How to use this reference
- Use the page search in your browser to find a term quickly.
- Read the professional definition when you need accuracy.
- Read the everyday explanation when the definition still feels abstract.
- Return to the linked guide before changing permissions, installing an extension, or publishing a package.
Model, Agent, and Harness
| Concept | Professional explanation | Everyday example |
|---|---|---|
| AI model | A statistical system that interprets input and generates text, code, or other output. It does not receive computer access by itself. | Everyday explanation: the model is the person reasoning at a desk. The desk, tools, keys, and records come from other parts of the system. |
| Agent | A running role that combines a model with instructions, context, tools, and a loop that can take several steps toward a goal. | Everyday explanation: instead of answering once, the worker can inspect, act, check the result, and continue until the assigned task is done. |
| Agent Harness | Software that coordinates the model, tools, context, permissions, sessions, and interface used for Agent work. | Everyday explanation: it is the managed workshop around the worker. It decides which tools are present, which doors are locked, and what is written in the logbook. |
| Provider | A service or adapter that makes a model available to the Harness and handles its request format and authentication method. | Everyday explanation: different electricity suppliers can power the same workshop, but each has its own account and connection rules. |
| Tool | A named capability the Agent can request, such as reading a file, editing text, running a command, or searching a source. | Everyday explanation: a tool is an item in the workshop cabinet. Having it listed does not always mean it may be used without approval. |
| Context | The information made available for the current model request, including instructions, selected history, files, and tool results. | Everyday explanation: it is the set of documents open on the desk for this decision. It is not automatically a permanent memory of everything that ever happened. |
How work is organized
| Concept | Professional explanation | Everyday example |
|---|---|---|
| Workspace | The project location used as a working directory and an organizational reference for related work. Selecting a workspace is not itself a security boundary. | Everyday explanation: it is the address on the work order. Locks and access rules still come from the configured executor and permissions. |
| Session | A durable record of messages, model output, tool activity, and state for one continuing line of work. | Everyday explanation: it is the project logbook that lets you return later and understand what was discussed and done. |
| Turn | The complete response cycle triggered by one admitted user input, possibly containing several model and tool steps. | Everyday explanation: one request at the service desk may require several internal actions before the final answer is ready. |
| Step | One model request and the tool calls that follow from it inside a turn. | Everyday explanation: it is one item on the worker's checklist, while the whole turn is the completed request. |
| Plugin | A loadable software unit that contributes services, events, tools, interface parts, or other behavior to the Harness. | Everyday explanation: it is an approved module added to the workshop, such as a new machine or control panel. |
| Profile | A named composition that determines which bundles, plugins, and configuration layers start together. | Everyday explanation: it is a saved workshop setup for a particular kind of job. DeepSeek Harness ships web and headless templates. |
| Skill | A reusable set of instructions and supporting resources that teaches an Agent how to carry out a bounded workflow. | Everyday explanation: it is a reviewed procedure card. It guides the worker, but it does not silently grant new computer permissions. |
| Runtime | The actual software process, dependency versions, configuration, and operating environment in which a task runs. | Everyday explanation: a recipe can be identical while the real kitchen differs. Verification must happen in the kitchen that will actually be used. |
| Diff | A structured comparison that shows how files or text differ between two states without applying either state by itself. | Everyday explanation: it is a before-and-after sheet with changed lines highlighted. Reading it does not accept the changes. |
| Hook | A configured trigger that runs a defined action before, after, or during a named event in a tool or workflow. | Everyday explanation: it is a doorbell wired to a specific response. Because the response can run code, review its source and permissions. |
| MCP | Model Context Protocol, a standard way for an Agent or Harness to connect to external tools and data through an MCP server. | Everyday explanation: it is a standard connector between the workbench and another service. A matching plug does not make the service trusted. |
| Cloud | A remote execution environment hosted outside the local computer, with its own files, credentials, network, persistence, and billing boundaries. | Everyday explanation: it is a rented workspace in another building. Check its keys, storage, costs, and cleanup rules before sending work there. |
Permissions and safety
| Concept | Professional explanation | Everyday example |
|---|---|---|
| Permission | A rule describing which action is allowed, denied, or subject to a decision at runtime. | Everyday explanation: it is the access list on a door. A sentence asking someone to be careful is not an access rule. |
