Add DeepSeek Harness (DSH) integration (#4336)

* feat: add DeepSeek Harness (DSH) integration

Add a skills-based integration for DeepSeek Harness (DSH), the agent
runtime behind the dsh CLI and Web GUI.

DSH discovers project skills one level deep from .dsh/skills (its
native root, provider rank 100) and .agents/skills (rank 200), using
speckit-<name>/SKILL.md directory bundles with name/description
frontmatter — the same agentskills.io layout SkillsIntegration already
scaffolds. Users invoke skills via the '/'-trigger input in the DSH
Web GUI (/speckit-specify <description>); the harness injects the
rendered <skill_content> into the turn. Project AGENTS.md guidance is
loaded natively by DSH, so no context-file handling is needed.

- Install skills to .dsh/skills/speckit-<name>/SKILL.md (isolated
  static root; multi-install safe alongside other agents)
- CLI dispatch via 'dsh --profile headless "<prompt>"' one-shot mode
- Register 'dsh' in ALWAYS_SLASH_AGENTS so init output and hook
  messages render /speckit-<command> invocations
- Add init next-steps hint, catalog entry, issue-template lists,
  devcontainer install step, docs table row
- Verify with the real dsh-skill-filesystem provider: all 10 skills
  discovered as user- and model-invocable, bodies loadable

Assisted-by: DeepSeek Harness (model: glm-5.3, supervised)

* Address review: bump catalog updated_at, assert DSH next-steps output

- Refresh integrations/catalog.json updated_at to the date of this
  catalog change (matches the bump convention in prior catalog PRs).
- Add TestDshNextSteps asserting the init next-steps panel shows the
  'Start DSH (dsh web)' guidance and .dsh/skills install path, and
  renders /speckit-<command> rather than dotted invocations — modeled
  on TestGrokNextSteps so a regression in the hint or path fails CI.

Assisted-by: DeepSeek Harness (model: glm-5.3, supervised)

* Address DSH review feedback

Place launcher extra arguments before the headless task positional and keep the discovery-depth test compatible with valid skill resource directories.

Assisted-by: OpenAI Codex (model: GPT-5, autonomous)

---------

Co-authored-by: sunxing <sunxing@qianxin.com>

Author: AnInteger

.github/ISSUE_TEMPLATE/agent_request.yml

@@ -8,7 +8,7 @@ body:
       value: |
         Thanks for requesting a new agent! Before submitting, please check if the agent is already supported.
 
-        **Currently supported agents**: Alquimia AI, Amp, Antigravity, Auggie CLI, Claude Code, Cline, CodeBuddy, Codex CLI, Command Code, Cursor, Devin for Terminal, Docker Agent, Factory Droid, Firebender, Forge, Gemini CLI, GitHub Copilot, Goose, Grok Build, Hermes Agent, IBM Bob, Junie, Kilo Code, Kimi Code, Kiro CLI, Lingma, Mistral Vibe, Oh My Pi, opencode, Pi Coding Agent, Qoder CLI, Qwen Code, RovoDev ACLI, SHAI, Tabnine CLI, Trae, ZCode, Zed
+        **Currently supported agents**: Alquimia AI, Amp, Antigravity, Auggie CLI, Claude Code, Cline, CodeBuddy, Codex CLI, Command Code, Cursor, Devin for Terminal, Docker Agent, Factory Droid, DeepSeek Harness, Firebender, Forge, Gemini CLI, GitHub Copilot, Goose, Grok Build, Hermes Agent, IBM Bob, Junie, Kilo Code, Kimi Code, Kiro CLI, Lingma, Mistral Vibe, Oh My Pi, opencode, Pi Coding Agent, Qoder CLI, Qwen Code, RovoDev ACLI, SHAI, Tabnine CLI, Trae, ZCode, Zed
 
   - type: input
     id: agent-name

.github/ISSUE_TEMPLATE/bug_report.yml

@@ -75,6 +75,7 @@ body:
         - Devin for Terminal
         - Docker Agent
         - Factory Droid
+        - DeepSeek Harness
         - Firebender
         - Forge
         - Gemini CLI

