dsh-plugin.md

InferRSI plugin for DeepSeek Harness

Use InferRSI in a DeepSeek Harness (DSH) session to research an inference bottleneck, prepare a controlled comparison, or inspect an existing run. The bundled CLI also serves InferRSI's live optimization dashboard for recorded runs and campaigns. The manual-install preview implements the DSH slice of RFC #152.

Build and install

From the InferRSI source checkout, build the native bundle with Python 3.10+.

python -m tools.build_dsh_plugin

The command creates dist/inferrsi-dsh.tgz and prints its SHA-256. Use --output <new-path>.tgz to build again. The archive includes the entry skill, launcher, Python runtime, knowledge, and license. Keep the printed checksum with the archive when moving it to another machine. The archive includes the pure JavaScript SSH transport with its license notices, Python runtime and knowledge resources.

On the DSH machine, select the profile you use and install the archive. The examples use web. Use an absolute archive path so it resolves from the profile directory.

dsh --version
dsh plugin --profile web add "<absolute-path>/inferrsi-dsh.tgz"
dsh --profile web

This preview targets DSH 0.2.1, starting at 0.2.1-rc.2. Initial Windows checks with DSH 0.2.1-rc.5 covered native installation, skill discovery, and removal. Put pnpm on PATH for plugin management. For DSH and Node.js setup, see the DeepSeek Harness bundle guide.

Python 3.10+ must be available to the session's terminal. Core InferRSI operations run from the bundled runtime. The optional skills CLI and explicit checkout-skill sources require PyYAML. Git and the prepared target environment are needed for comparisons.

Start an iteration from the form

In DSH Web, open a target repository and its conversation, then click InferRSI beside the message composer. The form has three sections.

1. Optimization target (优化对象). Enter the repository, model, framework, SIAS plane and objective. The form fills the local repository path from the current session when available.
2. Test environment (测试环境). Choose GPU or NPU and local or remote execution. After selecting remote execution, enter the remote repository path in the first section, then return here to test the SSH connection. Set the maximum optimization rounds and total time budget in minutes. The prepared device, benchmark and quality scripts, and campaign directory are optional.
3. Acceptance criteria (验收标准). Enter the agreed metrics, thresholds and constraints, and provide an acceptance issue or plan and baseline revision when available. The Agent checks missing inputs and freezes the gates before candidate experiments.

Start iteration sends one formatted prompt to the Agent in the current conversation. The optional prompt preview shows the complete instructions and structured task parameters before submission. Submission is disabled while the Agent is running or the request is being sent. A successful submission appears in the conversation; subsequent execution follows the selected Agent, its model connection and existing permissions. After confirmed remote delivery, the connection belongs to that task; the form blocks resubmission and connection changes even before the Agent's running indicator refreshes. Explicitly disconnect before reconfiguring, which can stop the accepted remote task. A local rejection before sending keeps the form available for correction; abandoning that rejected submission requests closure of its unused remote connection. An unconfirmed close retains its receipt for cleanup or retry.

If delivery has started but its result is uncertain, inspect the conversation before doing anything else. The form keeps the remote connection and blocks a second submission or connection-changing edits. Keep the connection if the task may have been accepted, or explicitly disconnect it when choosing to abandon the remote work. Disconnecting can stop active remote execution. There is no automatic resend or disconnection after an ambiguous send; a retained connection is released by session/plugin disposal if no explicit close is requested.

The request directs the Agent to use InferRSI, freeze the required gates before candidate experiments, stop at the submitted round/time budgets, record campaign history, and open the existing live dashboard. Fields stay in browser memory until submitted through the normal DSH message path. Reloading or switching sessions closes the unsent form. The form is a Web client contribution; the native entry skill remains available in other profiles.

Import a task from YAML

Open 给 Agent 的 YAML 模板 in the form and click 复制 YAML 模板. Give the template to the Agent and ask it to return a .yaml file with the task values. Choose the file under 导入 YAML, review the three sections and prompt preview, then click Start iteration.

The file uses schema_version: 1 and sections named optimization, environment and acceptance. Use acceptance.reference for an acceptance issue or plan and acceptance.constraints for metrics and thresholds. YAML block strings (|) support multiline objectives and criteria. Set maxRounds and maxMinutes as positive integers. Fields can be left null or omitted while drafting. Import replaces task fields, and ordinary form validation applies when starting. An omitted local repository defaults to the current workspace.

