DeepSeek Harness coding agents handling complex tasks often lack explicit engineering workflows and traceable evidence. The fullstack-expert plugin provides a Cordis-native workflow and safety layer for this, aiming to make agent operations, planning, and verification transparent and controllable.
Plugin Positioning¶
This is a Cordis-native, evidence-driven full-stack engineering specification for DeepSeek Harness agents. It does not introduce a new agent framework, shell, file system, browser, or database client; instead, it registers native capabilities to add a layer of engineering discipline to existing Harness workflows.
Core Features¶
- Cordis-native workflow and safety layer: Serves as a native workflow and safety layer for DeepSeek Harness coding agents.
- Repository inspection and planning: Supports repository inspection and vertical-slice planning.
- Sensitive operation boundaries: Clearly defines the scope of sensitive operations.
- Evidence recording: Records fresh verification evidence and ensures that verification evidence must come from native Harness checks, not self-attestation.
- Deterministic runtime seam: Provides a
FullstackExpertRuntimestate seam for managing phases, tasks, evidence, and compact event facts. - Embedded skills: Registers embedded narrow skills through native
ctx.skills. - Permissions and approval: Includes permission categorization and provides an approval-aware gate before execution; sensitive operations are delegated to native Harness approval, failing closed if unavailable.
- Supabase detection: Optional, read-only Supabase project evidence detection.
- Provider neutrality: Does not select or hard-code LLM adapters during operation.
Installation and Enablement¶
Use the DeepSeek Harness CLI to install this plugin. The installation command activates its declared patch layer.
dsh plugin --profile web add git+https://github.com/adithya-hmt/fullstack-expert.git#v0.1.1
dsh --profile web --dump-config
dsh --profile web
Replace web with the actual name of your profile.
Typical Usage¶
- Request a task: Ask the agent for a feature, bug fix, code review, refactor, or application build.
- Vertical-slice planning: For substantial work, the agent should invoke
fullstack_plan. This action performs repository inspection, states assumptions/unknowns, and uses native tools to implement vertical slices. - Structured checks: Use
fullstack_checkto perform structure, command, or Supabase checks and record non-verification review facts. - Verification for completion: Work may only be completed when native
verificationevidence from Harness is recorded as passing and no relevant failures exist.
Notes and Scope¶
- No additional components introduced: The plugin does not introduce a second agent framework, shell executor, file system backend, browser driver, database client, or model provider. It reuses existing capabilities exposed in the Harness composition.
- Browser verification: Browser verification currently consists of guidance plus evidence recording until stable browser tools are integrated.
- Supabase operations: Supabase operations are limited to detection and recommendations; migrations, linking, login, push, reset, pull, function deployment, and external writes remain explicitly sensitive operations.
- Configuration management: Provider configuration should remain outside this package. Use the Harness provider/model settings surface to configure existing OpenAI-compatible CLI agents (for example,
provider: <your proxy adapter>,model: GPT-5.6 Max). - License: The plugin implementation follows the MIT license.
Summary¶
fullstack-expert aims to make coding agents “show their work.” Through native Cordis capabilities and explicit workflows, it standardizes repository inspection, vertical-slice planning, sensitive operation boundaries, and verification evidence, making it suitable for developers who need rigorous engineering processes in DSH.
