Updated 2026-11-13 · Cameron Voss and West Mao
Super Kanban Turns DeepSeek Harness Sessions Into a Task Board (Nov 2026)
The board is not filled in by hand. It is a projection of what the agents already wrote down, which is both why it works and where it can mislead you.
Why do multi-agent sessions need a board?
One agent is easy to follow. Five are not. The session list shows you that work is happening, not what the work is.
The familiar failure looks like this. You start a frontend agent, a backend agent and a test agent. Two hours later you cannot tell which one is waiting on a migration, which one finished but left a failing test, and which one stopped because a command needed approval. The information exists, but it is spread across three transcripts and you have to read them to find it.
Task trackers do not fix this either, because they assume a human is updating tickets. In a coding agent session, the human is the least informed participant. The agent knows what it tried. The tracker knows what someone typed into it last Tuesday.
Super Kanban inverts that. It treats the transcript as the source of truth and generates the tracking view from it.
What does Super Kanban read, and what does it write?
It reads DeepSeek Harness session records inside the scan scope you configure. For each session it extracts a task goal, the current execution state, unresolved questions, and what was delivered.
It writes two things: the board view, and a resume draft when you ask it to continue an unfinished task. The board is derived data. Delete the scan scope and the board is empty, which is worth knowing before you build a habit around it.
That asymmetry is the design. A plugin that edits your project to maintain a board would create a second source of truth. A plugin that reads transcripts cannot drift from what the agents actually did.
What are the five capabilities?
The plugin advertises five behaviours. They differ in how much they depend on a model, which matters when you decide whether to trust a card.
| Capability | Input | What you get |
|---|---|---|
| Automatic board generation | Session records in scope | Task cards with goal, state and open questions, without manual entry |
| Multi-agent unification | Sessions linked to one task | One card that shows progress across frontend, backend and test agents |
| Blocker detection | Approvals, dependencies, acceptance items | A lane of work that needs a human decision |
| Resume drafts | Unfinished session state | A task draft with remaining items, acceptance conditions and context |
| Dual analysis mode | Model access, or local rules | Analysis when a model is available, basic tracking when it is not |
Automatic board generation is a summarization problem
The quality of every card depends on how well the session states its own goal. A session that begins with a clear request produces a clear card. A session that begins with "fix this" produces a card that says "fix this", because that is all the transcript contains.
The practical fix is not in the plugin. Write the first message of a session the way you would write a ticket title, and the board inherits the discipline.
Multi-agent unification is a linking problem
The plugin associates sessions with a task. How that association is made matters more than the board's visual design. If it links by time window, unrelated sessions started in the same hour will merge. If it links by explicit grouping, you have to do a small amount of setup.
Check this on your own workload before trusting a card that claims three agents are on one task. Open the linked sessions and confirm they belong to the same piece of work.
Blocker detection is where the board earns its keep
Most agent work does not stop because the model is incapable. It stops because a command needs approval, a dependency is missing, or the result needs a human yes or no. Those three conditions are exactly what the plugin scans for.
That gives you a lane that a normal task board does not have: waiting on you. Reviewing that lane once in the morning is a different activity from reading transcripts, and it takes a fraction of the time.
Resume drafts are only as good as the transcript
A resume draft collects what was unfinished, the acceptance conditions, and the context needed to continue. That is more useful than a fresh session that starts from zero, because the agent does not have to rediscover the project state.
It has a hard limit. If the original session trimmed its own context or never wrote down a decision, the draft cannot recover it. Treat the draft as a starting point and read it before you send it.
Dual mode is an availability feature, not a quality feature
Local rules can detect file changes, failed commands and approval prompts without a model. They cannot summarize intent or judge whether a task is genuinely complete.
So when the model is unavailable, expect the board to stay useful for status and blockers and to get vaguer about what the work means. That is the right trade for a tool that should not stop working when an API does.
How do I install it?
Add the project's plugin market index to dsh-market and install from there, or run the CLI command below. Then refresh DeepSeek Harness so the plugin registers, open the task board from the sidebar, and set the scan scope. Start with one project directory, not your entire workspace.
The version number comes from the project's own documentation and is an example. Pin whatever the current release is, and read the compatibility note before upgrading. A plugin that scans session files is coupled to the session format, so a harness upgrade can break it.
Session transcripts can contain secrets: environment variables printed during debugging, tokens in error messages, customer data in fixtures. The scan runs locally, but the board and any generated summary inherit that content. Set the scope deliberately.
dsh plugin --profile desktop add dsh-super-kanban@0.6.1Where can the board mislead you?
Every state on the board is an inference. The useful question is not whether it can be wrong, but which wrong states cause real damage.
| Failure | What it looks like | How to catch it |
|---|---|---|
| Stale cards | The board reflects the last scan, not the last second | Rescan before you trust a card on work you did outside the scope |
| False completion | An agent says it finished and the card moves to done | Keep CI as the authority on done; verify before you accept |
| Merged tasks | Unrelated sessions combine into one busy-looking card | Open the linked sessions and confirm they share a goal |
| Missing blockers | A blocker that never entered the transcript cannot be detected | Add manual notes to the task when a decision happens outside the agent |
| Summary drift | Two different tasks get similar generated titles | Read the card body, not only the title |
What does a working day look like with it?
Start with the blocked lane. Those are the items where a human decision unblocks an agent, so clearing them has the highest return. Then check the in-progress lane for tasks that have not moved since the last scan; those usually need a resume draft rather than more waiting.
Finish with the acceptance lane. This is where the agent believes it is done. Verify with tests or a manual check, then move the card yourself, because the board should never be the thing that declares work complete.
The order matters. If you start by reading what the agents did, you are back to reading transcripts with extra steps.
FAQ
Does Super Kanban change my project files?
It reads session records and generates board state. The resume action produces a task draft. It is not a build tool and does not modify source code as part of scanning.
Can I use it without a model API?
Yes, in local rules mode. Status, file changes and approval prompts still work. Summaries and intent judgments get weaker.
Does it replace Jira or Trello?
No. It covers the slice those tools miss: what the agents are actually doing right now. Team planning, estimates and cross-team reporting still belong in a tracker.
What license is it under?
MIT. You can read the source and fork it, which is worth doing if you need to change how sessions are linked to tasks.
Super Kanban is useful for one specific job: turning parallel agent transcripts into a list of decisions you owe. Use it to find blocked work and to draft resumes, not to declare victory. The board is a projection, and the projection is only as honest as the sessions it reads.