dsh-sessions
Session discovery, creation, message delivery, and slash commands for DeepSeek Harness
This is a DeepSeek Harness (DSH) plugin. Review its GitHub README, installation information, maintenance status, and public security signals here.
No Chinese README provided upstream — showing the repository’s original content.
dsh-sessions
Adds session discovery, bounded message inspection, fresh-session creation, and message delivery to DeepSeek Harness.
1. Release: 0.1.1
This release provides four public agent tools:
session_statusfor recent sessions or one exact session;session_readfor bounded canonical message reads;session_createfor creating a fresh session with an initial prompt;session_sendfor steering or following up an existing session.
It also provides the /sessions command family and supports explicit model
provider, model, and thinking-effort selection during session creation.
2. Install the plugin
Install the published package into a DSH profile:
# With the DSH CLI:
dsh plugin --profile web add dsh-sessions@0.1.1
# Without the `dsh` CLI:
npm install dsh-sessions@0.1.1
The DSH profile installation is required for Harness to load the plugin. After upgrading, restart DSH and create a new agent session so the current tool registry is loaded.
For local development, build the package and install the checkout into the target profile:
pnpm build
dsh plugin --profile web add C:\path\to\dsh-plugins\sessions
3. Quick start
Create a child session, inspect it, read its conversation, and send a message:
session_create({ "prompt": "Reply with exactly READY and nothing else." })
session_status({ "session_id": "<SESSION_ID>" })
session_read({ "session_id": "<SESSION_ID>", "offset": 1, "limit": 20 })
session_send({
"session_id": "<SESSION_ID>",
"message": "Continue with the next step."
})
session_create returns a queued result containing the new session ID. Use
that ID with the other tools.
4. Agent tools and command interface
Agent tools
| Tool | Arguments | Behavior |
|---|---|---|
session_status | session_id?, recent_n? | Lists recent sessions, or returns one exact status row. Defaults to the 50 most recently updated sessions. |
session_read | session_id, offset?, limit? | Reads canonical conversation blocks without resuming or mutating the session. |
session_create | prompt, preset?, model?, cwd? | Creates a fresh session and queues its initial prompt. |
session_send | session_id, message, mode? | Sends to an existing session; mode defaults to steer. |
An explicit creation model has this shape:
{
"provider": "<PROVIDER>",
"model": "<MODEL>",
"reasoningEffort": "<LEVEL>"
}
The adapter validates the effort identifier against the selected model. The
cwd option must be an existing absolute directory.
Slash commands
/sessions status [SESSION_ID] [--recent N]
/sessions read SESSION_ID [--offset N] [--limit N]
/sessions create PROMPT [--preset ID] [--provider PROVIDER --model MODEL] [--effort LEVEL] [--cwd PATH]
/sessions send SESSION_ID MESSAGE [--mode steer|followup]
The /sessions create flags map to the tool's nested model object. A JSON
object with the same shape as session_create is also accepted.
5. Session lifecycle and delivery
session_createonly creates a fresh session and queues its initial prompt; it does not wait for model completion.session_statusis inspection-only and defaults to 50 recent sessions.session_readuses a 1-based message-block offset and a maximum limit of 200. It omits trace-only chunks, token deltas, and lifecycle records.session_senddefaults tosteer, which wakes an idle agent and targets the nearest step of a running agent.session_send({ mode: "followup" })queues a separate next turn.- Cold sessions are resumed only for an explicit
session_send; status and read never resume them.
The returned send message_id confirms accepted inbox work. It does not mean
that the target agent has finished processing the message.
6. E2E testing
Reusable prompt fixtures are in
test/e2e/prompts. The recommended flow is create,
capture the returned session_id, and substitute it into the status, read, and
send prompts.
The exact registered names are required. session_send must not be replaced by
the built-in send_message, which targets subagents and has different
semantics.
7. Verify locally
pnpm typecheck
pnpm test -- --runInBand
pnpm build
pnpm pack --dry-run
The published package includes the generated lib directory, the plugin patch,
the README, the implementation specification, and the reusable E2E prompt
fixtures. Unit-test sources and development dependencies are not included.
8. Package scope
The plugin uses public DeepSeek Harness and Cordis APIs and does not modify
DeepSeek Harness source code. dsh-loop owns recurring self-prompts for the
current session; dsh-sessions owns cross-session inspection, creation, and
message delivery.
9. Documentation
Security & install evidence
This score is based solely on public repository metadata and the install evidence registered here — it is not a code security audit.
From a public plugin catalog, linked to a public GitHub repository.
The repository declares the MIT license.
Code updates within the last 180 days.
Exact npm version, integrity hash, and bundle checks are registered.
The inspected package metadata declares no install lifecycle scripts.