.github/ISSUE_TEMPLATE/feature_request.yml

@@ -69,6 +69,7 @@ body:
         - Devin for Terminal
         - Docker Agent
         - Factory Droid
+        - DeepSeek Harness
         - Firebender
         - Forge
         - Gemini CLI

docs/reference/integrations.md

@@ -16,6 +16,7 @@ The Specify CLI supports a wide range of AI coding agents. When you run `specify
 | [Codex CLI](https://github.com/openai/codex)                                         | `codex`          | Skills-based integration; installs skills into `.agents/skills` and invokes them as `$speckit-<command>` |
 | [Command Code](https://commandcode.ai/docs)                                          | `command-code`   | Skills-based integration; installs skills into `.commandcode/skills/` and invokes them as `$speckit-<command>` |
 | [Cursor](https://cursor.sh/)                                                         | `cursor-agent`   |                                                                                                                                           |
+| [DeepSeek Harness](https://github.com/deepseek-ai/deepseek-harness)                  | `dsh`            | Skills-based integration; installs skills into `.dsh/skills` and invokes them as `/speckit-<command>` |
 | [Devin for Terminal](https://cli.devin.ai/docs)                                      | `devin`          | Skills-based integration; installs skills into `.devin/skills/` and invokes them as `/speckit-<command>` |
 | [Docker Agent](https://docs.docker.com/ai/docker-agent/)                              | `docker-agent`   | Skills-based integration; installs skills into `.agents/skills/` (the same directory used by Codex and Zed). In the selected agent YAML, enable local skills with `skills: true` and provide filesystem read access. Detects either the standalone `docker-agent` binary or the Docker CLI plugin (`docker agent`). Configure workflow dispatch with `SPECKIT_INTEGRATION_DOCKER_AGENT_EXTRA_ARGS=./agent.yaml`; the Spec Kit prompt is appended after these arguments. Not multi-install safe by default because the skills directory is shared. |
 | [Factory Droid](https://docs.factory.ai/cli/getting-started/overview)               | `droid`          | Skills-based integration; installs skills into `.factory/skills/` and invokes them as `/speckit-<command>`                               |

integrations/catalog.json

@@ -1,6 +1,6 @@
 {
   "schema_version": "1.0",
-  "updated_at": "2026-08-26T00:00:00Z",
+  "updated_at": "2026-08-31T00:00:00Z",
   "catalog_url": "https://raw.githubusercontent.com/github/spec-kit/main/integrations/catalog.json",
   "integrations": {
     "alquimia": {
@@ -66,6 +66,15 @@
       "repository": "https://github.com/github/spec-kit",
       "tags": ["cli", "skills", "factory"]
     },
+    "dsh": {
+      "id": "dsh",
+      "name": "DeepSeek Harness",
+      "version": "1.0.0",
+      "description": "DeepSeek Harness (DSH) CLI skills-based integration",
+      "author": "spec-kit-core",
+      "repository": "https://github.com/github/spec-kit",
+      "tags": ["cli", "skills"]
+    },
     "amp": {
       "id": "amp",
       "name": "Amp",

src/specify_cli/_invocation_style.py

@@ -13,7 +13,7 @@
 
 # Agents that always render /speckit-<name>, regardless of ai_skills.
 ALWAYS_SLASH_AGENTS: frozenset[str] = frozenset(
-    {"devin", "droid", "grok", "qodercli", "trae", "zed"}
+    {"devin", "droid", "dsh", "grok", "qodercli", "trae", "zed"}
 )
 
 # Agents that render /speckit-<name> only when ai_skills is enabled.

src/specify_cli/commands/init.py

@@ -1013,6 +1013,7 @@ def init(
         devin_skill_mode = selected_ai == "devin"
         zed_skill_mode = selected_ai == "zed" and _is_skills_integration
         grok_skill_mode = selected_ai == "grok" and _is_skills_integration
+        dsh_skill_mode = selected_ai == "dsh" and _is_skills_integration
         cline_skill_mode = selected_ai == "cline"
         forge_skill_mode = selected_ai == "forge"
         bob_skill_mode = selected_ai == "bob" and _is_skills_integration
@@ -1028,6 +1029,7 @@ def init(
             or devin_skill_mode
             or zed_skill_mode
             or grok_skill_mode
+            or dsh_skill_mode
             or bob_skill_mode
         )
 