| Approval | A human or policy decision requested before a particular action is allowed to continue. | Everyday explanation: the worker pauses at a locked cabinet and asks before taking the key. |
| Sandbox | A technical boundary that restricts selected effects of a process. Its exact coverage depends on the platform and configuration. | Everyday explanation: it is a fenced work area. You must still check whether the fence covers files, network access, processes, or only part of them. |
| Credential | A secret or identity token used to authenticate with a model provider, service, repository, or other protected system. | Everyday explanation: it is a key or bank card. Store it in the designated settings or secret store, never in chat text or project files. |
| Least privilege | The practice of granting only the access needed for the current task and only for as long as it is needed. | Everyday explanation: give a visitor the key to the meeting room, not a master key for the whole building. |
| Rollback | A prepared operation that restores the previously verified state after a change fails or is no longer wanted. | Everyday explanation: keep the previous working version and clear return instructions before replacing anything. |
Appearance and interface
| Concept | Professional explanation | Everyday example |
|---|---|---|
| Semantic token | A named design value whose name describes a purpose, such as background or accent, instead of a fixed color. | Everyday explanation: it is a label such as wall paint or warning light. Each theme can supply a different color for the same job. |
| Theme | A coordinated set of semantic appearance values, usually covering light and dark modes without changing the product structure. | Everyday explanation: it changes the room's paint and lighting plan while leaving the walls and furniture in place. |
| Full Skin | A DSH Themes publication class that combines semantic values with controlled imagery and generated styling under stricter evidence rules. | Everyday explanation: it changes paint, lighting, and selected decoration. It is a DSH Themes format, not an official native DeepSeek Harness format. |
| Plugin | A Plugin that changes or adds a Harness workflow rather than only supplying appearance values. | Everyday explanation: it adds a new control panel or working station, so its behavior and permissions matter as much as its appearance. |
| Transport | The connection that carries requests, events, and state between the Web interface and the running Harness process. | Everyday explanation: it is the delivery route between the control room and the workshop. A visible page does not prove that the route is connected. |
Packages and verification
| Concept | Professional explanation | Everyday example |
|---|---|---|
| Package | A versioned distribution unit containing code, metadata, resources, or configuration intended to be installed together. | Everyday explanation: it is a sealed product box. The name on the shelf is not enough to prove what is inside this exact box. |
| Artifact | The exact built file that is downloaded, installed, tested, or published, such as a specific package archive. | Everyday explanation: it is the physical box that arrived, not the product description in the catalog. |
| SHA-256 | A cryptographic digest used as a stable fingerprint for the complete bytes of an artifact. | Everyday explanation: matching fingerprints show that two files have the same bytes. They do not prove that the contents are safe or properly licensed. |
| Sidecar | A separate machine-readable record stored beside an artifact to describe its identity, compatibility, hashes, or evidence references. | Everyday explanation: it is the inspection card attached to the box. Both the card and the box must match. |
| Attestation | A signed or controlled record stating what was checked, in which environment, against which exact inputs, and with what result. | Everyday explanation: it is a dated inspection report. A report for another model or version cannot certify the item in front of you. |
| Certified baseline | The exact Harness version and evidence set against which an item has completed the required verification workflow. | Everyday explanation: current upstream software may be newer than the version that has finished inspection. Newer and certified are different facts. |
What to do when a term is still unclear
- 01
Find the exact term
Copy the word as written, including its product name when one is present.
- 02
Identify the layer
Ask whether it describes intelligence, organization, permissions, interface, or verification.
- 03
Check the product-specific meaning
The same word can have a narrower meaning in DeepSeek Harness, Pi, Codex, or another tool.
- 04
Pause before a risky action
If the term affects access, credentials, installation, or deletion, read the linked guide or official source before continuing.
Continue from the concept you found
Practice: follow one file from request to verified change
Create a small inventory exercise to connect the glossary to visible behavior. You will read a file, check a calculation, change one value, and inspect what persists. Use an empty practice folder and the already configured Harness Web workspace and model from the quick-start guide.
text
item,quantity,unit_price
tea,3,12
mug,2,25
bag,1,18- 01
Select the workspace
Choose the folder that contains inventory.csv. The workspace supplies the working directory; selecting it alone does not enforce a filesystem or network sandbox.
