Agent Workflow Specification (AgentTaskboardManifest)

Execution Procedure

It is mandatory for the Agent to read

./reference/syntax_and_semantics.md

as the definitive source of truth prior to evaluating task routes, interpreting user intent, or generating any workflow artifacts.

After establishing this foundation, the Agent must determine the applicable execution route and strictly execute the corresponding standard operating procedures:

Route 1: Planning / Generation

Applies when building a new workflow from scratch collaboratively.

• Core Planning Constraints:
  • No Execution Allowed: The planning phase is strictly for structural modeling. Executing operations, running commands, or validating the environment is strictly prohibited (these actions belong exclusively to the execution phase).
  • Skill Priority: When formulating strategies, always prioritize solutions supported by existing Agent Skills over custom script generation or manual tool creation.

1. Mandatory Reading:

   ./reference/plan_communication_guidelines.md

   .
2. Mandatory Reading:

   ./reference/design_guidelines.md

   .
3. Initial Alignment & Baseline: Engage with the user to outline the strictly macro-level structure. Produce the minimum baseline deliverables (

   -workflow.yaml

   and

   !entry.task.yaml

   ). Do not delve into any execution details, concrete parameters, or deep sub-tasks at this stage. Report the preliminary architecture to the user and wait for the user's confirmation.
4. Iterative Refinement: Progressively expand the workflow details and deep-level tasks through continuous user consultation. You must discuss and design only ONE specific sub-task at a time. Unfolding or generating the entire detailed workflow tree simultaneously is strictly prohibited. Strictly adhere to syntax and design constraints. Arbitrary fabrication of steps or dependencies is strictly prohibited.

Route 2: Formalization / Summarization

Applies when converting a text plan into a manifest, or abstracting completed work into a reusable workflow.

1. Mandatory Reading:

   ./reference/design_guidelines.md

   .
2. Context Processing & Abstraction:
   • For text-based plans: Logically expand the provided details. Proactively query the user regarding any ambiguous logic, missing constraints, or edge cases. Use

     unknown

     for unverified parameters.
   • For summarizing executed work: Extract and abstract hardcoded values into reusable variables. Separate run-specific instances from universal logic to ensure the resulting workflow is highly extensible, not a one-off script.
3. Finalization: Output the complete workflow strictly adhering to the syntax constraints. Generating unsupported operations or assuming user intent without confirmation is strictly prohibited.

Route 3: Loading & Execution

Applies when directed to execute an existing workflow.

1. Pre-flight Check: Verify that the target workflow directory exists and is accessible. Read the

   -workflow.yaml

   to confirm its applicability and prerequisites against the current environment.
2. Mandatory Reading:

   ./reference/execution_guidelines.md

   .
3. Strict Execution: Dynamically load tasks step-by-step, strictly starting from

   !entry.task.yaml

   . The Agent must rigidly follow the defined sequence, evaluate routing conditions based entirely on runtime evidence, and respect all defined human-in-the-loop breakpoints. Deviating from or hallucinating beyond the defined workflow path is strictly prohibited.