@@ -1066,6 +1068,11 @@ def init(
                 f"{step_num}. Start Grok Build in this project directory; spec-kit skills were installed to [cyan].grok/skills[/cyan]"
             )
             step_num += 1
+        if dsh_skill_mode:
+            steps_lines.append(
+                f"{step_num}. Start DSH ([cyan]dsh web[/cyan]) in this project directory; spec-kit skills were installed to [cyan].dsh/skills[/cyan]"
+            )
+            step_num += 1
         if bob_skill_mode:
             steps_lines.append(
                 f"{step_num}. Start Bob in this project directory; spec-kit skills were installed to [cyan].bob/skills[/cyan]"

src/specify_cli/integrations/__init__.py

@@ -62,6 +62,7 @@ def _register_builtins() -> None:
     from .devin import DevinIntegration
     from .docker_agent import DockerAgentIntegration
     from .droid import DroidIntegration
+    from .dsh import DshIntegration
     from .firebender import FirebenderIntegration
     from .forge import ForgeIntegration
     from .gemini import GeminiIntegration
@@ -103,6 +104,7 @@ def _register_builtins() -> None:
     _register(DevinIntegration())
     _register(DockerAgentIntegration())
     _register(DroidIntegration())
+    _register(DshIntegration())
     _register(FirebenderIntegration())
     _register(ForgeIntegration())
     _register(GeminiIntegration())

src/specify_cli/integrations/dsh/__init__.py

@@ -0,0 +1,64 @@
+"""DeepSeek Harness (DSH) integration — skills-based agent.
+
+DSH discovers project skills from ``.dsh/skills`` (its native root, highest
+provider rank) and from the shared ``.agents/skills`` root, one level deep,
+using ``<name>/SKILL.md`` directory bundles with ``name``/``description``
+frontmatter — the same agentskills.io layout Spec Kit scaffolds for other
+skills-based agents. Skills are user-invocable through the ``/``-trigger
+input in the DSH Web GUI (and any TUI/ACP front end): typing
+``/speckit-specify <feature description>`` ships the literal token plus the
+user text, and the harness injects the skill's ``<skill_content>`` into the
+turn. Project guidance in ``AGENTS.md`` at the repo root is loaded
+automatically by DSH, so no context-file handling is needed here.
+
+See: https://github.com/deepseek-ai/deepseek-harness
+"""
+
+from __future__ import annotations
+
+from ..base import SkillsIntegration
+
+
+class DshIntegration(SkillsIntegration):
+    """Integration for the DeepSeek Harness (DSH) agent."""
+
+    key = "dsh"
+    config = {
+        "name": "DeepSeek Harness",
+        "folder": ".dsh/",
+        "commands_subdir": "skills",
+        "install_url": "https://github.com/deepseek-ai/deepseek-harness",
+        "requires_cli": True,
+    }
+    registrar_config = {
+        "dir": ".dsh/skills",
+        "format": "markdown",
+        "args": "$ARGUMENTS",
+        "extension": "/SKILL.md",
+    }
+    # ``.dsh/`` is a static, unique agent root that no other integration
+    # writes into, so co-installing DSH alongside other agents is safe.
+    multi_install_safe = True
+
+    def build_exec_args(
+        self,
+        prompt: str,
+        *,
+        model: str | None = None,
+        output_json: bool = True,
+    ) -> list[str] | None:
+        """Build non-interactive CLI args for DSH.
+
+        DSH's one-shot mode is ``dsh --profile headless "<task>"``: the
+        runner submits the task as an ordinary user message, waits for
+        quiescence, and prints the last assistant message to stdout. The
+        headless profile recognizes whitespace-bounded ``/name`` tokens
+        naming user-invocable skills, so a slash-command prompt such as
+        ``/speckit-specify build photo albums`` loads the skill exactly as
+        an interactive session would. The CLI has no structured-JSON output
+        flag, so ``output_json`` and ``model`` are ignored.
+        """
+        args = [self._resolve_executable(), "--profile", "headless"]
+        self._apply_extra_args_env_var(args)
+        args.append(prompt)
+        return args