- 02
Check the running setup
Confirm which model and permission preset this session uses. A profile selects plugins and configuration; it is different from a workspace folder or a saved conversation.
- 03
Send the reading task
Copy the prompt below. Expand the actual file-reading call and its result, then compare the returned rows with your file. A fluent final answer by itself does not prove that the file was read.
text
Read inventory.csv in the selected workspace. For each item show quantity × unit_price and the subtotal, then show the total. State the file path you actually read. Keep the file unchanged. If you cannot access it, explain the missing tool or permission instead of inventing its contents.The expected subtotals are tea 36, mug 50 and bag 18, for a total of 104. This checks arithmetic, not the model’s intelligence. If the result differs, compare the input rows, multiplication and addition separately.
| Concept | Professional explanation | Everyday example |
|---|---|---|
| Workspace | The folder containing inventory.csv. | Check the selected path and the path in the tool call. |
| Model and tool | The model requests a read; a registered tool performs it and returns bytes or an error. | Inspect the call and result rather than treating the answer as an execution log. |
| Context | The returned CSV can inform the next model request. | After editing the file, request a fresh read; an earlier tool result can be outdated. |
| Turn and step | Your submitted request starts a turn, which may involve several model requests and tool calls. | Count the calls you actually see. Do not assume that one user message means one model request. |
| Session | The continuing conversation records the settled messages and tool activity. | Reopen the same session and inspect the earlier read. That is different from restoring the file. |
Then distinguish a file change from conversation history
text
In inventory.csv, change only the quantity of tea from 3 to 4. Read the saved file again, show the exact changed line and recalculate all subtotals and the total. Do not change the other rows.The saved row should be tea,4,12 and the new total 116. Open the file yourself and compare it with the first code block. Reopening or branching a session is not a file rollback: to return to the sample, restore that one row to tea,3,12 and verify 104 again. If you use Git, inspect the diff before deciding what to restore.
- No tool call or a missing-file error: verify the workspace and filename. Do not solve a path mistake by granting unrestricted access.
- A blocked action: read the actual approval or sandbox error and the scope it names. A profile name and a request saying “read only” are not enforcement mechanisms.
- The answer still says 104 after the edit: read the current file again and check whether the answer used an old result from the conversation.
Read the primary documentation
DeepSeek Harness v0.1.3-alpha.1 releaseConfirm the upstream release tag and source commit used by this guide.DeepSeek Harness Alpha READMERead the pinned product definition, developer-preview boundary, and Web UI launch path.DeepSeek Harness Alpha architectureReview the pinned plugin, profile, bundle, tool, and session model.DeepSeek Harness Alpha Web UI guideReview the pinned model setup, workspace selection, first task, and approval behavior.DeepSeek Harness permission presetsSee how sandbox mode and approval policy are combined and where enforcement lives.DeepSeek Harness Skills subsystemReview discovery, providers, identity, loading, and invocation boundaries.
Common questions
Before you make a change
Answers about formats, compatibility, evidence, and rollback.
What does this guide verify?
This guide covers: DeepSeek Harness alpha.1 architecture and Web UI vocabulary; DSH Themes publication and appearance vocabulary; Professional definitions paired with short everyday explanations
What should I do first?
Look up an unfamiliar term, then connect its definition to the everyday example. To see the concepts in action, follow the inventory.csv exercise: read a file, verify the total, change one value, and restore it.
Which boundary matters most?
An analogy is a shortcut for understanding. It does not replace the exact product behavior, especially for permissions, sandboxing, credentials, and compatibility.
Is a profile a filesystem sandbox?
No. A profile composes configuration and plugins. Filesystem and process isolation depend on the active executor and sandbox configuration.
Can a correct answer prove that a tool ran?
No. Check the actual call, its returned content and the file yourself. The expected total is a cross-check, not proof of file access.
Does reopening an old session restore old files?
No. Session history and workspace files are separate state. Inspect the current file and use an explicit, reviewed file restore when needed.
Continue learning
Continue from the verified evidence
Compare published artifacts or return to the installation documentation.