you have questions

---

If you're also building DSH plugins, feel free to reach out. I've written separate docs for the aeval evaluation control integration — see docs/aeval-control.md if you need it.

_{🇨🇳 中文版见仓库 README：https://github.com/xyrrrrr-r/dsh-eval-control}

[davidfunn]

🚀 This is excellent work! Turning DSH session trajectories from a black box into a transparent, evidence-driven dashboard is a genuinely valuable contribution to the ecosystem! 👏

🔍 I especially appreciate the depth of the implementation: context usage analysis, multi-agent collaboration visualization, tool error diagnostics, smart alerts, and cross-session comparisons. The ability to trace conclusions all the way back to raw events makes this much more than a visualization tool — it provides a practical foundation for debugging, understanding, and optimizing complex agent workflows. 💡

🔒 The emphasis on local data processing and non-intrusive analysis is also a thoughtful engineering decision. It shows careful consideration of privacy, reliability, and real-world usability.

✨ What stands out most is how this project addresses real pain points that developers encounter when working with agent systems. Making these workflows observable and explainable can significantly improve developer productivity and help the community build more reliable agents.

🙏 Thank you for putting in the effort to build and share such a useful tool with the DSH community! This is a meaningful contribution, and I'm looking forward to seeing how it evolves. Keep up the great work! 🎉🔥

[pclin1212]

Great work — this fills a real gap in DSH observability.

I've been frustrated more than once by exactly the problems described here: unexplained token consumption, opaque sub-agent handoffs, and error logs disappearing into the void. Turning the session "black box" into a visual dashboard is exactly the kind of tooling the DSH ecosystem needs.

A few things that stand out to me:

• The evidence chain design is the killer feature. Most dashboards stop at "here's a number"; going all the way down to sessionId + seq -level raw events, with the rule, computation inputs, and evidence grade attached, makes the analysis actually trustworthy rather than just plausible. This is rare even in commercial observability tools.
• Distinguishing summary compression from history trimming is a subtle but important detail — those two behave very differently in practice, and conflating them makes context analysis nearly useless. Getting this right shows real understanding of how the session lifecycle works.
• The multi-agent collaboration graph (message vs. task-assignment edges, click-through to original text) addresses something I've never seen any other context plugin even attempt.
• Smart alerts with one-click config application is a nice touch — it closes the loop from "diagnosis" to "fix" instead of leaving the user to figure out what to change.
• Everything stays local and runs on event projections without intercepting model calls — zero performance impact and no privacy concerns. That's the right architecture choice.
  The comparison table is honest and the GIF demos in the README are unusually thorough — you can tell exactly what you're getting before installing. Session comparison + diff reports with PDF/MD/HTML export and read-only share links makes it practical for team use too.

Tried it with dsh plugin --profile web add dsh-eval-control and the onboarding was smooth. Looking forward to seeing where this goes — thanks for open-sourcing it!

[helloC-A]

This is really impressive! What sets it apart for me is the evidence chain design — most analysis tools just show you charts, but being able to trace every conclusion back to the raw session events is huge for debugging long sessions. And keeping all data local is a big plus. Looking forward to trying the session diff reports on my multi-agent setups!