Use one YAML 1.2 document up to 64 KiB. The import accepts the template fields and rejects unknown keys, duplicate keys, aliases and tags. Invalid files leave the form intact. Enter SSH connection settings and credentials in the form. Changing the execution location, repository or scripts through import follows the same connection-close and retest flow as editing those fields directly.

Run on a remote machine

Select Remote machine, then provide the remote repository's absolute Linux path, IP address, SSH port, username and authentication method. Upload an SSH private key and enter its passphrase when needed, or enter the account password. Use the local DSH page or an HTTPS deployment when providing credentials.

Click Test connection. For a new server, compare the displayed SHA-256 host key fingerprint against a trusted value, then choose Trust this fingerprint and test. A supplied fingerprint must match exactly. Successful testing checks SSH authentication, Linux command execution and access to the selected repository. GPU/NPU availability and benchmark readiness are checked by the Agent afterward.

After a successful test, Start iteration becomes available. Changing the host, authentication, repository, designated scripts or execution location first closes the tested connection. The original fields and connection receipt remain until the Host confirms closure. While closing, submission, retesting and further edits are blocked. If closure fails or its result is uncertain, use Retry disconnect; the original connection is retained, and the requested field change is not applied. After successful retry, edit the fields again and test the new connection. A same-session receipt already confirmed closed by the Host is acknowledged separately; unknown or other-session IDs remain errors.

If the Host explicitly reports no record of a queued ID, for example after a Host reload, the form also offers Forget this local receipt. This requires a separate confirmation and only discards the selected local cleanup record. It does not confirm that SSH or a remote task has stopped; check the remote target separately before forgetting it. Other-session IDs, network or permission failures and unconfirmed responses do not enable this action. Remaining cleanup records still block the form, and using a new remote connection always requires another connection test.

Unconfirmed close requests retain only opaque connection IDs in session-scoped client memory. Leaving and reopening the form can resume this cleanup, including after an unmount retry fails. Pending or accepted task connections are not included unless the user has explicitly requested disconnection. Passwords, uploaded keys and key passphrases are cleared from the form after authentication. An authenticated SSH client remains in Host memory for that Agent session. DSH shutdown, plugin removal, session disposal or an explicit connection close releases it. A lost connection needs a new test.

The prompt contains the hardware class, execution location, task inputs and an opaque connection ID. Credentials, IP address and SSH username use the authenticated owner-only connection endpoint and stay outside the prompt, dashboard and experiment records. Connection testing does not submit an Agent task. Connection failures return fixed diagnostic categories.

The Agent uses inferrsi_remote through DSH's normal tool execution pipeline. Its operations are repository file listing/reading/writing, the designated benchmark or quality script, canonical InferRSI CLI operations, a live dashboard tunnel, and connection close. Paths are checked against the selected remote repository. These fixed operation names are not an execution sandbox: write can change repository code, including designated scripts or Git hooks, and run_script / cli run can execute that edited code with the SSH user's permissions. Treat the repository, selected scripts and edits as trusted parts of the authorized task. There is no arbitrary shell-command operation, but path containment does not isolate the effects of code that is executed. Remote Python 3.10+, Git, the selected engine, accelerator runtime and model resources must be prepared separately.

For run_script, supply maxMinutes from the task's remaining time budget. Long CLI operations can use the same explicit limit; cli run also retains its own --max-minutes bound and uses the smaller limit when both are supplied. Do not reset the total task budget for each operation.

Canonical CLI operations unpack the bundled runtime into a private remote temporary directory. Dashboard mode keeps that runtime alive while the view is open and forwards the remote loopback listener to an OS-assigned local loopback port on the DSH Host. Use the returned local URL when the browser and DSH Host run on the same machine. For a remotely hosted DSH Web deployment, also forward that Host loopback port to the browser machine before opening it. Stopping or timing out an execution operation requests cleanup of that command's SSH channel while preserving the verified connection and any separate live dashboard. Explicit connection close, session disposal or transport failure still releases the whole connection. After interruption, inspect recorded state before a new run; no automatic retry follows a timeout.

Integration design

flowchart LR
    F["InferRSI form"] --> P["Structured task prompt"]
    P --> A["Current DSH Agent"]
    F --> C["Owner-only SSH connection test"]
    C --> R["Session-scoped connection ID"]
    A --> L["Bundled local CLI"]
    A --> R
    R --> E["Remote repository and bundled CLI"]
    L --> D["Existing dashboard"]
    E --> D

Loading

Owner	Responsibility
Form and prompt builder	Collect GPU/NPU, execution location, goal, SIAS plane, user budgets and optional acceptance inputs
DSH conversation	Deliver the prompt once to the selected session, preserving its model and permissions
SSH connection bridge	Authenticate the owner request, verify host key and login, retain a session-bound connection and return an opaque ID
inferrsi_remote	Carry explicit repository operations and cancellation through DSH tools without including authentication material
Canonical InferRSI runtime	Research, freeze, run, evaluate and record campaign rounds using the existing contracts
Existing dashboard	Display recorded progress, candidate history, results and evidence through local or forwarded loopback access

The connection route uses DSH's authenticated owner transport. DSH rejects non-owner peers before calling the plugin handler. The request's sessionId selects the owner's target session; it is not an authentication credential. DSH uses that session ID as its Agent ID, and remote tool calls obtain it from exec.agent.id rather than from model arguments. The manager rejects a receipt used from another session. This is one DSH owner's session partitioning, not isolation between mutually untrusted users of one SSH account.

User-supplied round and time budgets are part of the task prompt. Runtime command limits and recorded experiment verdicts retain their existing meaning. A form submission, successful SSH test or completed process does not accept a campaign.

Use in a target workspace

Open the repository you want to investigate in a new session and invoke /inferrsi or ask the agent to use InferRSI. For example, ask it to find relevant knowledge for your engine's memory bottleneck and preview a comparison. The skill tells the agent how to use its installed launcher while keeping the target repository as the working directory. From the installed skill directory, the launcher is scripts/inferrsi.py.

• Use --plugin-info to check the selected interpreter, bundled runtime version, and content SHA-256. An unbuilt source checkout can report 0+unknown as the runtime version.
• Use collect to find relevant packaged knowledge and read to inspect a selected revision and its resources.
• Use plan to preview a comparison. Running a comparison requires the accepted scope, frozen gates, and prepared target environment.
• Use status and report to inspect an existing run and retain its recorded verdict. report writes the HTML report into that run directory.
• Use dashboard to observe recorded work through InferRSI's existing live dashboard, following the workflow below.

If the skill is missing, verify that you opened the profile where you installed the bundle and that the profile provides the DSH skills service and skill tool. If Python is missing or incompatible, select a compatible interpreter in the same terminal environment. If the launcher reports an incomplete bundle or version mismatch, rebuild or reinstall the .tgz.

Open the live dashboard

In the DSH session, ask InferRSI to open the dashboard for the intended run or campaign. For a tested remote connection, use inferrsi_remote with operation dashboard, its connection ID and the remote campaign/run directory in path. The bridge starts the existing dashboard and returns a forwarded Host loopback URL; do not launch a second server through a local terminal.

For local execution, resolve <launcher> to scripts/inferrsi.py under the installed skill's resource directory and use the same Python interpreter selected for the other InferRSI commands:

python -X utf8 "<launcher>" dashboard --input "<CAMPAIGN_OR_RUN_DIRECTORY>" --live --port 0

The local command prints a JSON object whose dashboard field contains the selected loopback URL. Open that URL in an external browser and share it as a clickable link in the DSH session. Port 0 lets the operating system choose an available port. Keep the server running through DSH's terminal/process tools and retain the process identity so it can be stopped through that same process when the view is finished.

Choose a directory containing campaign.json to show its recorded rounds, candidates, optimization history, and linked runs. A directory containing run.lock.json shows that individual run. The page checks for changes about every three seconds and keeps the current selections where possible. The header shows the last successful check and refresh failures. Recorded verdicts, missing evidence, and partial results remain visible in the dashboard.

Populate the campaign as the work happens. Use campaign capture-context for the knowledge retrievals being prepared, campaign record to associate runs with candidates and rounds, and campaign log for the reviewed goal, search space, round hypotheses, decisions, and success/failure summaries. Follow the campaign integration guide for the exact fields and commands. These caller-supplied records are the dashboard's source for the optimization story. Record missing context or reasoning as missing when it was not captured during the work.

Run the dashboard server on the machine holding the evidence. It binds to loopback. The remote tool already forwards its port to the DSH Host; if the browser runs on another machine, forward that Host loopback port to the browser machine. For other remote execution mechanisms, arrange authorized forwarding before opening the URL. Follow the live dashboard guide for port forwarding, access boundaries, and shutdown. The same guide documents offline exports and reviewed translation overlays.

Disable, update, or remove

In Web's Plugin Manager, disable the bundle and restart the profile if prompted. To remove it, run the following and reopen the profile.

dsh plugin --profile web remove inferrsi-dsh

To replace a manual preview, remove the old package and install the new .tgz in the same profile. Run --plugin-info again to identify the selected runtime. Experiment records stay in their original output directories.